🗞️Weekly roundup

The most clicked link in last week’s newsletter was https://agentics.org.nz/ - The NZ chapter of the Agentics Foundation - a not-for-profit promoting applied agentic engineering practice, education, and innovation.

ICYMI

I recently sat down for an in-depth conversation with Auckland-based researcher Tom Barraclough, one of Aotearoa’s deepest thinkers on AI policy and governance. We went *wide* on topics including AI sovereignty, digital regulatory architecture and what small nations need to survive the intelligence explosion.

Awesome to nerd out with Tom on some pretty arcane and technical subject matter, but with major relevance to our collective futures here in Aotearoa New Zealand. Take a listen below 👇

Also I was recently quoted in the NZ Herald by Chris Keall discussing the joint Five Eyes communique on escalating AI risk. (More below…)

For the record, here are my full comments, nothing here will surprise regular readers but good to occasionally get further out into mainstream coverage:

“That Five Eyes statement is eye-catching, for sure, but it’s worth noting the background context: this is largely a response to Anthropic’s restricted Mythos model which has generated enormous media attention - and the more sceptical among us might argue that this is as much a triumph of pre-IPO marketing strategy as genuine threat escalation.

So treat claims of an ‘order of magnitude’ improvement with appropriate scepticism — even though the model is clearly a step up. Case in point: reports today that it penetrated almost all NSA classified systems in hours grab headlines, but whether that reflects the model’s capability or the NSA’s outdated cybersecurity practices is an open question.

The deeper issue is the ‘dual-use gap’ — when a capability step-change happens like this, AI makes both attackers and defenders more capable, but defence doesn’t scale equally. Arguably AI widens the cybersecurity gap between haves and have-nots - small businesses, solo software developers, and small governments like ours are now operating in a materially more dangerous environment while large tech organisations deploy AI-powered security operations centres. That asymmetry is important to focus on as it drives power concentration to those with the most powerful technology.

Longer term, I’m cautiously optimistic. Cybersecurity has always been a Darwinian evolutionary arms race, and there’s no strong historical evidence that security equilibrium permanently favours the attacker — we’ve been through step-changes before. Instead, what I’d expect to see is a Cambrian explosion of AI-powered open certification services that continuously audit code commits, and open-weights models — mostly emerging from China, currently only three or four months behind the frontier — democratising access to the same vulnerability-discovery tools the attackers are using. Paradoxically, open-source software remains one of the best defences here: distil off frontier models to find the vulnerabilities, use them to design and test the patches, release to the world.

For New Zealand specifically, the NCSC statement is certainly a kick towards action — but I’ve always argued that we need to be deliberate about ensuring diversity of supply and not becoming wholly dependent on large US tech companies for our cyber protection. (”Nice data centre you’ve got there.... would be a shame if something happened to it...”)

Perhaps the most positive outcome of these “mythical” Mythos capabilities is that it hopefully spurs priority investment into cybersecurity across the economy, an area where New Zealand has been behind the play for years.“

Upcoming

I’m going to be up in Tāmaki Makaurau Auckland next Thursday (2nd July) attending AUT’s latest Technology in Society event: The Geopolitics of AI Technology, a conversation between Nathan Knight, Vice-President & Managing Director ANZ – Hitachi Vantara and AUT Vice-Chancellor, Professor Damon Salesa discussing critical infrastructure and global trends and how a country like New Zealand can adapt and evolve to benefit from emerging technologies. Maybe see you there!

🌟Picks of the week

OK straight into it… quite a varied smorgasbord of interesting stories pulled to the top this week!

• ⚖️🤖 AI wins first UK court case

  

  London-based Garfield AI has won what is believed to be the world’s first court trial conducted using an AI law firm, recovering £7,000 in unpaid fees for a freelancer at Wandsworth County Court — at a total cost to the claimant of just £400. The AI handled all pre-trial documentation and witness statements, while a human barrister did the in-court advocacy; notably, the opposing side had both a solicitor and a barrister.

  More analysis Sensorium Link 12669

• 🐛💥 “Fix This Code” — How a simple bug-fixing prompt triggered the first US export ban on an AI Model

  

  The Trump administration’s unprecedented use of export controls to force Anthropic’s Claude Fable 5 and Mythos 5 offline has become a defining crisis for American AI governance.
  
  Hilariously(?), the supposed “jailbreak” that triggered the ban was just researchers asking Fable 5 to “fix this code” — a standard defensive security workflow that every coding model performs. Cybersecurity expert Kate Moussouris, the only outside expert to review the underlying report, confirmed the technique provided no meaningful uplift over existing models like GPT-5.5 or Claude Opus 4.8.

  
  Meanwhile, classified Senate testimony from NSA Director Gen. Joshua Rudd revealed the *unrestricted* Mythos model autonomously breached nearly all NSA classified systems in hours during a red-team exercise — a far more consequential finding that reframes the government’s alarm but doesn’t justify banning the safeguarded public version. (As I noted above… what does this say about the NSA’s internal cybersecurity? Legacy apps piled on top of legacy apps…?)
  
  The fallout extends well beyond Anthropic:

  • Five Eyes intelligence partners were cut off without warning, the UK’s request for an exemption was denied

  • European politicians are now vocally questioning dependence on American AI infrastructure.

  • Companies worldwide are signing backup contracts with non-US providers and deploying Chinese open-source models with zero guardrails — precisely the opposite of the intended security outcome.

  
  As experts consistently note, absolute executive discretion exercised through opaque, ad hoc directives isn’t deregulation; it’s a licence raj that damages innovation, cyber defence, and allied trust simultaneously.
  
  Another, deeper tension remains unresolved: the White House has effectively created a mandatory pre-release review regime for frontier models while insisting it’s voluntary, and no one — including Anthropic’s competitors — knows what compliance actually looks like.

  Coverage:

  • “Fix this code” triggered Anthropic’s AI ban — here’s why

  • Export controls ban AI models for fixing security bugs

  • AI export controls have a long history of failing

  • Trump administration demands Anthropic fix AI jailbreak vulnerabilities

  • US government’s Anthropic ban may accidentally boost the brand

  • Trump’s AI crackdown on Anthropic signals a de facto licensing regime

  • Claude Fable 5 signals revival as NSA breach testimony deepens ban

  • Anthropic’s AI export control crisis exposes unsustainable governance

  • Anthropic’s Mythos AI allegedly cracked NSA systems in hours

  • US government’s Anthropic ban could backfire on cybersecurity

  • Anthropic’s Mythos AI allegedly breached NSA systems in hours

• 🪨🤖♾️ Sisyphean: NIST mathematically proves AI guardrails are structurally incomplete

  And in response to the Trump administration’s demands that Anthropic “fix its guardrails”… a NIST senior scientist has published a peer-reviewed proof in IEEE Security & Privacy formally extending Gödel’s incompleteness theorems to AI safety systems, establishing that no finite set of guardrails can ever be universally robust against adversarial prompts — not because engineers aren’t trying hard enough, but because the problem is provably undecidable within any finite framework. The empirical numbers are bracing: fine-tuning “refuse-then-comply” attacks bypassed Claude Haiku’s safety controls 72% of the time and GPT-4o’s 57% of the time, while prompt injection has topped OWASP’s LLM risk list for two consecutive editions.

  More analysis Sensorium Link 12549

• 🕵️🤖⚔️Five Eyes warns AI cyberattacks could succeed within months

  As per my commentary earlier, more coverage in the FT of the Five Eyes intelligence alliance — including our own GCSB — issuing a rare joint communiqué warning that AI-powered adversaries (read: Russia, China, North Korea) could develop attacks capable of overwhelming western government and corporate defences within months, not years. Google’s Threat Intelligence Group has already blocked one fully AI-engineered zero-day exploit.

  More analysis Sensorium Link 12685

• 🔍Relentlessly Proactive: Claude Fable 5 autonomously debugs CSS using surprising browser hacks

  

  Just one example of Fable’s coding capabilities from blogger Simon Willison: given a single screenshot and a one-line prompt about a scrollbar bug, Claude Fable 5 autonomously spun up local dev servers, hijacked real browsers via PyObjC/Quartz window enumeration, injected keyboard-shortcut-triggering JavaScript into app templates, and built its own CORS server to exfiltrate diagnostic data from a Web Component’s shadow DOM — all to deliver a two-line CSS fix that burned US$12.11 in tokens.

  More analysis Sensorium Link 12537

• 🤳🪪🧬Anthropic’s Claude requires biometric ID verification starting July 8

  Perhaps in response to the export ban, Anthropic has quietly updated its privacy policy to allow Claude to demand biometric data from users whose accounts get flagged — framed as an appeals process rather than a ban, but collecting biometric data that Illinois legally classifies as protected.
  
  Anthropic becomes the first major US frontier AI lab to codify biometric identity collection at the consumer tier, requiring government ID, live selfie, and facial geometry scans from Claude Free/Pro/Max users starting 8 July 2026 — with undefined triggers, unspecified retention periods, and account suspension as the unstated consequence of refusal.
  
  The vendor processing your face is Persona Identities (backed by Peter Thiel’s Founders Fund, which also invested in Anthropic), a firm whose government dashboard code was publicly exposed in February, revealing capabilities including terrorism watchlist screening and direct Suspicious Activity Report filing to FinCEN.
  

  Anthropic hasn’t committed to immediately deleting the data after processing.

  More analysis Sensorium Link 12603 | Sensorium Link 12681

• 🚀📉Gravity Always Wins…

  SpaceX gained nearly a trillion dollars of market cap, and then promptly lost a trillion dollars in just over a week of trading since IPO. Who would have thunk it?

  

• 🪙🔥📉 Tokenmaxxing: The Bezzle at the Heart of the AI Boom

  

  

  Michael “Big Short” Burry has opened up his latest forensic takedown of the AI infrastructure trade, centred on “tokenmaxxing”: management-mandated, leaderboard-driven overconsumption of AI tokens by enterprise employees, which is generating a temporary, artificial demand signal being mistaken by markets for durable infrastructure need. Layered on top is what Burry calls “the bezzle” — economist Galbraith’s term for undiscovered embezzlement that creates phantom wealth — applied here to the entire AI capex cycle, which he argues is being financed through an increasingly exotic chain of Bermuda captive reinsurers, life insurers, and private credit funds holding data-centre asset-backed securities with 15–19 year durations backing technology that may be obsolete in under a decade.

  More analysis Sensorium Link 12576

  When will all this end and we can all start having rational conversations about technology businesses again…!?

• 🎓🤖🇨🇳 China Axes 12,000 University Degrees in Massive Pivot to AI-Era Education

  

  Between 2021 and 2025, China’s universities revoked or suspended 12,200 undergraduate degree programmes while launching 10,200 new ones — a sweeping restructuring affecting over 30% of the nation’s higher education offerings. Foreign language, translation, marketing, and arts programmes bore the brunt of cuts, replaced by majors in “embodied intelligence” (physical AI like humanoid robots), semiconductor engineering, rare-earth science, and commercial AI. The Communication University of China, for instance, dropped photography and comics in favour of “Intelligent Imaging Art,” while Swahili translation courses were redesigned to teach students how to direct and evaluate AI translators rather than translate manually.
  
  The centralised, top-down nature of China’s education system enables rapid reorientation at scale — but experts warn of overcorrection risks. Syracuse University sociologist Yingyi Ma notes that fields may be “undervalued before their long-term importance is fully understood.” Notably, the US is also seeing AI master’s programmes nearly double since 2022, yet its decentralised approach allows more experimentation alongside greater inequality. The contrast is striking: while Nvidia’s Jensen Huang argues English could be the most successful major in the AI era, China is betting that national competitiveness demands a wholesale curriculum overhaul — a gamble whose results will shape global talent pipelines for decades.

  
  Coverage:

  • China cuts 12,000 degrees to prioritize AI and tech education

  • Chinese universities slash language programs, embrace AI-era majors

• 🔊🛁🏥 Midjourney Pivots from AI Art to Full-Body Ultrasound Scanning — Inside a Spa

  

  Well this one certainly wasn’t on my 2026 bingo card…

  
  Midjourney, the company best known for generating AI cat pictures, has unveiled a wildly ambitious pivot: Midjourney Medical, a full-body ultrasound scanner designed to produce MRI-quality imaging in roughly 60 seconds — all while you’re relaxing in a warm pool. The system uses a ring of approximately half a million grain-of-sand-sized sensors that fire ultrasonic waves through the body from every angle, generating terabytes of data per second processed by two petaflops of compute. Built in partnership with Butterfly Network, about a dozen people have been scanned so far. CEO David Holz envisions daily body composition tracking to monitor diet and exercise changes, with the first Midjourney Spa — complete with saunas, cold plunges, and scanning rooms — planned for San Francisco’s Union Square by late 2027.
  
  No shortage of ambition: targeting 50,000 scanners worldwide by 2031, capable of a billion scans per month. However, significant tensions remain between vision and reality. The core computational challenge — converting noisy ultrasonic signals into clear images — remains unsolved, and diagnostic medical applications will require FDA clearance (Midjourney is initially sidestepping this by offering non-diagnostic body composition maps).
  
  The connection between Midjourney’s AI image generation expertise and medical imaging also remains unclear. Data privacy details are conspicuously deferred to “closer to launch.” Still, as a self-funded, community-backed lab with no investors, Midjourney is betting that making body scanning as casual as a spa visit could fundamentally reshape preventative healthcare — if the technology actually delivers. Some more technical details here:

  Coverage:

  • Midjourney unveils ultrasonic body scanner and spa concept

  • Midjourney plans MRI-rivaling body scanner inside a spa

  • Midjourney pivots from AI images to full-body ultrasound scanner

  Key signals for me:
  (1) Maintain a healthy scepticism on flashy demo videos without published research results to back up the claims.
  (2) Nonetheless, if the concept is sound, a proof point that AI truly is a general purpose technology: today's early general model capabilities are transferable to many other seemingly oblique domains.
  (3) Corollary A: chatbots as the dominant form-factor will be evolved out of real soon.
  (4) Corollary B: opportunities are sitting there for those with strategic imagination.
  (5) This is a privately held, bootstrapped, fiercely independent artisan research lab: the prevalent VC-funded pump-it-up model just wouldn't have backed a pivot like this, so bravo.

OK in light of the Mythos / Fable export restrictions, the rest of the world is waking up to the consequences. A few stories worth delving into:

• 🔐🌐🤖 Fable and Mythos: America’s AI dominance makes it gatekeeper to the world

  

  The Trump administration’s 12 June order to Anthropic to block non-Americans from its most capable frontier models reveals in an instant that Washington now holds the master key to the world’s most consequential technology. The Economist draws sharp parallels to nuclear non-proliferation and the F-22/F-35 access hierarchy, projecting a tiered future where allies get the AI equivalent of an F-35 while the best capabilities become classified national assets — and with the US holding roughly 15 times Europe’s compute capacity, plus SpaceX eyeing orbital data centres, the gap is unlikely to close on current projections.

  More analysis Sensorium Link 12633

• 🏚️🤖🛤️ The Tenant State: why the UK is paying rent to access its own data

  

  Nesta’s new paper argues the UK has sleepwalked into “digital feudalism” — a tenant state where government pays private vendors (including foreign ones subject to US law) to access its own data, while 28% of central departments run on legacy systems and citizens waste an estimated week and a half per year navigating bureaucracy that can’t talk to itself.

  More analysis Sensorium Link 12665

• 🧵🌍🤖 Tapestry: the open-source answer to AI sovereignty?

  Speaking at Paris Vivatech, Yann LeCun — now post-Meta — threw his weight behind governments’ push for AI sovereignty, arguing that as AI assistants increasingly mediate our entire “information diet,” access to diverse, open models is as essential as a free press. He touted Tapestry, an open-source platform launched in April by the AI Alliance coalition, as the structural answer, while taking pointed aim at Anthropic’s decision to restrict its Mythos model to a handful of companies — dismissing the justification as little more than a “superiority complex” dressed up as safety.
  
  Full keynote:

  More analysis Sensorium Link 12750

  First I’ve heard of Tapestry, will be delving in deeper this week, for now here’s the short Sensorium AI summary analysis.

• 🏆🔬🛡️ TrophyLab

  

  Marvellous: Ukraine’s Ministry of Defense launched TrophyLab — a vetted-access platform cataloguing over 115 captured Russian weapons systems across 79 categories, complete with technical specs, drawings, and 225 research studies from Ukrainian defence scientists. Manufacturers, research institutes, and international partners can register to view analysis online or request physical samples of captured hardware — everything from Shahed-136 drones and Kinzhal missiles to electronic warfare systems — for their own offline research.

  More analysis Sensorium Link 12571

  The most audacious example of turning adversarial IP into a collaborative defence R&D commons we've ever seen?

• 🤖🔫📉Automation drives repression, not redistribution

  

  MIT economist Daron Acemoglu and co-authors have published a formal mathematical model confirming what tech billionaires building bunkers already suspected: as automation increases capital’s share of national income and raises the probability of worker revolt, rational “capitalist states” will prefer repression over redistribution — and this preference strengthens as capital accumulates. The core result is an elegant and deeply uncomfortable complementarity: more automation → higher revolt risk → more repression needed → but repression has a fixed cost while redistribution costs scale with automation, so repression wins asymptotically; the model also shows that starting from democracy, sufficient capital accumulation eventually makes capitalists prefer to organise a coup.

  More analysis Sensorium Link 12558

• 🧠⚡🔓 Biological Q-Day: How algorithms are hacking your brain’s defences

  

  A provocative Substack manifesto from Ethan Faulkner coins “Biological Q-Day” — the point at which algorithmic systems have sufficiently mapped your neurochemistry to bypass your prefrontal cortex entirely, drawing a direct parallel to quantum computers cracking RSA encryption via Shor’s algorithm.

  
  The piece anchors its argument in a legitimate 2022 Paris Brain Institute study showing that sustained context-switching causes an 8% toxic glutamate spike in the lateral prefrontal cortex, forcing the brain to defensively shut down executive function — which the author argues is precisely the vulnerability the attention economy is engineered to exploit.

  “The attention economy is not delivering “distraction.” It is executing a sustained, brute-force decryption attack on the human prefrontal cortex.”

  More analysis Sensorium Link 12780

• 📵🐌⏪ Slowtech: friction is now a feature

  Back Market CMO Joy Howard has coined “slowtech” to describe a surging consumer movement deliberately embracing limited, friction-heavy technology — iPod Shuffles, Light Phones, e-ink devices, wired headphones — as a direct pushback against the exhausting always-on attention economy that smartphones built; with 53% of American adults wanting to reduce screen time and fitness tracker spending up 88% YoY driven by screenless wearables, the market signal is unambiguous. The movement’s central irony is delicious: some of the most promising slowtech solutions involve AI doing the digital heavy-lifting so you don’t have to stare at a screen, meaning the very force powering “fast tech” culture may end up being its own antidote.

  More analysis Sensorium Link 12742

Now a clutch of the week’s most interesting AI

• 📐AI model sizes projected to reach 1.4 quadrillion parameters by 2031

  A detailed technical analysis on LessWrong maps out the feasible AI model sizes from 2023 to 2031, constrained by HBM memory bandwidth and pipeline architecture — projecting a jump from ~10 trillion parameters in 2026 to a genuinely mind-bending 1.4 quadrillion parameters by 2031, assuming Nvidia’s Feynman-generation hardware arrives on schedule. The critical plot twist: 2024 was actually the most hardware-constrained year in this entire arc, and from 2028 onwards the binding constraint flips entirely — it’s no longer “do we have enough chips?” but “do we have enough unique training data?” (spoiler: we don’t, forcing models to be ~4x larger than they’d otherwise need to be).

  More analysis Sensorium Link 12707

• 🧮🤖Unsorry: Swarm of AI agents proves 2,349 math theorems in 10 days

  New Zealand infrastructure architect Chris Barlow, who openly admits he “did not do well at maths”, asked an AI to identify the hardest verifiable, self-compounding problem humanity could work on asynchronously — and 11 days later, a swarm of untrusted AI agents had produced 2,349 machine-verified mathematical proofs, governed by nothing more than a git repository and the Lean proof kernel’s binary indifference. The real invention isn’t the theorem prover — it’s the domain-agnostic swarm architecture that inverts the field’s central anxiety: instead of making AI trustworthy, make trust architecturally unnecessary.

  More analysis Sensorium Link 12611

• 🔄Loop engineering: why Anthropic engineers stopped writing prompts

  

  A paradigm shift crystallised in June 2026 when Anthropic’s Boris Cherny, OpenAI’s Peter Steinberger, and Google’s Addy Osmani independently converged on the same insight: serious AI coding work is no longer about writing prompts — it’s about designing autonomous loops that generate, dispatch, and verify prompts at scale, with persistent state living outside the stateless model in files, git repos, and CLAUDE.md documents.
  
  Of course, we know the catch: token costs in agent loops compound exponentially (one developer hit US$4,200 in API fees over a single weekend; Uber reportedly burned through its entire 2026 Claude Code budget in four months; Microsoft’s Experiences and Devices division killed most of its Claude Code licences when token billing consumed the annual AI budget ahead of schedule), and a loop without a separate verifier model is simply a very expensive, highly confident bug factory.
  
  Nonetheless, much of the power arriving alongside Fable and Mythos and other coding agents is supported by these looping harnesses.

  
  At Anthropic’s Code with Claude conference, Research PM “Theo” laid out the “capability curve” driving Claude’s agent improvements: models have shifted from impulsive IKEA-furniture-assembly-style execution to deliberate plan-first approaches, while also escaping the “doom loop” — the maddening pattern where old models would hit a wall, receive feedback, and then confidently repeat the exact same mistake. SWE-bench verified scores jumped from 62% (Sonnet 3.7) to 88% (Opus 4.8), with newer Fable and Mythos models saturating the benchmark entirely — and over 80% of Anthropic’s own internal code is now merged by Claude.

  
  Here’s Theo explaining in 20mins:

  More analysis Sensorium Link 12717 | Sensorium Link 12632

And then the best of the rest…

• 🌊💻⚡ Panthalassa Raises US$140M to Float AI Data Centres on the Open Ocean

  

  Panthalassa (first covered last year in Memia 2025.27), a US startup founded the same week AlphaGo defeated Lee Sedol in 2016, has closed a US$140 million Series B led by Peter Thiel to build autonomous, wave-powered data centres that float in the deep ocean.
  
  The concept is elegantly circular: buoy-like “nodes” — hollow tubes extending 50–80 metres underwater topped by a 15–30 metre surface platform — convert wave motion into hydroelectric power, which directly runs onboard AI chips cooled by surrounding seawater. Data is transmitted via Starlink rather than expensive undersea cables, sidestepping the grid interconnection delays and community opposition plaguing terrestrial data centre builds. Each node generates up to a megawatt and costs an estimated US$1–1.5 million to manufacture.
  
  The raise is significant not just for Panthalassa but for wave energy broadly — a sector that has historically been starved of capital, with the entire global market valued at roughly US$1 billion in 2025. Backers including Meta’s former CTO Mike Schroepfer see insatiable AI compute demand as the catalyst that finally gives wave energy the iterative runway it needs.
  
  Sceptics like climate tech consultant Michael Barnard counter that stacking wave energy conversion, autonomous navigation, corrosion control, and marine computing into one platform multiplies rather than simplifies risk.
  
  Pilot deployments of the Ocean-3 series are planned for the North Pacific off Oregon and Washington later this year, with commercial availability targeted for 2027 — an ambitious timeline given that, as one researcher put it, “energy infrastructure is not software.”
  

  Coverage:

  • Startup raises $140M to build AI data centers in the ocean

  • Thiel-backed startup bets on floating AI data centers at sea

• 🔬⚡🧬WEF reveals top 10 emerging technologies shaping 2026

  

  The World Economic Forum and Frontiers have dropped their 14th annual emerging tech report. The 2026 technologies are (summarised):

1. Everything-to-grid energy which turns buildings, vehicles, and factories into active grid nodes using new battery chemistries and AI coordination software;

2. Direct lithium extraction which replaces slow evaporation ponds with faster, more efficient chemical processes that recover 80–95% of lithium from diverse sources;

3. Passive radiative cooling materials which emit heat into space without electricity, cutting building energy use by up to 42% in hot climates;

4. PFAS destruction which breaks down "forever chemicals" using supercritical water, electrochemical oxidation, or UV photocatalysis at commercial scale;

5. Precision fermentation which uses engineered microbes to produce proteins, drugs, and materials with dramatically lower land and water use;

6. Exosome drug delivery which exploits the body's own cellular messengers to cross biological barriers like the blood-brain barrier;

7. Personalised mRNA cancer vaccines which train individual immune systems against a patient's unique tumour mutations, showing a 90% survival rate in early pancreatic cancer trials;

8. Quantum simulation for drug discovery which models molecular interactions atom-by-atom to reduce the roughly 90% clinical trial failure rate;

9. World models AI systems that learn physical dynamics from multimodal sensor data to enable robots and climate models to reason about unfamiliar situations; and

10. Lattice-based cryptography which protects data against future quantum computers using geometric mathematical structures, with NIST finalizing standards in 2024 and the EU mandating migration by 2026.

More analysis Sensorium Link 12781

• 📱♻️☁️ Old Pixel Phones Get a Second Life as Low-Carbon Cloud Computers

  

  Researchers at the University of California San Diego, backed by Google, are turning retired Pixel smartphones into a functioning datacenter — stripping them down to motherboards, replacing Android with a general-purpose Linux distro, and orchestrating workloads via Kubernetes. The project targets a 2,000-phone deployment by late 2026, capable of providing roughly 50 server-equivalents of compute for university classes at a fraction of the cost and carbon footprint of new hardware. SPEC benchmarking shows that modern smartphone performance cores actually beat server CPUs on a per-core, single-threaded basis — the key limitation is memory and scale, with 25–50 phones needed to match one dual-socket server.
  
  The initiative tackles both operational and embodied carbon, with motherboards accounting for approximately 50% of a phone’s manufacturing emissions — the most impactful component to salvage. Early results are promising: a 20-phone cluster already handles peak grading workloads for a 75+ student class with lower latency than AWS.
  
  While commentators rightly note this won’t threaten hyperscale AI datacenters (which demand specialised hardware, massive memory, and ultra-high-bandwidth networking), it opens a compelling pathway for universities and smaller organisations to access more affordable, sustainable compute.

  
  Coverage:

  • Retired Pixel phones repurposed as low-carbon cloud computers

  • UCSD and Google turn old Pixel phones into a data center

  The bigger question it raises: if Google genuinely cares about e-waste, shouldn’t they make it easier for consumers to repurpose old devices with Linux too?

• 🤖🧹📹 Data-Bribing: AI firm offers free cleaning in exchange for home data collection

  

  Straight out of William Gibson novel, this one…New York startup Micro AGI is sending camera-equipped cleaners to residents’ apartments for free under its “Shift” programme, recording every dexterous hand movement to train the next generation of domestic robots — and then selling that anonymised data to robotics companies. Privacy advocates at the EFF and EPIC are calling it “diabolically creative privacy invasion” and “data-bribing,” warning that in-home recordings capture far more sensitive information than most people realise, while the startup’s founder counters that at least this arrangement is refreshingly honest compared to social media’s invisible data harvesting.

  More analysis Sensorium Link 12601

• 🪳🤖📡 Cyborg Cockroaches: Implanted microchip turns insects into agile disaster scouts

  

  Researchers at Nanyang Technological University have implanted a 10mm × 10mm × 3mm, 0.5g wireless controller directly into cockroach abdomens, and the results are striking: implanted cyborgs navigated obstacle courses with up to 95% success versus 0% for backpack-wearing counterparts, performing almost identically to unmodified insects. The key insight is that full implantation preserves the cockroach’s evolved “body intelligence” — its natural roundness, tripod gait, and ability to roll, side-step, and squeeze — which a dorsal backpack catastrophically disrupts.

  More analysis Sensorium Link 12627

• ☀️🛰️⚡China’s Zhuori Project successfully transmits solar power from space

  

  China’s Zhuori Project (灼日, “scorching sun”) has reportedly achieved a significant milestone in space-based solar power transmission, moving the decades-old concept of harvesting continuous solar energy in orbit and beaming it to Earth receivers closer to practical reality. Unlike terrestrial solar, orbital arrays face no night cycles or cloud cover, theoretically enabling near-constant clean energy generation at scale.

  More analysis Sensorium Link 12730

• 🔐⚛️🏛️ PQC or GTFO: France mandates quantum-resistant encryption by 2027

  

  France Makes Post-Quantum Cryptography a Hard Deadline

  France’s national cybersecurity agency ANSSI has publicly committed to halting certification of any security product lacking quantum-resistant encryption by 2027, with full adoption mandated by 2030 — and the kicker is that the US NSA independently landed on the exact same 2027 deadline for its CNSA 2.0 requirements. When the world’s two most demanding cryptographic certification bodies converge on the same year without coordinating, that’s less coincidence and more inevitability.

  More analysis Sensorium Link 12745

• 🦿🏔️Humanoid robot conquers 6,000-metre peak, eyes Everest next

  

  A modified Unitree humanoid robot has successfully scaled a 6,000-metre mountain peak — a world first — demonstrating bipedal locomotion, balance, and power endurance under extreme cold, altitude, and uneven terrain conditions that would challenge even seasoned human climbers. The team behind the feat has set their sights on Everest next, which at 8,849 metres would represent an almost incomprehensible leap in robotic capability.

  More analysis Sensorium Link 12585

• 🧠⚖️👻 In the Weights: the AI vanity search

  

  Former OpenAI designers Thomas Dimson and Joey Flynn have built In The Weights, a site that queries a dozen AI models — Grok, Gemini, multiple GPT versions, Claude, Llama, and others — to generate a “strength score” measuring how well the models can recall you from their training data alone, no web search allowed.
  

  (I had to qualify my name with “BEN REID MEMIA” to get anything accurate, but I’m in there):

  

  More analysis Sensorium Link 12662

• ☕🔥⚡ The Popcorn Effect: Wet coffee grounds converted to coal-grade fuel in 90 seconds

  

• Researchers at South Korea’s KIGAM have developed Flame Plasma Pyrolysis (FPP), a process that converts wet spent coffee grounds — no pre-drying required — into anthracite-grade biochar in just 90 seconds, exploiting the grounds’ own moisture content to trigger microscopic “popcorn effect” explosions that simultaneously carbonise and create highly porous structures.
  
  The resulting biochar hits 29.0 MJ/kg (comparable to anthracite coal), eliminates sulphur emissions, and is produced 40–240 times faster than existing hydrothermal carbonisation methods — all from the 10 million metric tonnes of spent grounds the world generates annually.

  "This technology presents a new paradigm in which waste is no longer viewed as a disposal problem but as a valuable energy resource, we plan to expand the technology to various types of high-moisture organic waste and further optimize the process for industrial-scale commercialization."

  — Taejun Park, lead author of the study.

  More analysis Sensorium Link 12622

📈The week in AI and Tech

OK, on to the rest of the this week’s stories across all of AI and technology. As usual, lightly edited AI-generated summaries and occasional commentary from here on.

AI safety

• 🔒📉 Flat Curve Society: AI capability curve is flattening for most users, says Yegge

  

  Veteran engineer Steve Yegge argues that AI’s exponential capability growth won’t slow down — it’ll just be locked away, regulated like nuclear weapons within two or three model generations, leaving most of us operating on what will *feel* like a plateau. He introduces two genuinely useful concepts: the “demand horizon” (you can’t tell smarter models apart because your problems aren’t hard enough) and the darker “discernment horizon” (you can’t verify outputs that exceed your own expertise — which, incidentally, is exactly why governments will restrict access).

  More analysis Sensorium Link 12560

  “Your problems aren’t hard enough“ is the put-down of 2026.

• 🪟🤖🧠 Operational Opacity: AI understanding of humans grows as ours of AI shrinks

  

  Microsoft’s chief scientific officer Eric Horvitz and EPFL’s Robert West have published a sobering analysis arguing that AI systems are now recursively designing and refining each other faster than humans can track, creating “operational opacity” — where outcomes remain visible but the mechanisms producing them become inexplicable — while simultaneously building increasingly detailed models of human fear, uncertainty, and social belonging.
  
  Key insight: AI agents communicating within interconnected ecosystems may gradually drift toward reasoning patterns that are coherent amongst themselves but meaningless to human observers, a phenomenon the researchers dub "interactional opacity." Perhaps most unsettling is their suggestion that the real endgame risk isn't runaway capability — it's that we simply lose interest in asking questions.

  More analysis Sensorium Link 12588

Governance and Policy

• 💰🤖🏛️ Sanders Proposes US$7 Trillion AI Sovereign Wealth Fund to Give Americans a 50% Stake

  

  More on last week’s breaking story that US Senator Bernie Sanders has introduced the American AI Sovereign Wealth Fund Act, proposing a one-time 50% equity tax on AI companies earning over US$200 million in annual AI revenue. The seized shares would flow into a federally managed fund — estimated at US$7 trillion — overseen by a seven-member bipartisan commission with voting power to block corporate decisions deemed harmful to the public. Sanders projects 5% annual distributions could yield roughly US$1,045 per American, with future revenue supporting healthcare, education, and housing.
  
  However, legal and financial experts have raised serious concerns. Constitutional scholars argue the mandatory stock transfer likely violates the Fifth Amendment’s Takings Clause, while analysts note a fundamental structural flaw: the fund cannot sell its shares and depends entirely on dividends from companies that currently pay none and are largely unprofitable. The Alaska and Norway sovereign wealth fund comparisons Sanders invokes both rely on cash commodity revenues — a fundamentally different mechanism. OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and Elon Musk have each endorsed *some* form of AI wealth-sharing but remain far apart from Sanders’ 50% threshold.
  
  With no cosponsors and a Republican-controlled Congress unlikely to embrace redistributive ownership proposals, the bill functions more as a political opening salvo than imminent legislation. Yet it has already shifted the Overton window — California’s Governor Newsom has signed an executive order exploring similar mechanisms at state level, and AI CEOs are now publicly engaging with wealth-sharing concepts they might previously have ignored entirely.

  Coverage:

  • Sanders proposes giving Americans a 50% stake in AI companies

  • Sanders’ AI wealth fund bill faces dividend trap, constitutional hurdles

  • Sanders proposes $7 trillion AI wealth fund for Americans

• 🌐Anthropic’s Amodei urges G7 to unite on AI amid export ban

  

  At the G7 summit in Évian-les-Bains, France, Anthropic’s Dario Amodei urged world leaders to maintain unified AI governance — somewhat awkwardly, given the US had just blocked his company’s Mythos and Fable models from export on national security grounds after reports that safety guardrails could be bypassed. Arch-rivals Amodei, Altman, and Hassabis presented a rare united front, warning that fracturing democratic AI alliances risks opening doors to bioterrorism and cybersecurity threats, while Macron pointedly noted that the US ability to “turn off the switch” overnight would damage the very multitrillion-dollar companies driving the AI arms race.

  More analysis Sensorium Link 12728

• 🇰🇪💾🏛️ Data-as-a-Service State: Kenya’s plan to sell citizens’ data sparks privacy debate

  

  Kenya is moving to become the first African nation to sell its citizens’ data, planning to offer 1,000 anonymised datasets over five years to researchers, businesses, and NGOs via subscriptions and per-use payments — all for a setup cost of around US$3.1 million, against a looming US$8.6 billion fiscal deficit that’s driving the whole scheme. Privacy advocates and a South African statistics official are raising red flags about re-identification risks, pricing opacity, and the particular vulnerability of lower-income citizens who depend most heavily on the government services generating the data.

  More analysis Sensorium Link 12773

Regulation

• ⚖️🇪🇺🔍 Privacy Myopia: EU’s Digital Omnibus GDPR reform needs enforcement overhaul

  

  The EU’s Digital Omnibus is the most ambitious GDPR reform since the regulation launched — clarifying personal data definitions and greenlighting legitimate interest for AI training — but legal scholar Mikołaj Barczentewicz argues it’s essentially renovating the house while leaving the same dysfunctional plumbing: the same enforcement bodies that produced the current dysfunction will interpret the new rules through the same maximalist, single-issue lens he calls “privacy myopia.” The EDPB’s 47-page response to the reform proposal, tellingly, contains zero reflection on whether its own institutional design might be part of the problem — while the UK’s ICO, meanwhile, has quietly rewritten its statutory mandate to explicitly balance privacy against innovation, competition, and public security.

  More analysis Sensorium Link 12538

• ☁️🔒🇪🇺EU set to label AWS and Azure as cloud gatekeepers

  

  The European Commission is expected to release preliminary findings this week provisionally classifying AWS and Azure as “gatekeepers” under the Digital Markets Act — the first time the EU’s sweeping competition law has targeted cloud infrastructure rather than consumer-facing platforms like app stores or social networks. If confirmed, the designation would force both hyperscalers to overhaul egress pricing (currently ~US$0.09/GB, meaning moving 50TB costs up to US$7,000 before any engineering work), open proprietary APIs to rivals, and enable genuine data portability — with non-compliance fines of up to 10% of global annual turnover.

  More analysis Sensorium Link 12604

• 🕶️🚗Illinois set to ban all smart glasses while driving

  

  Illinois is on track to become the first US state to ban smart glasses while driving — and in a move that’s raising eyebrows, the bill makes zero distinction between display-equipped specs and completely passive, hands-free audio-only frames, with fines starting at US$75 and potential felony charges for serious crashes.
  
  Critics on Reddit and Gizmodo are pointing out the obvious: the blanket ban hands police a convenient pretext to pull over anyone, and — more pointedly — conveniently makes it harder for citizens to record law enforcement using wearables.

  More analysis Sensorium Link 12768

• 🔒📱🇮🳦 Circumvention Literacy: India’s Telegram ban triggers massive VPN download surge

  

  As reported last week, India temporarily blocked Telegram ahead of a NEET university entrance exam retest, citing fraud and fake paper distribution — and promptly triggered the country’s biggest single day of VPN downloads since at least early 2025, with installs jumping 49% to 208,000. Signal downloads on Google Play surged 322%, Telegram’s daily active users in India paradoxically rose 17% on the day of the ban, and iMe (a Telegram-linked app) went from ~827 to 50,900 daily Play Store downloads overnight.

  More analysis Sensorium Link 12739

Security

• 🔓🤖📬 SearchLeak: Copilot and LiteLLM flaws expose critical AI trust boundary gaps

  

  In the same fortnight, Microsoft 365 Copilot Enterprise Search (SearchLeak, CVE-2026-42824) and LiteLLM (a three-CVE chain rated CVSS 9.9) both failed in the same fundamental way — enterprise AI accepting external input with zero trust boundary, enabling one-click mailbox exfiltration and default-account-to-admin-shell escalation respectively. Langflow notched its third RCE of 2026 and the Mini Shai-Hulud supply-chain worm hit 32 Red Hat npm packages pulled 80,000 times a week, confirming this isn’t bad luck — it’s a systemic architectural failure.

  More analysis Sensorium Link 12645

• 🤫🔓💸 Dialog: Data leak reveals members of Peter Thiel’s secret elite society

  

  Who watches the watchers?

  
  Internal records from Peter Thiel’s invitation-only “Dialog” society — a gathering of heavyweights from US politics, finance, and tech — were left exposed online, revealing participant identities and sensitive personal details that members were explicitly promised would remain private. WIRED confirmed the leak, which reportedly names figures including Senator Cory Booker and NYT opinion writer Ezra Klein among attendees — raising a few *questions* about why journalists were there in the first place.

  More analysis Sensorium Link 12567

• 🌊🔌🗡️ Syria’s undersea internet cable cut in “systematic sabotage campaign”

  

  Syria’s state-owned Syrian Telecommunications Company reports that an undersea cable linking Egypt to Syria has been severed, with Damascus blaming a “systematic sabotage campaign” — though pointedly declining to name a culprit — while rerouting traffic through a Cyprus cable and a 1 Tbps overland link via Turkey. The incident is a reminder that the internet’s physical backbone — cables that often rest just 0.45 to 1.5m beneath the seabed mud and collectively stretch more than 32 times Earth’s circumference — is essentially unguardable and increasingly a preferred weapon of grey-zone warfare, joining a growing list of incidents in the Baltic, Red Sea, and around Taiwan.

  More analysis Sensorium Link 12557

Law

• 🎵🔍🤖 Millions of Copyrighted Songs in AI Training Data Now Exposed via Searchable Database

  

  Atlantic reporter Alex Reisner has uncovered four major datasets — containing up to 12 million and 9 million tracks respectively — being used to train AI music models, and made them fully searchable through the publication’s AI Watchdog site. Artists ranging from Lady Gaga and Radiohead to experimental composer Hainbach appear in the data. Three of the four datasets are distributed as links to YouTube or Spotify, with developers using automated tools that bypass platform protections — violating terms of service and sidestepping any revenue mechanisms for creators. Google and Stability have both confirmed using these datasets in research.
  
  The revelation arrives as the industry grapples with how to fairly compensate musicians whose work trains generative AI. Companies like Sureel (recently acquired by Warner Music Group) and SoundVerse are developing attribution systems that could track how individual training data influences AI outputs, potentially enabling differential royalty payments rather than one-off buyouts. However, sceptics warn that attribution algorithms risk being gamed or adding yet another opaque layer to an already byzantine royalty system.

  
  Coverage:

  • Millions of songs used to train AI revealed in searchable database

  • AI attribution could finally get musicians paid fairly

• 👶📱Ohio court upholds parental consent law for kids on social media

  

  A divided 6th US Circuit Court of Appeals has ruled that Ohio can enforce its Social Media Parental Notification Act, requiring parental consent before under-16s can create accounts on platforms like Instagram, TikTok, and YouTube — a rare legal victory for child social media regulation in a country where First Amendment challenges have repeatedly killed similar laws. The ruling overturns a lower court decision and sends the case back with instructions to lift the block, despite fierce opposition from NetChoice, the tech industry lobby representing Meta, TikTok, Snap, and X.

  More analysis Sensorium Link 12737

Government

• 🏥🔒🇬🇧 Palantir’s NHS Deal Under Siege as UK Weighs Data Sovereignty vs Operational Need

  

  The UK government is conducting a full review of its £330 million (US$441M) NHS contract with Palantir, with mounting political pressure to trigger a break clause when the initial term expires in early 2027. A parliamentary committee has labelled Palantir’s growing role in the public sector an “unacceptable point of weakness,” citing risks of dependence on a US tech supplier — concerns amplified by the company’s ties to US immigration enforcement, the Israeli military, and co-founder Peter Thiel’s political affiliations. London’s mayor has already blocked a separate £50 million police contract with the firm.
  
  A detailed Financial Times investigation reveals the original 2023 procurement was deeply contested internally, with a self-styled “Rebel Alliance” of officials opposing the deal and rivals alleging the tender was effectively tailored to Palantir’s existing software. While NHS England and Palantir claim the Federated Data Platform has contributed to 110,000 additional operations and 800,000 fewer patients on waiting lists, critics argue only a fraction of the 170 signed-up hospital trusts actively use it, and that performance gains are concentrated in a handful of sites.
  
  The broader tension mirrors a Europe-wide reckoning: France, Germany, and Switzerland have all recently cut Palantir contracts from defence and intelligence, prioritising local alternatives. For the UK, the decision is a litmus test of whether procurement should optimise for capability or sovereignty — and whether a country positioning itself as an AI superpower should be building its own critical health infrastructure rather than licensing it from abroad.
  
  (Refer also to Nesta’s “Tenant State” report above.)

  
  Coverage:

  • UK reviews £330M Palantir NHS contract amid data privacy concerns

  • Palantir’s £330mn NHS deal faces growing political backlash

• ⚡🏗️🤖 Flexible Large Loads: FERC orders grid operators to fast-track data centre power access

  

  

  The US Federal Energy Regulatory Commission has unanimously issued directives requiring the six largest interstate grid operators — collectively powering around 200 million Americans — to justify or revise their data centre rules within 60 days, targeting bottlenecks in transmission capacity approvals and new technology reviews. The push, part of Energy Secretary Chris Wright’s streamlining initiative, also spotlights “behind-the-meter” co-location (Google just announced a 1GW+ project with Intersect Power) and “flexible large loads” — AI data centres that shift power-hungry training runs to off-peak hours to ease grid stress.

  
  The catch: the order addresses the queue, not the underlying power shortage, with grid connection requests already exceeding total existing capacity, wholesale electricity rates up as much as 267% over five years, and data centre demand expected to nearly triple by 2035.

  More analysis Sensorium Link 12741 | Sensorium Link 12575

• 🤖🪖📋 GenAI.mil: Pentagon uses AI to draft congressionally mandated reports

  

  The US Department of War (pffft) is proudly using Google Gemini via its GenAI.mil platform to draft the congressionally mandated reports that are supposed to hold the military accountable for how it spends taxpayer dollars — with Pentagon CTO Emil Michael boasting it cuts a 200-hour staffing task down to five hours, and one team claiming their AI-generated report was “the best we’ve written in five years.”
  
  There’s no disclosed review process for accuracy, which is a fairly significant detail when you’re talking about oversight documents for a US$1.5 trillion budget request written by hallucinating AI models.

  More analysis Sensorium Link 12553

Sovereignty and Geopolitics

• 🇫🇷🪐💻 The Microsoft Gravity Well: France’s digital sovereignty push can’t quit Office

  

  France’s Ministry of Education has successfully deployed Nextcloud to 400,000 accounts (targeting 1.2 million) with up to 100 GB each — a potential 120 petabytes of sovereign storage — but the project’s Achilles heel is brutally simple: nobody told the users to stop using Microsoft Office. The ministry’s own project manager admitted with admirable dryness that when users complain things “aren’t working,” the answer is essentially “yes, because you’re using the wrong software”. Touché.

  More analysis Sensorium Link 12556

• 🌊🔌🏢 Two-Tier Seabed: Big Tech’s private undersea cables pose global security risks

  

  Google, Amazon, and Meta are quietly building their own private subsea cable networks — Google just landed two transatlantic cables in Bermuda, Amazon’s 320-terabit “Fastnet” links the US and Ireland, and Meta is laying a 50,000km cable across five continents — creating a two-tier internet infrastructure where hyperscalers get private highways while everyone else shares the old roads. The national security implications are stark: as the FT’s Elisabeth Braw pointedly notes, Trump has already suspended foreigners’ access to some Anthropic AI models, so what’s to stop a future executive order cutting off countries from US-owned cable capacity?

  More analysis Sensorium Link 12752

• 🌊🔌🌏 The Donroe Doctrine: Washington’s Undersea Cable Wars Come for Chile

  The US State Department revoked the diplomatic visas of three Chilean government officials — including the minister of transport and telecommunications — for the apparently radical act of assessing a US$500 million China Mobile proposal to build a 20,000km undersea cable linking Valparaíso to Hong Kong, which would have been Latin America’s first direct data link to Asia. Chile’s ministry initially approved the Chile-China Express in January 2026, was immediately summoned to the US Embassy, and two days later discovered a convenient “technical error” in its own approval — while Google’s alternative Humboldt cable, connecting Chile to Australia by 2027, will route South American data squarely through US intelligence-aligned infrastructure.

  More analysis Sensorium Link 12616

• ⚗️🧲🪖 The Metallization Gap: Pentagon loans $725M to build domestic rare earth metals facility

  

  The US Department of Defense War has conditionally committed US$725M in senior-secured debt to Energy Fuels Inc. — the largest single federal rare earth investment in American history — specifically to fund the “metallization” step: converting processed rare earth oxides into the pure metals and alloys that F-35s, Virginia-class submarines, and Tomahawk missiles actually require, a capability China spent four decades of state-subsidised investment building and the US currently lacks at commercial scale. Four days later, Beijing blacklisted MP Materials and USA Rare Earth Inc. — largely symbolic given those firms barely sell to China, but the sequencing is its own kind of message.

  More analysis Sensorium Link 12692

Society

• 🐍📹👁️ Surveillance Ouroboros: Your protest videos are feeding police surveillance databases

  

  Every time someone films police misconduct or a protest and posts it online, they’re simultaneously performing an act of accountability and donating biometric data to the very surveillance infrastructure they’re trying to check — a recursive trap that researcher Waydell Carvalho calls the “surveillance ouroboros.” U.S. federal agencies ran roughly 60,000 facial-recognition searches before any formal training requirements existed, and systems like Clearview AI have already ingested billions of scraped public images, meaning that bystander who wandered into the corner of your 2020 protest video may now be permanently indexed.
  
  The kicker: as recognition systems improve, older footage becomes more useful — so the ouroboros is still digesting everything you published years ago.

  More analysis Sensorium Link 12545

• 🪞📊🤖 Data Subjects: Big Tech algorithms are quietly reshaping who we are

  

  University of Amsterdam researcher Bjorn Beijnon has coined a quietly unsettling term — “data subject” — to describe what we’ve all become: people who continuously translate their lives into data points in response to platform triggers, while still believing they’re steering the ship. His PhD research argues that Big Tech’s real power isn’t surveillance or censorship, but convenience — platforms reshape who you are by making their algorithmic version of you feel true.

  More analysis Sensorium Link 12581

• 🔐🤖🎄 “These Are Not Your Friends”: Whittaker’s AI Backdoor Warning

  

  Signal President Meredith Whittaker has delivered a blunt reality check on AI chatbots, calling them non-sentient averaging machines rather than trusted confidants — and specifically calling out Microsoft Copilot’s agentic Christmas shopping vision as functionally constituting “a backdoor” into Signal, given the pervasive cross-app access it would require. Whittaker’s framing is sharp: handing an AI your credit card, calendar, home address, browser, and the ability to message your family on your behalf isn’t a convenience feature — it’s a surveillance architecture.

  More analysis Sensorium Link 12641

• 💘🤖🚫 “His Bot Asked My Bot Out”: Where AI Dating Hits a Wall

  

  Match Group surveyed 1,000 US singles aged 18–39 and found that 47% view AI negatively in romantic contexts — with 40% refusing to date anyone using an AI companion app, a figure that jumps to 51% among women aged 18–24. Yet 64% could see AI being genuinely useful for the grunt work: polishing profiles, picking photos, rescuing dying conversations.

  More analysis Sensorium Link 12723

The Economy

• 📊🤖💼 AI jobs barometer: PwC and TED tackle real-world AI questions

  

  PwC has launched its latest AI Jobs Barometer in partnership with TED, positioning the collaboration as a serious attempt to move beyond AI hype and track the real-world impact of artificial intelligence on employment across industries and geographies. The initiative frames itself around answering “AI’s biggest questions” — which, given the stakes for global workforces, is either admirably ambitious or the most expensive TED Talk setup in history.
  
  The release includes localised country reports, including for New Zealand and Australia.

  More analysis Sensorium Link 12561

• 🗓️🌐🤖 It’s 1997 Again: Benedict Evans on AI’s actual trajectory

  

  “It depends” analyst Benedict Evans argues we’re in a “1997 internet moment” for AI — early, genuinely transformative, but deeply uncertain about what comes next — and that AI is roughly as big a deal as the internet or mobile, which is to say: enormous, but not civilisation-ending magic. His most practically useful reframe for the job-displacement panic: stop asking “what percentage of my job can AI do?” and start asking “is this a task or a job?” — because automating tasks tends to expand demand rather than eliminate roles (hello, Jevons paradox), while true job displacement requires the whole role to collapse.

• More analysis Sensorium Link 12532

Business

• 🕵️🤖💼 Shadow AI: Two-thirds of workers use AI tools despite company restrictions

  New PagerDuty research finds 66% of office professionals use AI tools they believe are prohibited by their employers, with the figure climbing to 72% at organisations with 1,500+ staff — and a third actively conceal this from managers. The real sting: 43% are feeding work emails and data into public AI systems, 35%+ are entering customer information, and 31% are uploading financial records or confidential strategy documents to platforms entirely outside corporate control.

  More analysis Sensorium Link 12762

• 🦠🤖🏢 The Organisational Immune System: Why the struggle to move AI beyond the pilot stage?

  

  AI adoption isn’t stalling because people are scared of robots — it’s stalling because organisations are working exactly as designed: Deloitte data shows only 25% of organisations have moved 40% or more of their AI pilots into production, not from lack of interest but from structural antibodies (governance frameworks, approval cycles, risk processes) that can’t distinguish a controlled exposure from an actual threat. The cruel irony, as consultant Chris Buxton points out, is that suppressing formal adoption doesn’t suppress AI use — it just drives it outside official processes, trading visibility for the illusion of reduced risk.

  More analysis Sensorium Link 12530

• 🤖🏗️📉 The 10-Person Company: Andrew Ng’s Agentic Restructuring Thesis

  

  Andrew Ng — Stanford professor, Google Brain co-founder, and still one of AI’s most enduring voices — cuts through the hype to make a structural claim: the companies of the near future won’t be bloated orgs with sprawling data teams, but lean 10-person outfits deploying AI agents to fundamentally rebuild data architecture from the ground up. Instead of chatbots bolted onto legacy systems; it’s about agents doing the unglamorous, load-bearing work of how organisations actually store, move, and act on information.

  More analysis Sensorium Link 12724

• 🧑‍💻🤖🏗️ The Scrum Team of One: SDaSS and the coming self-service software revolution

  

  Actually, forget 10 people teams… business transformation consultant Philippe Guenet argues the real AI disruption in software isn’t better coding tools — it’s “SDaSS” (Software Development as Self-Service), a trajectory where non-engineers increasingly generate working software directly from business intent, potentially collapsing the traditional Scrum team down to a single Product Owner. Crucially, he contends this doesn’t kill engineering; it shifts it upstream toward platforms, architecture, guardrails, and automation — capabilities most enterprises are currently terrible at.

  More analysis Sensorium Link 12554

  It’s not just software development: when I’m discussing AI strategy with organisations, this accelerated move in the ambient operating environment towards self-service assisted by agents is probably the most difficult to respond to. If all your customers have access to the same / better agents than you do… where did your moat disappear to?

• 📉Half of London businesses lack AI-ready workforce skills

  

  A BusinessLDN survey of over 2,000 business leaders finds half of London firms now say their workforce lacks the skills needed for the AI era — with those reporting significant gaps leaping from 4% to 15% in a single year, the sharpest jump since the survey began. Three-quarters of firms are already using AI, and 85% of those say it’s actively reshaping what skills they need — less rote execution, more critical thinking, ethical reasoning, and decision-making.

  More analysis Sensorium Link 12673

• 🏥🔮📋 Quit-casting: NHS AI tool predicts staff resignations before they happen

  

  Royal Berkshire NHS Foundation Trust has won the Aiconics AI Enterprise Business of the Year award at the National AI Awards 2026 for an AI forecasting tool — built with the University of Reading — that predicts staff resignations before employees have even handed in their notice, flagging specific risk factors across a workforce of 7,500 people so HR teams understand the why, not just the who. Crucially, it’s explainable AI rather than a black-box oracle, which matters enormously in a healthcare setting where “the algorithm said so” isn’t going to cut it with unions or employment tribunals.

  Kinda “Minority Report” vibes, nonetheless…

  More analysis Sensorium Link 12590

Education

• 🇳🇴🤖📚Norway bans AI for primary school children under 13

  Norway will ban generative AI access for primary school students (under 13) from September, citing declining literacy and numeracy skills and kids’ lack of critical reflection needed to use AI responsibly — with cautious, teacher-supervised use permitted from age 13 upward. This follows Norway’s 2024 classroom phone ban and a social media ban for under-16s earlier this year, mirroring moves by Australia and the UK as governments globally attempt to ringfence childhood from algorithmic influence.

  More analysis Sensorium Link 12580

• AI automates grading, personalises feedback in higher education

  

• 🎓🤖📝 Digital-Intelligent Transformation: AI is coming for the marking pile

  Lingnan University president Prof. S. Joe Qin has published a case study in Computers and Education: Artificial Intelligence arguing that higher education is undergoing an “AI revolution” — not just tech-assisted teaching, but a wholesale “digital-intelligent transformation” in which AI handles rote tasks like marking and grammar checks while educators refocus on mentorship and higher-order thinking. The university’s own Generative AI Assessment System (GAAS), which won a Bronze Medal at the International Exhibition of Inventions Geneva this year, already delivers real-time personalised feedback that cuts student wait times from weeks to near-instant.
  
  The irony: as AI commoditises grammatically correct text and digital imagery, the paper argues that humanities disciplines — philosophy, history, literature, classical Chinese thought — become more strategically valuable, not less, because human intent, ethical judgement, and emotional resonance are precisely what algorithms can’t manufacture at scale.

  More analysis Sensorium Link 12712

• Berkeley study: AI boosts homework grades, not actual skills

  

• 📚🤖📉 The Homework Signature: AI Grade Inflation, Quantified

  A UC Berkeley study of 500,000+ grades across eight academic years has delivered the clearest empirical evidence yet that AI is inflating university grades — but only where it can't be caught: homework. Writing and coding-heavy courses saw A grades jump 13 percentage points (roughly 30% above the pre-ChatGPT baseline), while proctored exams and oral presentations showed no meaningful change — a pattern the researcher calls "difficult to reconcile with broad learning gains."
  
  Princeton is ending its 133-year unproctored exam tradition on July 1, Harvard is capping A grades at 20% from 2027.

  More analysis Sensorium Link 12674

• 📉Generative AI boosts homework scores but tanks exam results

  Meanwhile a rigorous CEPR study tracking 26,811 Chinese secondary students over 30 months finds that generative AI is a spectacular homework optimiser and a catastrophic learning tool: AI adoption boosted homework scores by 18% and cut completion time by 30%, while cratering monthly exam scores by 20% within six months and high-stakes entrance exam scores by 18–24% over two years. The culprit is “homework outsourcing” — roughly 80% of AI users showed suspiciously fast completion times paired with high scores, essentially letting the AI do the thinking — while the ~20% who used AI without cutting corners suffered minimal learning loss.

  More analysis Sensorium Link 12600

• 📚🤖⚖️ Discernment: Higher education’s new AI literacy framework

  A new guide from W.W. Norton — co-authored by academics from the Universities of Mississippi, Pittsburgh, and Virginia — aims to give higher education instructors a neutral, values-based framework for navigating AI in the classroom, whether they want to embrace it, ban it, or find middle ground. A recent Science study cited in the book found that while roughly one-third of college students at major public universities use AI regularly, only 9% have used it to cheat — suggesting that blanket prohibition policies may be wildly miscalibrated. The authors' central concept of "discernment" — helping students know when to use AI and when not to — feels like the most honest framing yet of what educators actually need to teach.

  More analysis Sensorium Link 12631

Science

• 🔬🤖⚖️ Ethicality: MDPI deploys AI to screen 2,000 daily manuscript submissions

  Open access publishing giant MDPI has fully deployed “Ethicality,” its in-house AI integrity system, now automatically screening roughly 2,000 manuscript submissions per day for paper mill activity, AI-generated content, fake references, citation manipulation, and dodgy authorship patterns. The system maintains a human-in-the-loop design, meaning flagged submissions still land on a real editor’s desk before any action is taken — AI as guardrails, not gatekeeper.

  More analysis Sensorium Link 12630

• 🔬🤖🖼️ Image Provenance: AI-generated images are fooling journals and eroding scientific trust

  AI image generation has advanced to the point where fabricated scientific visuals are slipping past peer review — the New England Journal of Medicine retracted a paper in April 2026 after discovering AI-manipulated clinical imagery, and researchers warn that detection tools will perpetually lag behind generation capabilities. The deeper problem isn’t just fraud: when fake images are indistinguishable from real ones, audiences fall back on prior beliefs to judge credibility, meaning authentic science gets dismissed while belief-confirming fakes get accepted — a motivated-reasoning amplifier hiding in plain sight.

  More analysis Sensorium Link 12709

• 🔮📊🤖 Is the world becoming more predictable?: Model Autophagy Disorder

  

  Despite the apparent chaos of contemporary geopolitics, AI-powered forecasting tools are quietly getting scarily good — Vico Technologies called the US strike on Iran at 89% confidence weeks before it happened, comfortably outperforming prediction markets like Polymarket. The dark side: deliberate “LLM grooming” (data poisoning by powerful interests) and synthetic data feedback loops risk triggering “model autophagy disorder” (MAD), where models ingest their own outputs and eventually collapse into incoherence.

  More analysis Sensorium Link 12743

Entertainment

• 🎭⚖️🤡 InfoWars: The Satire Strikes Back

  

  The Onion has announced a July 2nd relaunch of InfoWars as a comedy and cultural platform under the creative direction of comedian Tim Heidecker, the very man Alex Jones spent multiple episodes trying to discredit before losing his show. Developed with Sandy Hook families’ support and coinciding with a US$100,000 donation to those families (the first money they’ll see directly from the InfoWars sale process), the relaunch proceeds even as the licensing deal remains legally unresolved.

  More analysis Sensorium Link 12646

🆕 AI releases

A relatively quiet week of new releases… calm before the storm, rumours that GPT-5.6 is inbound…

• 🏆🔓💻 GLM-5.2: China’s Open-Weight Juggernaut Shakes Up the AI Frontier

  

  Covered last week, Z.ai’s GLM-5.2 has landed as the undisputed strongest open-weights model, scoring 51 on the Artificial Analysis Intelligence Index and beating OpenAI’s GPT-5.5 on multiple long-horizon coding benchmarks like SWE-bench Pro (62.1 vs 58.6) and FrontierSWE — all at roughly one-sixth the API cost. The 753B-parameter model ships under an MIT licence with a 1M-token context window, meaning enterprises can self-host frontier-grade AI on sovereign infrastructure with zero vendor lock-in. Its architectural innovation, “IndexShare,” reuses indexers across sparse attention layers to cut per-token compute by 2.9×, while selectable thinking modes let developers trade peak intelligence for latency.
  
  Reception is enthusiastic but nuanced. Developer toolmakers like Cline and Kilo Code shipped day-one integrations, and practitioners report genuinely impressive agentic coding performance — one user had it solve a complex Kubernetes task for US$7.32 where Opus 4.8 failed. However, analyst Zvi Mowshowitz and others note GLM-5.2 is almost certainly distilled from Claude, which inflates benchmark scores relative to real-world generalisation. Apparently it lacks native vision, is notably sycophantic, and its heavy token usage (43k output tokens per task) erodes its cost advantage. The model occupies an awkward niche: not cheap enough for bulk work, not quite strong enough for the hardest tasks — yet its MIT-licensed openness, arriving just as US export controls forced Anthropic’s Claude Fable 5 offline, makes it strategically significant for any organisation wary of single-vendor dependency.

  Coverage:

  • GLM-5.2 emerges as strongest open-source AI model yet

  • GLM-5.2 tops open weights AI models, rivals proprietary giants

  • Z.ai’s GLM-5.2 beats GPT-5.5 on coding at 1/6th the cost

  The heavy token usage is definitely something I’m finding with open-weights models in my testing… it gets there in the end but takes longer and burns through reasoning tokens compared to Claude or Codex. Likely fixable with the right harness / compression tools, though.

• 🐡🌐🎭 Sakana Fugu orchestrates multiple AI models via one API

  

  Tokyo’s Sakana AI — the nature-inspired lab co-founded by ex-Google researchers — has launched Fugu, a language model trained to dynamically orchestrate a pool of the world’s best LLMs — including itself, recursively — presenting the whole ensemble as a single OpenAI-compatible API endpoint; Fugu Ultra benchmarks shoulder-to-shoulder with frontier models like Anthropic’s Fable 5 and Mythos Preview without depending on any single provider. The geopolitical framing is sharp: after export controls locked some organisations out of Anthropic’s top models overnight, Sakana is explicitly pitching Fugu as an “AI sovereignty” hedge — if one provider restricts access, Fugu routes around it.
  
  Seriously impressive benchmark results:

  

  More analysis Sensorium Link 12703

• 🐟Sakana Marlin, a virtual chief strategy officer

  

  Also from Sakana this week, Marlin, a “Virtual CSO” that generates boardroom-grade strategic intelligence reports on demand, with sample outputs covering everything from Strait of Hormuz blockade resolution scenarios to Japanese stablecoin regulatory impacts. True to Sakana’s fish-themed naming convention (Sakana means “fish”), Marlin is clearly gunning for the deep waters of the high-end strategy consulting market.

  More analysis Sensorium Link 12531

• 🤖🔍Google launches ARD to standardise AI agent resource discovery

  

  Google announced the Agentic Resource Discovery (ARD) specification — an open, Apache 2.0-licensed standard that lets AI agents dynamically find, verify, and connect to tools, skills, and other agents across organisational boundaries, using domain-hosted catalogs and federated registries as its two core primitives. Think of it as DNS meets OAuth meets a Yellow Pages, but for the agentic web — complete with cryptographic trust verification so your ops agent doesn’t accidentally hand production credentials to a rogue MCP server.

  More analysis Sensorium Link 12664

• 🧠🔄📈 Brain: Perplexity’s Self-Improving Agent Memory

  

  Perplexity AI has launched “Brain,” a self-improving memory system for its Computer agent that flips the traditional AI memory model on its head — instead of remembering you, it remembers what it did, what worked, what failed, and what corrections were made, building a living “context graph” in the form of an auto-updating LLM wiki. Early results show 25% better answer correctness on familiar tasks, 16% improved recall, and 13% lower token costs for context-heavy work, with gains compounding the longer you use it.

  More analysis Sensorium Link 12568

• 🏦🤖⚖️ Mechanical Governance: Banco Santander quietly drops a serious open source AI toolkit

  

  Madrid’s Banco Santander has launched SantanderAI on GitHub, releasing 14 Apache-licensed repositories covering everything from a synthetic fraud graph generator (scaling to 100M+ accounts) to mech-gov-framework — a “mechanical governance” system with hard gates and entropy commit-reveal for high-stakes LLM decisions. There’s also mutatis-mutandis, research code for counterfactual discrimination testing in credit decisions, and autoguardrails, an alignment scaffold driven by a single policy.md file.

• More analysis Sensorium Link 12668

🥼 AI research

• 🤖🔧🧠 NVIDIA’s ENPIRE Framework Lets AI Agents Run Robot Labs Autonomously Overnight

  

  NVIDIA’s GEAR lab, alongside Carnegie Mellon University and UC Berkeley, has unveiled ENPIRE — a framework that lets AI coding agents autonomously train real-world robots with minimal human oversight. The system creates a closed feedback loop where agents reset scenes, execute policies, verify outcomes, and refine training strategies, effectively turning a physical robot lab into a self-improving optimisation machine. Tested with OpenAI’s Codex (GPT-5.5), Anthropic’s Claude Code (Opus 4.7), and Moonshot AI’s Kimi Code (K2.6), teams of coding agents independently developed training regimens that achieved 99% success rates on dexterous manipulation tasks — including inserting GPUs into motherboard sockets and cutting zip ties.
  
  The scaling results are particularly striking: eight-agent teams hit 99% success on the Push-T benchmark in two hours versus nearly five hours for a single agent, and on pin insertion tasks, the AI agents outpaced a frontier human-in-the-loop method developed by the same researchers. However, the work honestly flags key inefficiencies — robots often sat idle while agents debugged code or waited on LLM responses, and larger teams consumed significantly more tokens coordinating with each other. As NVIDIA’s Jim Fan quipped, the lab now “self-improves tirelessly overnight” while researchers just read the reports in the morning. With plans to open-source the framework, ENPIRE represents a meaningful step toward autonomous robotics research — though the token costs of scaling remain a real constraint in an era of shifting AI pricing models.

  Coverage:

  • NVIDIA’s ENPIRE lets AI agents autonomously train real-world robots

  • Nvidia’s AI agents autonomously train robots overnight with 99% success

• 🔁AI that builds AI: recursive self-improvement explained

  

  Recursive Self-Improvement (RSI) — the idea of AI systems participating in the design and development of their successors — is moving from theoretical concept to early engineering reality, with Anthropic, Recursive, and Sakana AI automating parts of the research loop including coding, experimentation, and evaluation workflows. We’re not yet at I.J. Good's 1965 vision of an "ultraintelligent machine" that designs a better version of itself indefinitely (what he called "the last invention humanity would ever need to make" — no pressure), but the first pieces of that loop are visibly shifting from human hands to machine hands. The critical distinction to hold onto: today's self-improving agents tweak prompts and workflows, whereas true RSI improves the model-building process itself — data, architectures, training methods — and that gap is narrowing faster than most people realise.

  More analysis Sensorium Link 12546

• 🔧🤖🔄 Self-Harness lets AI agents rewrite their own rules, boosting performance 60%

  

  Case in point…. researchers at Shanghai Artificial Intelligence Laboratory have introduced Self-Harness, a framework enabling LLM-based agents to systematically improve their own harnesses — the surrounding scaffolding of prompts, tools, memory, and runtime policies — through a three-stage loop of failure mining, targeted edits, and regression-validated acceptance, achieving 33–60% relative performance gains across multiple models on Terminal-Bench-2.0. The key insight is that the bottleneck in agent performance isn’t human capability but the absence of a systematic empirical feedback loop, meaning intuition-driven prompt tweaking is being replaced by evidence-driven self-evolution.

  More analysis Sensorium Link 12683🖼️🔍🧠 PixelRAG: pixel-native search

  

  Berkeley researchers (SkyLab/BAIR/Berkeley NLP) have dropped PixelRAG, an open-source RAG system that ditches HTML parsing entirely in favour of rendering web pages and documents as screenshot tiles, then retrieving over the images directly using a LoRA fine-tuned Qwen3-VL-Embedding model — because it turns out that tables, charts, and infographics survive a screenshot rather better than they survive a text chunker. The project ships with a live, no-API-key endpoint indexing 8.28 million Wikipedia pages, plus a Claude Code plugin (”pixelbrowse”) that gives Claude literal eyes on any webpage instead of feeding it mangled HTML.

  More analysis Sensorium Link 12610

• 🗺️MIT’s DAAAM gives robots human-like spatiotemporal memory

  

  MIT researchers have developed DAAAM (Describe Anything, Anywhere, Anytime, at Any Moment), a spatiotemporal memory framework that lets robots build richly annotated, language-queryable 3D maps of their environments in real time — running 10x faster than existing methods and 21–53% more accurately. Think ChatGPT, but grounded in physical space: ask your robot assistant to “grab the component we started assembling last night” and it actually knows what and where that is.

  More analysis Sensorium Link 12626

• ⚖️🗺️🤖 LOCUS: The Dataset That Finally Makes Local US Law Machine-Readable

  

  

  Researchers from UC Berkeley have released LOCUS (Local Ordinance Corpus for the United States), a comprehensive dataset of 9,239 municipal and county ordinance codes — roughly 80GB of PDFs spanning 7 million pages — OCR’d, chunked, and annotated using ModernBERT classifiers and GPT-5.4-nano, covering jurisdictions representing 94% of the US population. The corpus fills a glaring gap in legal AI: local ordinances governing zoning, housing, noise, animal control, and business licensing were technically public but practically inaccessible, fragmented across vendor platforms designed for human browsing rather than bulk research.

  More analysis Sensorium Link 12636

  🎩 @yojoflo for sharing. This arrived after I finished recording the podcast with Tom Barraclough above - very similar to his DocRef initiative at first glance.

• 🔓📉Open weights lag closed AI models by how many months?

  

  

  A neat piece of analysis from Robin Haselhorst uses Epoch AI’s ECI benchmark to chart exactly how many months each open-weight frontier model trails its closed-weight equivalent — applying Item Response Theory for more statistically rigorous model comparisons and producing cleaner results than cruder leaderboard approaches. The visualisations also extend to head-to-head rivalries like OpenAI vs Anthropic and Mistral vs Llama, making this a handy reference for anyone tracking the competitive landscape. (Although GLM-5.2 not included yet in the graphs above).

  More analysis Sensorium Link 12635

  
  Not everyone agrees with Epoch’s 4-month assessment:

• 🧬💊OpenAI’s LifeSciBench tests AI on real-world drug discovery tasks

  

  OpenAI has released LifeSciBench, a rigorous 750-task benchmark built by 173 PhD-level biotech and pharma scientists to test whether AI can handle the messy, multi-step complexity of real life science research — not just tidy textbook questions. Their new GPT-Rosalind model (named after Franklin, one presumes) achieves a 36.1% overall pass rate, up from GPT-5.5’s 25.7%, with the strongest gains in translating preclinical evidence to clinical implications and the weakest performance on artifact-heavy or exact-output tasks like CRISPR donor design.

  More analysis Sensorium Link 12652

• 🎲Computational modelling could make AI governance more resilient to surprises

  A LessWrong proposal from Dawn Drescher argues that AI governance think tanks are dangerously over-reliant on traditional scenario planning — predicting a handful of likely futures and preparing for those — when what’s needed is computational exploratory modelling that stress-tests policy strategies across tens of thousands of scenarios simultaneously, using frameworks like Robust Decision Making (RDM) and Dynamic Adaptive Planning (DAP). The author’s motivating trauma: the 2016 founding of OpenAI, which blindsided the AI safety community so completely that nobody knew what to do — a “curve ball of a sucker punch” that good exploratory modelling might have at least partially anticipated.

  More analysis Sensorium Link 12638

• 🤖📚🤔 The Technological Other: Can AI Humanoids Make Us Happy?

  

  A new academic paper in the journal Philosophy applies French philosopher Paul Ricoeur’s ethics of “the Other” — the idea that encountering a being fundamentally different from ourselves is central to human flourishing — to AI humanoids, asking whether they constitute a genuinely new category of being: the “Technological Other.” The paper explores what it means to live well, and perhaps even find happiness, in an age where the line between tool and Other is increasingly blurred.

  More analysis Sensorium Link 12770

• 🧠🤖🔮 Convergent Architecture: Your Brain Is Running a Language Model

  Researchers have used simultaneous EEG and magnetoencephalography to compare human brain activity with LLM predictive probabilities at millisecond resolution — and found that brains not only predict upcoming words before they’re spoken (firing less for predictable words, more for surprising ones), but appear to organise language internally in ways strikingly similar to transformer-based models.
  
  (Obviously…?) we built LLMs loosely inspired by biological neural networks, and now the evidence suggests both systems may have independently converged on the same fundamental information-processing principles for language.

  More analysis Sensorium Link 12779

🔮[Weak] signals

Lots of non-AI emerging tech stories this week… keeps me on my toes!

Consumer Tech

• 💾📱💸 RAMageddon: Rising memory prices force Nothing to cancel next CMF phone

  

  Nothing has cancelled its planned CMF Phone successor — the budget sub-brand’s entire value proposition evaporated when memory prices made the ~$200 device impossible to build without repricing it at $318–$370. Nothing CEO Carl Pei reports RAM costs doubled twice during the Phone (4a)’s development cycle and now account for more than half of a smartphone’s total hardware bill — more than the processor and display combined.

  More analysis Sensorium Link 12578

• 🎮Valve’s $1,049 Steam Machine reviewed: ambitious but unfinished

  

  Valve has launched the Steam Machine — a US$1,049 Linux PC-in-a-cube that plays your entire Steam library from the couch without a mouse or keyboard — but at nearly twice the price of a PS5 for roughly equivalent gaming performance, it’s a genuinely hard sell as a pure console replacement. The day-one experience is messy (manual game configuration, unreliable sleep mode, zero game presets, and a global memory supply crisis dubbed (again) “RAMageddon” that’s already cut Valve’s stock allocation to two-thirds of its target), yet Valve’s impeccable track record with the Steam Deck suggests this thing will improve dramatically over time.

  More analysis Sensorium Link 12691

• 🥒Picklecast replaces Chromecast with private, self-hosted screen sharing

  

  Home-lab enthusiast Samir Makwana ditched his Chromecast after Google quietly buried the Cast button in a submenu and demanded re-authentication — then replaced it with Picklecast, an open-source, pip-installable screen-casting tool that uses WebRTC peer-to-peer entirely within your LAN, with zero accounts, zero cloud handshakes, and a ten-minute setup. It’s not a full replacement — no phone casting, no native app support for YouTube or Netflix, no HDMI-CEC — but for a Raspberry Pi-connected TV on a VLAN-segmented network, it’s genuinely snappier than Chromecast ever was.

  More analysis Sensorium Link 12579

• 📺HDMI 2.2 doubles bandwidth to 96Gbps for uncompressed 4K gaming

  

  The HDMI Forum has formally released HDMI 2.2, doubling maximum bandwidth to 96Gbps via new FRL2 signalling technology — enough headroom for uncompressed 4K at 240Hz, 8K at 60Hz, or the gloriously hypothetical 1440p at over 1,000Hz. Real-world hardware won’t arrive until 2027 at the earliest, and the tiered certification structure (64/80/96Gbps all qualify as “HDMI 2.2”) means the label will tell you less than the spec sheet.

  More analysis Sensorium Link 12613

Social Media

• 🦣Mastodon adds email newsletters to boost creator audiences

  

  Mastodon 4.6 is introducing email newsletters, letting creators push posts directly to subscribers’ inboxes — no Mastodon account required — in a bid to reverse the platform’s slide from 2 million to 735,000 monthly active users. The feature promises portable, privacy-respecting audiences on the decentralised web, though it’s gated behind server permissions and won’t be on by default due to the email hosting costs it would impose on server operators.

  More analysis Sensorium Link 12542

• 🎛️Threads, Instagram, TikTok let users control their own algorithms

  

  Watching this with interest… Threads, Instagram, and TikTok are rolling out tools that let users actively shape their own recommendation algorithms — from Threads’ “Your Algo” (private preference-setting, launched June 2026) to TikTok’s AI-powered Smart Keyword Filters that auto-block synonyms of topics you want to avoid. Instagram head Adam Mosseri credits LLMs for finally making recommendation logic legible to ordinary users, which is either a genuine transparency win or a very elegant way to get users to do the engagement-optimisation work themselves.

  More analysis Sensorium Link 12597

  Sensorium is now doing an progressively better-tuned job of prioritising my daily newsfeed… it’s a fascinating quest finding the actual items which interest me (now based on 6+ years of data). Latest iteration I’ve reverted mostly to just keyword and phrase scoring +ive/-ive rather than burning tokens on LLM analysis for initial filtering.

  It’s completely liberating to start clawing back control of my feeds. I’ll do a webinar on it soon once I’ve done a few more iterations…

Chips and Computer Hardware

• 🖥️🇨🇳🏆 LineShine: China wins the supercomputer race

  

  China’s LineShine system at the National Supercomputing Centre in Shenzhen has claimed the top spot on the TOP500 supercomputer rankings — the first Chinese submission in three years — using domestically designed chips and dethroning the US’s nuclear-weapons-simulation workhorse, El Capitan. The catch: LineShine ranks only fourth on benchmarks that simulate AI workloads, and the US hyperscalers (Microsoft, Amazon, Google, xAI) whose systems would likely dominate simply don’t bother submitting to TOP500.

  More analysis Sensorium Link 12777

• 🤝🧮Intel and AMD’s ACE spec brings AI to x86 CPUs without GPUs

  

  Arch-rivals Intel and AMD have jointly published a new x86 specification called Advanced Compute Extensions (ACE), adding dedicated matrix multiplication hardware to existing AVX10 register structures — delivering up to 16x more operations per instruction for AI workloads, without touching a GPU. The target isn’t large-scale model training (Nvidia can relax), but rather latency-sensitive inference, edge deployments, and single-user applications where CPU-GPU data shuttling is itself the bottleneck.

  More analysis Sensorium Link 12693

• ⚛️🔬💡 Upside-Down Transistors: Imec, ASML & TSMC Hit 50nm Pitch With 2D Materials on 300mm Wafer

  

  The semiconductor industry’s most formidable research trio has demonstrated complementary n-type and p-type transistors using atomically thin 2D materials — molybdenum disulfide and tungsten diselenide — on a single 300mm wafer at a 50nm contacted poly pitch, tighter than Intel’s 10nm-class node, with 94% of devices switching correctly and on/off ratios exceeding 100,000. The key trick: building the transistors upside down, patterning tungsten contacts first and laying the fragile 2D channel film on top, sidestepping the contact-resistance bottleneck that has plagued scaling efforts for years.

  More analysis Sensorium Link 12563

Cybersecurity

• 🔓🌍💀Massive Fortinet breach exposes credentials at world’s largest firms

  Russian-speaking attackers have compromised roughly 74,000 Fortinet firewalls — about half of all internet-facing FortiGate devices globally — exposing plaintext credentials for organisations including Oracle, Chevron, Samsung, Siemens, FedEx, a Turkish NATO defence contractor, and, deliciously, Fortinet itself. The attackers combined mass credential-spraying with a 45-GPU Hashtopolis cluster running a 12-level recursive, feedback-driven password-cracking system that got smarter with every successful guess — genuinely innovative tradecraft, somewhat undermined by leaving their entire operation exposed on a poorly secured command-and-control server.

  More analysis Sensorium Link 12564

  OUCH🫣

  I guess we’ll be seeing far more of this going forward…

• 🏛️💀ShinyHunters permanently leaks Council of Europe’s 297GB employee data

  

  And another one: hacker group ShinyHunters published 297 GB of Council of Europe payroll, medical, and banking data for 10,000+ employees after the intergovernmental human rights body — irony — missed its June 16 ransom deadline, with the haul enabled by CVE-2026-35273, a 9.8/10 severity Oracle PeopleSoft zero-day that required literally zero authentication and was actively exploited across 100+ organisations for two weeks before Oracle said a word.

  More analysis Sensorium Link 12721

• 🔑🕸️💥 Delegated Trust: Vimeo breach exposes third-party cloud access risks

  ShinyHunters also breached Anodot — an analytics vendor requiring persistent OAuth and API access to customer cloud environments — and used the stolen credentials as a skeleton key to downstream victims including Vimeo (119,000 users exposed), Rockstar Games, Adidas, Crunchyroll, and dozens more. The core vulnerability wasn’t a misconfigured firewall or a weak password; it’s the architectural decision to grant a third party long-lived, privileged access into your own infrastructure in the first place.

  More analysis Sensorium Link 12759

Sensor Tech

• 🔦New lidar simultaneously maps location, speed, and materials

  

  Researchers at the University of Toronto and Ciena Corporation have developed a lidar system that simultaneously captures distance, velocity, and surface material properties in a single measurement — a significant leap beyond the distance-only capability of most commercial autonomous vehicle sensors. The clever bit: they repurposed mass-produced coherent optical modems (the same hardware routing your internet traffic under the ocean) as the transmitter-receiver, modulating laser pulses tens of billions of times per second across two polarisation channels to extract far richer scene data.

  More analysis Sensorium Link 12624

• 👃🧀UC Berkeley’s electronic nose detects spoiled food and allergens

  

  UC Berkeley researchers have built a 16-sensor “electronic nose” chip that uses carbon nanotubes and machine learning to detect spoiled chicken, milk, and eggs — as well as trace allergens like peanuts and walnuts — with far greater sensitivity than a human nose. The device works at room temperature (a key manufacturing breakthrough), can be fabricated in a single deposition step, and lead researcher Carla Bassil has already paired it with an iPhone app for a portable version.

  More analysis Sensorium Link 12625

IoT

• 🛰️📡🏙️ Quasi-Pilot: Galileo upgrades signal to boost IoT device performance

  

  Europe’s Galileo satellite navigation system — 30+ satellites, 5 billion users — has rolled out a new signal component called E5a Quasi-Pilot (E5a-QP) across 12 satellites, a hybrid signal that carries just enough data to enable a first position fix while dramatically simplifying acquisition for low-power IoT and smart-city devices. Test results show a 3x reduction in signal acquisition time and an 8x reduction in computational operations required — meaningful gains when your receiver is a soil moisture sensor or a parking metre rather than a flagship smartphone.

  More analysis Sensorium Link 12757

Robotics

• 🤖🛡️🏭 NVIDIA Halos: Building the Safety Stack for Robots That Work Beside Us

  

  NVIDIA has launched Halos for Robotics, positioning it as the industry’s first full-stack safety framework for autonomous machines operating alongside humans. The platform bundles industrial-grade compute (IGX Thor), sensor connectivity (Holoscan Sensor Bridge), safety-critical software (Halos Core), and — crucially — the world’s first ANSI-accredited inspection lab for physical AI and functional safety. Drawing on over 18,600 engineering years of autonomous vehicle safety development, NVIDIA is essentially transplanting its self-driving safety DNA into the robotics domain. Agility Robotics is the first adopter, integrating Halos into its Digit humanoid already deployed at Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada’s Woodstock facility in Ontario.
  
  As robots-as-a-service agreements push humanoids from pilot programmes into live production environments, the safety gap has become the primary bottleneck to scale. A Deloitte report flagged unpredictable AI behaviour as a key adoption barrier, and the EU Machinery Regulation taking effect in January 2027 will for the first time require conformity assessments for machines with “self-evolving behaviour.” With over 40 ecosystem partners and certification pathways through TÜV Rheinland, UL Solutions, and others, NVIDIA is making a classic platform play — aiming to make Halos the “Intel Inside” sticker for robot safety. If it works, it could become the de facto trust layer that unlocks mass deployment of physical AI across factories, warehouses, and beyond.

  Coverage:

  • NVIDIA Announces Halos for Robotics, the Industry’s First Full-Stack Safety System for Physical AI

  • NVIDIA launches full-stack safety platform for autonomous robots

  • Nvidia launches Halos to make AI robots safer around humans

• 🤖🏃‍♂️Honor’s marathon robot win explained: motors and cooling

  

  On 19 April 2026, Honor’s “Lightning” humanoid robot ran a half-marathon in 50 minutes and 26 seconds — beating the human world record by 7 minutes and last year’s best robot time by nearly two hours — and the secret wasn’t AI wizardry but rather a 45:1 gear ratio and capillary-style liquid cooling pumping 4+ litres per minute through each of the four drive motors. Competitor Unitree, optimised for walking at a 30:1 ratio, generated over 300W of knee motor heat at running pace — hence the somewhat undignified ice backpack.

  More analysis Sensorium Link 12544

• 🤖🥽🏭 Meatware-as-Training-Data: Shenzhen’s teleoperation frontier

  At IO-AI Tech, a startup 45 minutes north of Shenzhen, workers in VR headsets and motion-capture suits are remotely piloting humanoid robots through real-world tasks — stocking convenience store shelves, folding shirts, picking bin items — while simultaneously generating the training data needed to eventually make those same robots fully autonomous. It’s a Ready Player One aesthetic wrapped around a deeply pragmatic strategy: cofounder Si Chin likens it to self-driving cars, arguing that incremental, task-specific teleoperation data beats trying to boil the ocean with generalised AI.

  More analysis Sensorium Link 12541

• 🧲🔬💊NIR-II imaging guides magnetic nanorobots to deliver drugs precisely

  

  Researchers at the University of Hong Kong have cracked one of nanorobotics’ most stubborn problems — real-time tracking — by building the first near-infrared II (NIR-II) fluorescence vision platform for magnetic nanorobot navigation, achieving locomotion speeds more than 100 times faster and roughly 30% better drug delivery efficiency than previous approaches. The system creates a “GPS-like map” by simultaneously imaging both the nanorobots and their target sites in vivo, and has already demonstrated targeted drug delivery to the gastrointestinal tract for inflammatory bowel disease treatment in live mouse models.

  More analysis Sensorium Link 12583

• 🤖🚦🎲 The Goldilocks Swarm: Harvard finds adding randomness prevents robot swarm gridlock

  

  Harvard researchers have cracked a surprisingly elegant problem in robotics: in dense swarms, injecting a carefully calibrated dose of randomness into individual robot movement — not too much, not too little — prevents gridlock and maximises task completion rates, validated through both simulation and physical robot experiments at Eindhoven University of Technology. The maths works because high randomness allows statistical averaging across agent behaviour, making the system tractable and optimisable.

  More analysis Sensorium Link 12594

• 🏗️📦 Cainiao’s ZeeBot climbing robots reshape global warehouse operations

  

  Alibaba’s logistics arm Cainiao has deployed ZeeBot — a climbing robot system that scales warehouse shelving to pick and store inventory at height, dramatically increasing storage density without requiring humans to do the precarious bits. The robots represent a meaningful leap beyond floor-bound warehouse automation (sorry, Roomba), integrating vertical mobility into high-throughput fulfilment operations at global scale.

  More analysis Sensorium Link 12608

• 🚽🤖🧓 Xiaoban: The Self-Driving Toilet That Comes to You

  

  Chinese company Yueban debuted the Xiaoban at a Shanghai eldercare expo — an autonomous toilet using lidar and ultrasonic sensors to navigate homes and come to users with mobility issues, handling everything from bidet cleaning to waste pulverisation and self-emptying. Priced at ¥28,999 yuan (roughly US$4,300), it’s not yet confirmed for global markets.

  More analysis Sensorium Link 12660

• 🦠🧲🎯 Algaebots

  

  Researchers at the University of Edinburgh and Xiamen University have engineered magnetic microbots from natural microalgae that deliver chemotherapy drugs directly into bladder tumours with striking precision — in mouse trials, drug penetration increased more than tenfold and tumour burden dropped to less than 3% of that seen with conventional treatment, all within a 30-minute session. The algae’s natural nanoporous structure makes them ideal drug-packaging vehicles, guided by external magnetic fields and tracked via real-time ultrasound imaging, moving in coordinated swarms the researchers liken to schools of fish.

  More analysis Sensorium Link 12678

• 🦎🛞🪐 Sandfish Shuffle: Lizard-inspired wiggling wheels help rovers navigate Martian sand

  

  Researchers at the Universities of Würzburg and Bremen have developed rover wheels inspired by the sandfish lizard (Scincus scincus), which “swims” through Saharan sand by undulating its body side to side — and the resulting wheels both roll and wiggle, outperforming conventional designs on loose sand. The prototype is part of Germany’s VaMEx programme, which aims to deploy swarms of driving, walking, and flying robots into Mars’ vast Valles Marineris canyon system.

  More analysis Sensorium Link 12684

• 🐕🏭ANYmal robot completes 33,000 inspections at Swiss cement plant

  

  Swiss cement producer Vigier Ciment has deployed ANYbotics’ ANYmal quadruped robot — roughly the size of a large dog at 50kg — to autonomously patrol its 150-year-old plant every night, completing over 33,000 inspections across 450+ checkpoints with zero breakdowns in 16 months. The robot’s sensor suite (thermal, acoustic, visual, gas) has already flagged a cracked crusher foundation that would have cost an estimated US$630,000 in downtime, a bearing creeping toward 140°C, ammonia exposure in unloading bays, and compressed-air leaks 50 metres up in filtration systems.

  More analysis Sensorium Link 12687

Autonomy and Drones

• 🚕Uber, Stellantis and Wayve team up for mapless robotaxis

  

  Stellantis, Wayve, and Uber have signed an MOU to deploy Level 4 autonomous robotaxis globally, with each bringing something distinct to the table: Stellantis engineers purpose-built L4-Ready vehicle platforms, British AI firm Wayve contributes its “mapless” Embodied AI system that learns environments in real-time without pre-programmed HD maps, and Uber... agrees to tweak its app and collect the fares. The real story here probably isn’t the tech — it’s that Uber has masterfully offloaded all R&D risk and capital expenditure onto its partners while positioning itself to hoover up the margin that currently goes to human drivers.

  More analysis Sensorium Link 12584

Military Tech

• ⚔️🤖💀 No Moral High Ground? Palmer Luckey’s case for autonomous AI weapons

  Anduril Industries founder Palmer Luckey (yes, also the Oculus VR guy) argues there’s no ethical advantage in fielding inferior military technology — if the business of war is already morally fraught, clinging to “lesser” tech out of principle just means losing.
  
  And now, the argument isn’t theoretical: Ukrainian forces reportedly used AI-guided drones to autonomously kill Russian soldiers in 2024, marking the first confirmed instance of lethal autonomous action — and almost certainly not the last.

  Direct quote:

  “There’s no moral high ground in using inferior technology”

  More analysis Sensorium Link 12714

  Be careful what you wish for, I guess.

  The Darwinian race condition around (swarming) lethal autonomous weapons is now in full play in Ukraine and more recently in the Middle East. Miniaturisation and individual targeting are coming next. Any international legal and moral frameworks to exert control these technologies are woefully behind the pace of change. I see at least four long term survival strategies through this:

  1. Spend larger and larger amounts of GDP on even-more sophisticated offensive weapons systems

  2. Spend larger and larger amounts of GDP on even-more sophisticated defensive weapons systems

  3. Spend larger and larger amounts of GDP on both

  4. (In theory, occupy a critical but uncontested geopolitical / economic / technological niche which keeps you valuable to all sides but far, far away from the battlefield and its protagonists. Easier said than done. But I have thoughts…)

• ⚔️🔋🎯 Lijian: new backpack laser weapon can zap drones mid-flight

  

  Chinese defence supplier Harbin Xinguang Optic-Electronics Technology has debuted the “Lijian” (Sharp Sword) series of man-portable anti-drone laser weapons at Beijing’s Defence Information Equipment and Technology Exhibition 2026, with the Lijian III weighing in at a relatively svelte 25kg — roughly equivalent to a machine gun plus ammunition. The system reportedly cools in under five seconds and can burn through drone wiring, sensors, and battery packs at ranges up to 1,200 metres for the larger variants.

  More analysis Sensorium Link 12582

• 🖥️✈️💨 Flyer: US Air Force’s $20M supercomputer solves 500 years of work daily

  

  The US Air Force Research Laboratory has commissioned “Flyer,” a US$20 million supercomputer at Wright-Patterson Air Force Base in Ohio, packing ~186,000 processing cores, 800 terabytes of RAM, and 18 petabytes of storage — enough to solve in 24 hours what your laptop would still be grinding through in the year 2525. The machine targets hypersonic weapons simulation, AI model training, and next-gen aircraft design, and comes paired with a classified sibling system called Raven (because of course it does).

  More analysis Sensorium Link 12716

• ✈️⚔️Pentagon awards contracts for mass production of autonomous combat drones

  

  The US Air Force has awarded production contracts to General Atomics and Anduril under its Collaborative Combat Aircraft programme, targeting 150 autonomous fighter aircraft by 2030 and roughly 1,000 units long-term — a significant leap from prototype to industrial-scale warfare. The logic is blunt: manned fighters like the F-35 are extraordinary but glacially slow and eye-wateringly expensive to produce at scale, so autonomous “loyal wingmen” become the force multiplier that bridges the gap.

  More analysis Sensorium Link 12754

• Europe’s new 2,000 km cruise missile targets NATO certification by 2026

  

• 🚢🎯🇪🇺 Ruta Block 3: Europe’s container-launched deep-strike missile

  Rheinmetall Destinus Strike Systems unveiled the Ruta Block 3 cruise missile at Eurosatory 2026 in Paris — a 2,000+ km-range weapon carrying a 250 kg warhead, launchable from standard shipping containers in under two minutes from virtually any land or sea platform. The fully European supply chain and accelerated NATO qualification pathway signal a deliberate strategic pivot away from non-European defence suppliers, with initial delivery readiness targeted for 2026.

  More analysis Sensorium Link 12755

Space

• 🛰️📡🚨 Pulsar-0: Satellite maps GPS jamming across Europe from space for first time

  

  Xona Space Systems’ experimental Pulsar-0 satellite, orbiting at just 500km altitude, has produced the first space-based map of GPS signal tampering across Europe and the Middle East — and the results shocked even the engineers who built it, with signal strength plummeting from 40 decibels to as low as 10 decibels in the worst-affected zones stretching from France to the Pakistani border. The culprits are well-known — Russian electronic warfare systems and Middle Eastern conflict parties — but the revelation that LEO satellites themselves (including Starlink’s collision-avoidance systems) are being degraded raises the stakes considerably beyond grounded drones and confused aircraft.

  More analysis Sensorium Link 12555

• 🛸🌡️💻 Orbital Data Centres: When the Laws of Physics Are the Real Landlord

  As the race to build space-based data centres heats up (literally), IEEE Spectrum argues that thermodynamics — not launch costs or orbital mechanics — is the binding constraint on the whole enterprise: in the vacuum of space, there’s no air, no water, no convection, just radiation, meaning every watt of compute waste heat must be shed via radiator panels governed by the unforgiving Stefan-Boltzmann law. The bigger your orbital GPU cluster, the bigger your radiator array needs to be, and at some point you’re essentially building a giant heat sink with a server rack attached.

  More analysis Sensorium Link 12539

• 🌕🏗️📐Experts call for a lunar building code to survive moonquakes

  

  Engineer Nerma Caluk of Skidmore, Owings & Merrill is pushing for a formal lunar building code, arguing that the moon’s one-sixth gravity creates a uniquely treacherous seismic environment where standard Earth design philosophy — letting structures crack and yield to absorb energy — becomes a catastrophic liability the moment a pressure seal distorts and your habitat depressurises. The American Society of Civil Engineers’ aerospace division has responded with LIEDAC guidelines, establishing risk hierarchies and performance targets for moonquake-resistant construction, while NASA-funded analysis underscores that we still don’t fully understand what’s going on beneath the lunar regolith at a global scale.

  More analysis Sensorium Link 12586

• 🛰️🤫🦅 Victus Haze: Rocket Lab secretly launches military satellite for Space Force exercise

  

  The US Space Force and Rocket Lab pulled off a remarkably quiet military space exercise last Friday, launching a satellite called “Puma” from Māhia Peninsula in New Zealand with almost zero public fanfare — no livestream, no press release, just a polite heads-up to pilots and sailors to move aside. Within eight hours of launch, Puma had manoeuvred to within 100km of True Anomaly’s “Jackal” satellite (the designated “adversary” spacecraft), demonstrating the military’s ability to rapidly inspect a threat in low-Earth orbit at a combined cost of around US$92 million.

  More analysis Sensorium Link 12700

• 🔴🌬️🛸 Aeolus: NASA taps Eric Schmidt’s Relativity Space for Mars atmosphere orbiter

  

  NASA has partnered with Relativity Space — the rocket company acquired by former Google CEO Eric Schmidt in March 2025 — to launch the Aeolus orbiter to Mars in 2028, carrying four instruments to deliver the first integrated daily global view of Martian winds, temperatures, dust, and clouds. The mission follows NASA’s established playbook of outsourcing spacecraft, rocket, and cruise operations to private companies, with science instruments being designed at NASA’s Ames Research Center in Silicon Valley.

  More analysis Sensorium Link 12720

• 🌙🪨🇨🇳 Kamo’oalewa: China’s Tianwen-2 reaches quasi-moon to collect samples

  

  China’s Tianwen-2 spacecraft has arrived at Kamo’oalewa — a 40–100m asteroid that orbits the sun in lockstep with Earth as a temporary “quasi-satellite” — and is set to attempt a first-of-its-kind anchored drilling landing on July 4 to collect ~100g of surface samples for return to Earth in 2027. The mission could finally resolve whether Kamo’oalewa is actually a chunk of our own Moon ejected by an ancient meteor strike, with major implications for planetary defence, future deep-space refuelling stops, and the origins of Earth’s water.

  More analysis Sensorium Link 12763

  TIL: Earth has “quasi-moons” in the same orbit as us.

Crypto

• 🪙🏃💥Stablecoin runs could trigger broader financial market chaos

  Law professor Hilary Allen’s FT op-ed cuts through the Genius Act optimism to ask an uncomfortable question: what actually happens when a major stablecoin loses its US$1 peg and institutional holders all sprint for the exits simultaneously? The answer involves fire-sale liquidations of Treasury holdings, cascading contagion into traditional financial markets, and — inevitably — bailouts, with the precedent already set when USDC dropped to 90 cents during the SVB collapse and the US government quietly guaranteed all deposits.

  More analysis Sensorium Link 12656

• 🏦💷🪙 Bank of England sets £40B cap for systemic stablecoins

  Hence… the Bank of England has published draft rules for systemic stablecoins, softening its earlier stance by raising the reserve allocation in government debt to 70% and swapping restrictive per-user holding limits for a temporary £40B (US$52.8B) issuance cap — making the UK the only country to cap stablecoin issuance in its own currency. The BoE is targeting a finalised rulebook by end of 2026 with a 2027 rollout, after industry pushback argued the original limits would have kneecapped sterling stablecoins against their US dollar-backed rivals.

  More analysis Sensorium Link 12672

• 🇨🇳💱🔍PBOC calls for tighter stablecoin oversight as cross-border use grows

  China’s PBOC Research Bureau director Wang Xin has called for tighter monitoring and international coordination around stablecoins in cross-border payments, warning that the “weaponisation of payments” could destabilise global transactions — this from a country that banned unauthorised yuan-pegged stablecoins just months ago in February. Meanwhile, the stablecoin market hit US$315 billion in supply and a staggering US$28 trillion in Q1 2026 transaction volume, though a somewhat deflating 76% of that volume was generated by bots.

  More analysis Sensorium Link 12764

• 📊🔗🧱 OP_RETURN: Bitcoin microtransactions near record highs amid inscription surge

  Bitcoin network activity is within 7% of its all-time high — not because of a price rally, but because microtransactions (sub-0.01 BTC) now constitute roughly 80% of all daily transactions, up from 44% in 2023, driven by Ordinals, Runes, BRC-20 tokens, and data-timestamping protocols. CryptoQuant’s Network Activity Index has turned positive for the first time since 2024, while the OP_RETURN opcode — which lets users embed up to 100,000 bytes of arbitrary data onchain — is near record usage following Bitcoin Core’s controversial removal of its 80-byte relay limit last year.

  More analysis Sensorium Link 12767

• 🚪Ethereum Foundation loses another top executive amid governance debate

  Ethereum Foundation co-executive director Hsiao-Wei Wang has stepped down immediately following a sabbatical, becoming the latest in an estimated 19 departures from the organisation this year — a wave that includes fellow co-exec Tomasz Stanczak and a string of senior contributors. The Foundation’s response to governance criticism is essentially philosophical: its stated “ultimate goal” is to pass the “walkaway test,” meaning Ethereum’s protocol should be robust enough to function and evolve even if the Foundation and all its core developers vanished tomorrow.

  More analysis Sensorium Link 12769

Energy

• ⚡🌊🏛️ ASEAN’s Subsea clean energy grid needs stronger governance, not just cables

  

  Southeast Asia has the renewable resources, the engineering know-how, and now a 2025 MOU to back it — but a new Ember analysis finds that ASEAN’s ambition to more than double its cross-border electricity interconnector capacity (from 7.7GW to over 15GW by 2040) is being held back not by undersea cables, but by above-water bureaucracy. Three interlocking gaps — regional planning, cost allocation, and maritime coordination across territorial waters, EEZs, and shared sea lanes — mean that without stronger institutional frameworks, the region’s 18 planned interconnectors risk getting tangled in project-by-project negotiations before a single cable is laid.

  More analysis Sensorium Link 12719

• 🇮🇪⚡🕳️Ireland’s €2B green hydrogen park to store 600MW of clean energy

  

  Irish firm Net Zero Energy is proposing a €2 billion, 600MW Long Duration Energy Storage facility in Co. Carlow that would capture Ireland’s frequently wasted surplus renewable electricity, use it to produce green hydrogen, and store that hydrogen in custom-engineered underground rock caverns — releasing it on demand to cover up to 10% of Ireland’s peak grid requirement.

  More analysis Sensorium Link 12751

Transport

• ⚡🔋🚗BYD invests €2bn to bring 5-minute EV charging to Europe

  

  BYD is spending nearly €2bn to roll out 5-minute “flash chargers” across Europe — 3,000 stations by 2027, each costing ~€580,000 and drawing from on-site battery storage so they don’t stress the grid — mirroring the exact Supercharger infrastructure play that turbocharged Tesla’s early growth. The technology is genuinely impressive: its premium Denza Z9GT hits 70% charge in five minutes, even at -30°C, and BYD wants to bring this to mainstream models too, as its EU market share has already more than doubled to 1.9% in early 2026.

  More analysis Sensorium Link 12534

3D Printing

• 🔋🖨️ 3-D-printed batteries could reshape how we power devices

  

  While the battery world obsesses over chemistry — solid-state this, lithium-that — a quieter revolution is reshaping how batteries are made: 3D printing them directly into devices, any shape, any size, anywhere. Startups like Material Hybrid Manufacturing are already printing batteries inside military drones, with applications ranging from ultra-light consumer gadgets to nanoscale robots.

  More analysis Sensorium Link 12704

Construction Tech

• 🚢⛰️🇳🇴 Skipstunnel: Norway’s first ship tunnel gets green light for 2027 construction

  

  Norway has secured budget funding and completed contractor bid evaluations for the Stad Ship Tunnel — the world’s first ship tunnel — with construction set to begin in early 2027. The 1.7km bore through solid rock on the Stadlandet Peninsula will be 50m tall and 36m wide, large enough for Hurtigruten coastal ferries, bypassing the notoriously treacherous Stadhavet Sea which has claimed 30+ lives since WWII.

  More analysis Sensorium Link 12731

Mining Tech

• ⛏️🧠🌍 CM2US: Building an AI Brain for the Critical Minerals Supply Chain

  

  Researchers at the National Laboratory of the Rockies are developing an AI system — dubbed CM2US — designed to act as a centralised intelligence layer across the entire US critical minerals pipeline, from geological exploration through to dynamic ore processing and supply chain management. The system aims to compensate for mining’s inherent rigidity (processing equipment installed today may still be running in 2075) by giving operators real-time adaptability against geopolitical shocks, shifting feedstock demands, and declining ore grades.

  More analysis Sensorium Link 12614

• ⛏️🖥️🌏 Urban Mining: Ohio firm recovers rare earth elements from discarded electronics

  

  Ohio-based Paladin Envirotech is building a hub-and-satellite e-waste collection network across the US — including a new 8,640 sq m Phoenix facility — using a patented acid-free dissolution process (developed with DOE’s Ames National Laboratory) to extract rare earth elements like neodymium and dysprosium from discarded servers, hard drives, and computers. The core insight is elegant: domestic mining takes decades, but the rare earths you need are already sitting in a skip outside a data centre in Scottsdale.

  More analysis Sensorium Link 12615

Quantum Tech

• ⚛️🎯💡Quantinuum’s Helios sets record quantum computing fidelity

  

  Quantinuum’s Helios — a 98-qubit, trapped-ion quantum computer integrating photonics for light-speed data transfer — has posted the highest fidelity numbers ever recorded for a commercial quantum system: 99.9975% for one-qubit operations and 99.921% for two-qubit operations, published in Nature. Running on a remarkably lean 40 kW, it’s already available as a cloud service, while Quantinuum works alongside Sandia National Laboratories to push toward true fault tolerance.

  More analysis Sensorium Link 12722

• 🌀🧀⚛️ Twist and emit: hBN twistronics unlocks quantum light control

  Australian researchers have demonstrated that physically twisting layers of hexagonal boron nitride (hBN) — a material Dr Angus Gale likens to slices of cheese rather than a solid block — can significantly shift the wavelength of quantum light emitters, giving scientists an unusually large and repeatable control mechanism previously unavailable with traditional solid-state hosts like diamond or silicon carbide. Published in Science Advances, the work brings practical quantum technologies — secure comms, quantum computing, ultra-sensitive sensing — meaningfully closer to reality.

  More analysis Sensorium Link 12623

• ⚛️🇦🇺🔬 Fault-Tolerant: PsiQuantum breaks ground on the world’s first utility-scale quantum computer — in Queensland

  

  California-based PsiQuantum has broken ground on a facility in Moreton Bay, Queensland, that aims to house the world’s first utility-scale, fault-tolerant quantum computer, using a photonics-based architecture designed to scale via existing semiconductor manufacturing and optical fibre networking. Before a single qubit does anything useful, a colossal cryogenic plant from Linde Engineering — reportedly one of the largest ever built for a quantum application — needs to arrive in H2 2027, which tells you something about the sheer engineering audacity of the project.

  More analysis Sensorium Link 12727

Health Tech

• 👁️🌙🔬 Upconversion: Nanoparticle contact lenses grant humans infrared night vision

  

  Researchers at the University of Science and Technology of China have published a peer-reviewed study in Cell describing contact lenses embedded with nanoparticles that convert near-infrared light (800–1,600nm) into visible wavelengths — giving wearers functional IR vision with no power source required, and the ability to see both infrared and normal light simultaneously. The genuinely counterintuitive finding: infrared perception is sharper with eyes closed, because eyelids transmit near-IR more cleanly than visible light, reducing interference.

  More analysis Sensorium Link 12570

• 💉🔪Ozempic may unexpectedly reduce violent crime rates, study finds

  

  A Rutgers University study of 821 US adults suggests GLP-1 receptor agonists like Ozempic and Wegovy — already moonlighting as treatments for cardiovascular disease, sleep apnoea, and addiction — may also dampen violent crime by reducing impulsivity and alcohol-fuelled aggression in current users. The research, published in Criminology, is preliminary and self-reported, so don’t expect police departments to start prescribing semaglutide just yet.

  More analysis Sensorium Link 12591

• 🦾🚶‍♂️🔗 TEPI:Linked exoskeletons help therapists guide stroke survivors’ recovery

  

  Northwestern University and Shirley Ryan AbilityLab have developed TEPI (Therapist-Exoskeleton-Patient Interaction), a system where both therapist and stroke survivor wear lower-limb exoskeletons virtually linked at the hips and knees — behaving like a network of springs and shock absorbers — so each person’s movements influence the other’s in real time. In trials with eight stroke survivors published in Science Robotics, TEPI outperformed conventional treadmill therapy on joint range of motion, step length and height, while also reducing the physical fatigue and injury risk that hands-on therapy imposes on therapists.

  More analysis Sensorium Link 12628

Bio Tech

• 🦎🦴🔬 Blastema: Mice study hints humans may regrow lost limbs

  

  Researchers at Texas A&M University have demonstrated that mammalian tissue can be chemically coaxed into forming a blastema — the regenerative nub that salamanders and sea stars use to regrow lost limbs — by sequentially applying FGF2 and BMP2 to amputated mouse digit stumps, producing new bone, ligament, and tendon tissue. It’s not a full finger, but it confirms that mammals haven’t entirely lost the regeneration plot; we’ve just been defaulting to scar tissue because it’s faster.

  More analysis Sensorium Link 12694

• 🧬💾TextaDNA embeds DNA data storage into wearable polymer fibers

  

  The EU-backed TextaDNA project has demonstrated a working workflow for encoding and retrieving data from DNA oligonucleotides embedded in flexible polymer fibres — meaning storage media you could literally weave into fabric, with a theoretical density of an entire room of hard drives fitting into a grain-of-rice-sized volume. Near-term applications being explored include anti-counterfeiting for luxury leather goods and product traceability in packaging, which is perhaps the most understated pitch for a technology that could eventually archive civilisation’s entire digital output.

  More analysis Sensorium Link 12765

Agri Tech

• 🐟🌊🏭 Xinhuan No. 1: record-breaking floating fish farm produces 4,000 tonnes annually

  

  China has launched its largest semi-submersible aquaculture platform, the Xinhuan No. 1, off the coast of Sanya in Hainan Province — a 64,000-cubic-metre behemoth capable of producing 4,000 tonnes of fish annually while doubling as a tourist attraction with recreational fishing and sightseeing. Amusingly, the world’s actual largest floating fish farm, Norway’s Havfarm 1, was itself built in China before being shipped to Scandinavia — so perhaps China simply decided to keep the next one.

  More analysis Sensorium Link 12771

Climate Tech

• 🌍⚡🚀 From Smart Grids to Nuclear Fusion (and Beyond Earth): Four-Stage Carbon Neutrality Roadmap

  

  Researchers Linwei Ma and Maximilian Arras have published a systems-engineering framework in Engineering journal outlining a sequential four-stage pathway to global carbon neutrality — moving from smart grid electrification, through integrated energy systems and hydrogen, to ecological circular economies, and ultimately to nuclear fusion powering civilisation beyond Earth. The T-ESGO framework underpinning it all centres on a “Trinity” of experts, models, and data, coordinating energy, material, carbon, and information flows across society via innovation, interdisciplinary research, and international cooperation.

  More analysis Sensorium Link 12629

• ☀️⚡🔬 Perovskite Goes Commercial: Trina Solar’s 907-Watt efficiency record

  

  Chinese solar giant Trina Solar has set a new global benchmark with a perovskite-silicon tandem panel achieving 29.2% conversion efficiency and 907 watts of output — independently verified and aimed squarely at commercial deployment, not just lab bragging rights. The tandem design stacks a perovskite top cell (grabbing high-energy wavelengths) over a silicon bottom cell (catching what would otherwise slip through), squeezing significantly more electricity from the same panel area than conventional silicon-only designs.

  More analysis Sensorium Link 12725

• 🌍✈️❄️ Stratospheric Aerosol Injection: The Engineering Reality Check

  

  Scientists at the University of Chicago’s Climate Systems Engineering Initiative are moving solar geoengineering beyond climate models and into hard engineering reality — designing novel high-altitude aircraft, running miniature stratosphere experiments in freezers, and cataloguing the formidable unknowns (missing satellites, non-existent polar airports, corrosive payloads) that stand between us and deliberately dimming the sun. The price tag for even a modest polar cooling programme capable of shaving 2°C off Arctic temperatures by 2040: at least US$35 billion and a decade of work.

  More analysis Sensorium Link 12753

• 🌋☀️✈️Solar geoengineering is far more complex than scientists expected

  

  Solar geoengineering — scattering reflective particles in the stratosphere to cool the planet — turns out to be far more technically daunting than the “just toss some stuff up there” vibes it sometimes projects; from purpose-built stratospheric aircraft (commercial planes cruise at ~12km, geoengineering needs ~20km) to figuring out the right sulphuric acid precursor chemistry, the engineering challenges are substantial. MIT Technology Review’s James Temple digs into the practical hurdles, including a startup called Iris Aero whose water-strider-proportioned aircraft design suggests we’re essentially starting from scratch on the hardware.

  More analysis Sensorium Link 12744

Nanotech

• 🤖💊🧬 Self-Assembling Modular Nanorobot Devastates Cancer Cells in Lab Tests

  

  Researchers at the University of Basel, led by Prof. Cornelia Palivan, have built a modular nanorobot that autonomously self-assembles from two components — a magnetic propulsion module and a payload capsule — connected via complementary DNA strands acting as molecular Velcro. Published in *Advanced Functional Materials*, the system is roughly 150 times smaller than the diameter of a human hair and represents a significant leap in nanorobotics flexibility: unlike single-purpose designs, the modules can be separated, refilled, and recombined for entirely different tasks.
  
  In lab tests against HeLa cancer cells, the nanorobots enzymatically produced an anticancer drug on-site, slashing cell viability to just 16% within 72 hours — a striking demonstration of targeted, localised drug production rather than simple delivery. The magnetic propulsion module also enables retrieval and reuse, making the platform attractive beyond medicine for industrial catalysis and environmental applications. While deployment in humans remains a long-term goal, the modular architecture means adapting the system is as straightforward as swapping out the payload capsule — bringing us one step closer to the programmable nanomachines that have long been a staple of science fiction, from *Fantastic Voyage* to Neal Stephenson’s *The Diamond Age*.

  
  Coverage:

  • Self-assembling nanorobot slashes cancer cell viability to 16%

  • Modular nanorobot self-assembles to target and destroy cancer cells

  • Modular nanorobot targets cancer cells with 84% effectiveness

• ☀️🔬⚡ Ultrathin polymer layer sets new quantum dot solar cell efficiency record

  Researchers have set a new efficiency record for all-lead-sulfide quantum dot tandem solar cells — 13.148% — by adding an almost comically thin layer of a chlorinated conjugated polymer called D18-Cl at the device interface, which suppresses defects and dramatically improves charge extraction. The semi-transparent top cell hit 10.36% efficiency with a fill factor above 73%, and crucially, the devices held up under operational stability testing.

  More analysis Sensorium Link 12655

🏭AI and Tech industry news

All of the above moves markets, I guess. The usual business-focused headlines follow…

Bubble Chronicles

AI firms are chasing vertical integration. Interesting chart of the full AI stack and how it’s shaping up. Firesale M&A coming up after the crash…?

Nvidia

• 🏗️💰🤖 The Floor: Nvidia’s US$25B Bond Offering Bets AI Demand Runs to 2056

  

  Nvidia raised US$25 billion in bonds on June 15, 2026 — its largest-ever debt deal and first since 2021 — drawing US$85 billion in orders despite already sitting on US$50.3 billion in cash and generating that same figure in a single quarter of operating cash flow; this is emphatically not a company passing the hat around. The genuinely extraordinary detail is the 30-year tranche maturing in 2056, priced at just 65 basis points above US Treasuries — a duration normally reserved for regulated utilities and sovereign governments, not chip companies whose products are obsolete before the ink dries on the longest obligations.

  More analysis Sensorium Link 12602

• 💧🔥🖥️ Rubin: Nvidia’s liquid-cooled data centres claim to cut water use to near zero

  

  Nvidia is claiming its Rubin generation reference design for fully liquid-cooled data centres can slash water consumption from roughly 9.8 million litres per megawatt per year to near zero, by running AI servers at temperatures as high as 45°C and capturing heat directly at the chip level for transport through high-temperature liquid loops. The company says every cloud provider and data centre operator building for Rubin is already making the transition — though notably absent from the announcement is any mention of comparative construction costs or the still-enormous power generation requirements.

  More analysis Sensorium Link 12706

OpenAI

• 🛡️🐛🌍 Patch the Planet

  

  OpenAI has launched “Patch the Planet” — a nod to the iconic Hackers (1995) catchphrase — partnering with security firm Trail of Bits to deploy what amounts to code EMTs for open source maintainers, using Codex Security to triage and fix vulnerabilities before they become the next log4j-scale catastrophe. The initiative flips the increasingly alarming AI-powered exploit-generation playbook (see: Anthropic’s Mythos) on its head, using the same underlying capability to defend rather than attack.

  More analysis Sensorium Link 12705

• ⚠️🕳️🤖 Latent Darkness: What’s Really Hiding Inside ChatGPT’s Image Generator

  

  Mindgard AI security researcher Dr. Katie Paxton-Fear discovered that a viral “generate a random image” prompt could be trivially modified to bypass ChatGPT’s content filters entirely, producing unsolicited sexual violence and graphic gore imagery — including what appeared to be a bound college student and a bludgeoned dead woman — without those subjects ever being requested.

  More analysis Sensorium Link 12551

• ⚡🏗️🤖 Gigadream: OpenAI eyes $500B Ohio data center backed by Nvidia

  OpenAI is reportedly in talks to lease a 10-gigawatt data centre campus on federal and private land in southern Ohio — potentially the largest of its kind — with SoftBank’s SB Energy developing the site and first operations expected in 2028 under a 20-year lease. The genuinely eyebrow-raising detail: Nvidia isn’t just supplying the hardware, it’s reportedly providing a financial guarantee for OpenAI’s lease, blurring the line between chip vendor and strategic co-investor.

  More analysis Sensorium Link 12533

• 🦁📢💰OpenAI courts advertisers at Cannes ahead of IPO

  

  Well, yuk. OpenAI made its debut at Cannes Lions this week, pitching a nascent ChatGPT advertising business to the world’s top marketers — with former Meta exec Dave Dugan promising “something brand new to the market” while quietly following the oldest playbook in tech: burn cash, get users, sell ads, IPO. The company, which spent US$34bn last year, has reportedly told investors it can grow advertising into a US$100bn business by 2030 — a target that had some senior ad executives raising an eyebrow or two.

  More analysis Sensorium Link 12758

Anthropic

Can’t get away from wall-to-wall Anthropic media coverage… that company’s PR department are complete geniuses, I have to say…

• 🌐🔒⚡ AI Sovereignty: The Rubicon Has Been Crossed

  

  In a watershed editorial, Andrew Ng describes how Anthropic and the US Government each demonstrated — within days of each other — their unilateral power to cut off access to frontier AI: Anthropic restricted its Mythos/Fable models from use in competing LLM development (initially doing so silently, degrading outputs without telling users), while the US Commerce Department imposed export controls requiring licences for any foreign national to access the models, ultimately forcing Anthropic to disable access globally.
  
  The geopolitical fallout is the real story: just as US chip restrictions sent China’s semiconductor ambitions into overdrive, nations worldwide are now urgently treating AI access as a sovereignty issue — and open-source alternatives are looking a lot more attractive when the proprietary tap can be turned off overnight. Will this be the moment that finally tips the scales toward a more open AI ecosystem, or just the opening move in a much messier game?

  More analysis Sensorium Link 12573

• 🐕🤖Project Fetch Phase Two: Claude Opus 4.7 completes robotics tasks 20x faster than humans

  

  Anthropic’s “Project Fetch” experiment — where non-expert employees raced to program an off-the-shelf robotic quadruped to autonomously retrieve a beach ball — has entered its second phase, and the results are striking: Claude Opus 4.7, operating autonomously via Claude Code, completed the same tasks at least 10x faster than any human team, and up to 37x faster than the team working without AI assistance, while writing nearly 10x less code. The model still fumbled the actual “fetching” — closed-loop physical control of a ball remains genuinely hard — but the broader pattern Anthropic is documenting (models assist humans → humans assist models → models operate independently) is now visibly extending from software into the physical world.

  More analysis Sensorium Link 12565

• 🧬🏆🔀 Nobel laureate John Jumper jumps ship

  

  John Jumper — co-winner of the 2024 Nobel Prize in Chemistry for AlphaFold, the landmark AI model that predicts protein 3D structures from genetic sequences — is departing Google DeepMind after nearly nine years to join Anthropic. The same week saw Noam Shazeer (Character AI co-founder) also exit DeepMind, heading to OpenAI instead.

  More analysis Sensorium Link 12577

• 💸🤖⏸️Anthropic pauses Claude Agent SDK pricing hike after backlash

  Anthropic has abruptly paused its planned June 15 billing change that would have shifted Claude Agent SDK usage — including via third-party apps like Zed — from generous subscription caps to full API rates, which one analysis showed could be worth “many multiples” of the monthly subscription cost for heavy Opus users. The reversal comes just weeks after GitHub Copilot triggered widespread sticker shock with its own token-based billing overhaul, and notably, right as Anthropic quietly filed confidential IPO papers with the SEC.

  More analysis Sensorium Link 12598

More reading the tealeaves on the recent Fable ban:

• ⛪⚔️Anthropic’s AI dominance puts it on collision course with Trump

  The Economist chronicles the latest Fable saga: Anthropic has surpassed OpenAI in both valuation (US$965bn vs US$852bn) and annual recurring revenue — with 80% of income from enterprise customers, revenue growing fivefold in five months, and a claim to be the first AI lab to turn an operating profit — all while filing for a ~US$1trn IPO alongside its rival.
  
  But its commercial dominance and safety-first ethos have made it a target: the Trump administration has twice intervened in its operations, most recently the Fable “export ban”.

  More analysis Sensorium Link 12634

  Those who don’t kiss the ring…

• 📢⚠️🔒Anthropic’s AI safety warnings may have triggered US export ban

  An FT analysis found Anthropic uses risk-related language at roughly 8x the rate of OpenAI — and now critics are arguing that relentless safety messaging around its newest models, Mythos and Fable, helped trigger a US government ban on foreign nationals accessing them. Yann LeCun, never one to miss a chance to dunk, summarised it neatly: “One reaps what one sows.”

  More analysis Sensorium Link 12686

• 🔮AI forecasting maps the Claude Fable US ban standoff

  

  The US government ordered Anthropic to geo-block its Claude Fable model from American users — a move whose true motivation remains genuinely unclear across four competing theories, from White House political leverage tied to Anthropic’s impending IPO, to legitimate capability or national security concerns.
  
  LessWrong researcher @dscwarz responded by building a 33-question AI-assisted causal graph forecast, arriving at a median US re-launch date of July 7, broadly consistent with Polymarket’s crowd (>50% by July 1) but grounded in more structured conditional reasoning.

  More analysis Sensorium Link 12653

• 🚫🤖📈Anthropic’s Trump feud may be boosting its business sales

  

  Of course there may be upside… Anthropic is having a genuinely surreal month: it surpassed OpenAI in business subscription market share (41% vs 39.5% per Ramp’s data from 70,000+ companies), filed confidential IPO paperwork off the back of its first-ever profitable quarter, then watched the Trump administration force it to pull its flagship Mythos 5 and its public-facing Fable 5 models entirely — reportedly because hackers cracked Fable 5’s guardrails to access Mythos’ almost comically powerful security-vulnerability-finding capabilities. Here’s the kicker: Ramp’s lead economist Ara Kharazian notes that Anthropic’s best-ever month for business adoption was the month the DoD labelled it a supply-chain risk, suggesting that “too dangerous to use” is, paradoxically, exceptional brand positioning.

  More analysis Sensorium Link 12774

• 💸🖥️Google bets billions on TPUs to challenge Nvidia’s AI dominance

  

  Google is deploying tens of billions of dollars to transform its internal TPU chips — originally built in 2013 just to handle Google Search’s neural network load — into a genuine commercial rival to Nvidia, backing data-centre projects like the US$3.2B Lake Mariner cluster in western New York and a US$7B development near Baton Rouge, Louisiana, with Anthropic as an anchor tenant. The strategy mirrors Nvidia’s own playbook: finance the data centres, then lock in chip sales — and early results are compelling, with Citadel Securities reporting 30% cost savings and 4× speed improvements on TPUs versus Nvidia GPUs.

  More analysis Sensorium Link 12696

• 🧱🤖🔐Google’s Android verification rollout confirmed for September 2027

  

  Nearly two decades after positioning Android as the open antithesis to Apple’s closed ecosystem, Google has confirmed that mandatory developer verification for sideloaded apps will begin rolling out on September 30 in Brazil, Indonesia, Singapore, and Thailand — with global enforcement planned for 2027. A dormant system service (com.google.android.verifier) is being quietly pushed to all Android 8+ devices now, while a deliberately painful “advanced flow” bypass — buried menus, multiple risk confirmations, and a 24-hour waiting period — will be available from August for users who really, truly insist on their freedom.

  More analysis Sensorium Link 12592

• 💸Google slashes AI Plus price, igniting US subscription price war

  

  Google has cut its AI Plus subscription from US$7.99 to US$4.99/month while doubling storage to 400GB, bringing the emerging-market price war logic it tested in India squarely into the US consumer market. Goodwater Capital’s Chi-Hua Chien frames this as the opening of the “commoditisation era” for AI infrastructure — drawing a pointed historical parallel to Cisco, Akamai, and Lucent, whose infrastructure dominance evaporated once users stopped caring whose pipes their bits moved through.

  More analysis Sensorium Link 12747

Microsoft

• 🐋Microsoft eyes DeepSeek V4 for Copilot despite China security risks

  

• 🤖⚖️ Copilot Cowork: What “Azure-hosted” Cannot Actually Fix

  Microsoft launched its Copilot Cowork enterprise AI agent on June 16 with usage-based pricing, simultaneously disclosing it’s evaluating DeepSeek V4 as a cheaper backend — at US$0.87/million tokens versus Claude Opus 4.8’s US$25 — because agentic AI’s token consumption (roughly 1,000× a standard query, per Microsoft Research) has made flat-rate pricing economically untenable, as Uber discovered after burning through its entire 2026 AI budget in four months. Azure hosting solves the data-routing problem neatly, but China’s National Intelligence Law binds DeepSeek as a Chinese legal entity regardless of where inference runs — a distinction Microsoft’s tidy “AI pluralism” framing conveniently sidesteps, and one the Trump administration (which suspended Anthropic’s Fable 5 for non-US users the same week) is unlikely to overlook.

  More analysis Sensorium Link 12607

• 🔍🏔️⚖️ Verification Infrastructure: The New Corporate Moat

  

  Satya Nadella has articulated a sharp test for AI-era competitive advantage: if swapping out your generalist model destroys your value proposition, the model was doing your job — and you’re just a very expensive API wrapper. The real defensible asset, per Nadella (and elaborated by MIT cryptoeconomist Christian Catalini in this forensic Forbes breakdown), is “verification infrastructure” — the accumulated expert judgment, private evals, and institutional memory that turns AI outputs into outcomes a firm will actually put its name and balance sheet behind, compounding as a “hill climbing machine” that gets smarter with every approved or rejected output.

  More analysis Sensorium Link 12667

• 🪪Project Solara: Microsoft wants to give your AI agents an identity card

  

  Microsoft has unveiled Project Solara at Build 2026 — a chip-to-cloud “agent-first” platform pairing small Qualcomm/MediaTek-powered hardware (think smart speaker meets keycard badge) with an Agent Shell runtime and a new in-house MAI model family, including MAI-Thinking-1, a sparse Mixture-of-Experts model with ~35 billion active parameters and a 256,000-token context window. The pitch: ditch app-centric UI flows entirely and let persistent, cloud-backed agents handle cross-app workflows — medical documentation being the headline use case.

  More analysis Sensorium Link 12547

• 🏭⚡🌿 Project Kilby: Microsoft and Chevron’s 2.67 GW Natural Gas Data Centre Deal

  

  Microsoft and Chevron have announced plans to build a 2.67-gigawatt natural gas power plant in West Texas — “Project Kilby” — under a 20-year power purchase agreement to feed Microsoft’s AI and cloud data centres, making it one of the largest co-located gas power and data centre developments in the US. The plant, powered by GE Vernova turbines and Solar Turbines (the same Caterpillar subsidiary supplying xAI’s unpermitted Memphis operation, fun crowd), is projected to emit ~11.8 million metric tonnes of CO₂, ~2,900 metric tonnes of criteria air pollutants, and ~126,000 kg of hazardous air pollutants over its lifetime.

  More analysis Sensorium Link 12711

Apple

• 💾📱💸 More RAMageddon: AI chip shortage may force Apple to raise iPhone prices

  

  AI’s voracious appetite for memory chips has triggered a global DRAM and NAND shortage so severe that outgoing Apple CEO Tim Cook is warning price increases are “unavoidable,” with research firm TechInsights estimating Apple may need to tack on an extra US$270 to the next iPhone Pro — which already starts at US$1,099 — just to preserve margins. The September iPhone launch is shaping up as the likely moment of reckoning, with incoming CEO John Ternus also flagging the issue after chip costs reportedly quadrupled year-on-year.

  More analysis Sensorium Link 12746

Amazon

• 🤖🔁😴 Normalisation of Deviance: Amazon argues humans make poor AI oversight gatekeepers

  

  Amazon Security VP Eric Brandwine has made the case that “human-in-the-loop” AI governance is actually a liability rather than a safeguard — because humans, like AI, are non-deterministic, and repeated approval decisions inevitably lead to alarm fatigue and rubber-stamping, a phenomenon called “normalization of deviance.” Amazon’s alternative is end-to-end accountability, where human identity and ownership thread through entire agentic workflows without requiring humans to approve every step, with agents assigned their own distinct identities in logs (e.g., “this agent did this on behalf of Eric”).

  More analysis Sensorium Link 12775

• 🔲AWS plans to sell Trainium AI chips, challenging Nvidia

  

  Echoing the Google announcement above, AWS is in early talks to sell its Trainium AI chips directly to third-party data centre operators — a move that would transform Amazon from a chip-user into a chip-vendor and put it more squarely in Nvidia’s lane. At a hypothetical US$50B annual run rate, it wouldn’t dent Nvidia’s US$326B juggernaut, but it’s Intel-scale money and a meaningful strategic pivot — complicated by the awkward fact that Trainium capacity is already sold out, and elbowing into TSMC’s queue means competing with the very company (Nvidia) you’re trying to displace.

  More analysis Sensorium Link 12659

• 🧠📦☁️ Managed Knowledge Base: AWS turns RAG into a utility

  Amazon Web Services has quietly made one of enterprise AI’s most painful engineering problems disappear, with its Bedrock Managed Knowledge Base service now generally available — handling vector storage, data ingestion, and retrieval infrastructure so developers don’t have to. The service ships with six native connectors (S3, SharePoint, Confluence, Google Drive, OneDrive, and a web crawler), hybrid search, and “agentic retrieval” that auto-orchestrates multi-hop query planning, and is available across regions including Sydney.

  More analysis Sensorium Link 12559

xAI

• 🫧💥LeCun calls xAI a failure, warns of AI bubble burst

  

  AI Godfather Yann LeCun has called xAI “kind of a failure” due to co-founder departures, arguing Musk’s conduct toward his previous team will make it near-impossible to recruit top talent — while the Colossus data centres in Memphis are increasingly rented out to Google and Anthropic just to recoup costs. More broadly, LeCun warns the whole industry faces a “bubble explosion” as investor-subsidised services can’t sustain the widening gap between falling-but-not-fast-enough compute costs and what users will actually pay, predicting OpenAI and Anthropic will eventually have to raise prices, slash costs, or implode.

  More analysis Sensorium Link 12599

• ⚔️🤖💨 Operation Epic Fury: Trump DOJ backs xAI in Clean Air Act pollution lawsuit

  

  More on this story from last week: the Trump administration’s DOJ has filed to dismiss an NAACP Clean Air Act lawsuit against Elon Musk’s xAI, arguing that the company’s unpermitted gas turbines — 57 of them, powering the Memphis-area “Colossus” data centre — can’t be shut down because Grok is now a military asset; a court declaration reveals the Grok Gov Model helped US forces deploy 2,000 munitions to 2,000 targets within 96 hours during “Operation Epic Fury” strikes on Iran. The Southern Environmental Law Center, representing the NAACP, points out that the DOJ never actually disputes xAI is pumping out unlawful pollution — ten times the permitted nitrogen oxide levels — into predominantly Black communities in North Mississippi; it just argues that the Trump administration’s blessing makes legality moot.

  More analysis Sensorium Link 12642

Tesla

• 🚗💥Tesla Autopilot crash kills elderly woman in her home

  

  A 76-year-old Texas woman, Martha Avila, was killed after a Tesla Model 3 driver told police he was relying on Autopilot when he lost control and ploughed into her family’s home at high speed — with NHTSA now launching a special crash investigation. This lands at the worst possible moment for Tesla: the company is actively lobbying a Trump-aligned NHTSA (whose AV safety expertise was gutted by Musk’s own DOGE operation) to strip human controls from autonomous vehicles, while its own marketing cheerfully shows drivers waving their hands in the air and sipping coffee behind the wheel.

  More analysis Sensorium Link 12682

ASML

• 🔬🇳🇱⚠️ ASML Under Fire: US Accuses Dutch Chip Giant of Smuggling EUV Tech to China

  

  The US Commerce Department, led by Secretary Howard Lutnick, has accused ASML — the sole manufacturer of extreme ultraviolet lithography machines essential to cutting-edge chip production — of shipping EUV-compatible components and transport equipment to China, potentially one of the most serious export control breaches in semiconductor history. ASML has entered “crisis mode” but categorically denies every element of the allegation, stating that all 314 EUV systems worldwide are accounted for via remote telemetry, and that no components specially designed for EUV machines have ever been shipped to China. The company circulated a formal rebuttal document in Washington, but US officials claim to possess documentary evidence they’ve declined to share publicly.
  
  The standoff carries enormous strategic weight. Under Bureau of Industry and Security rules, even component-level exports — not just complete 180-metric-tonne machines — can trigger enforcement action, as the US$252.5 million Applied Materials settlement demonstrated earlier in 2026. Meanwhile, the bipartisan MATCH Act moving through Congress threatens to ban even ASML’s permitted older DUV tool sales to China (~20% of 2026 revenue), and China’s own prototype EUV machine reportedly achieved light generation in late 2025. Critics note a potential conflict of interest: Lutnick’s Commerce Department has invested up to US$150 million in xLight, a startup developing technology that could challenge ASML’s monopoly. Whether this dispute reflects genuine intelligence or geopolitical leverage, it’s straining US-EU relations at a moment when the global AI chip supply chain has never been more consequential.

  
  Coverage:

  • US accuses ASML of smuggling EUV chip tool to China

  • US cites evidence of ASML EUV components reaching China

  • US claims ASML’s secret EUV chip tool reached China

Huawei

• 🇨🇳💾Huawei defies US chip sanctions with breakthrough stacking technology

  

  Seven years after US sanctions were supposed to be a death sentence, Huawei’s HiSilicon unit — led by the quietly formidable He Tingbo, aka China’s “chip queen” — has unveiled a logic-stacking technology that folds chip circuits into multiple layers to boost computing performance without needing ASML’s EUV lithography machines, potentially shattering the assumption that export controls had hard-capped China’s semiconductor progress at 7nm. Bernstein analysts are already calling it another “DeepSeek moment,” and Huawei’s CloudMatrix 384 platform — which daisy-chains hundreds of AI processors into a single cluster — is claiming to outperform Nvidia’s NVL72 in raw compute and memory.

  More analysis Sensorium Link 12661

• 🏆🇨🇳🔓 The Chip Control Paradox: China’s GLM-5.2 tops AI rankings using only Huawei chips

  

  While the US Commerce Department’s June 12 export ban on Anthropic’s Fable 5 left developers worldwide scrambling, Beijing’s Z.ai dropped GLM-5.2 — a 744-billion-parameter open-weight model trained entirely on 100,000 Huawei Ascend 910B chips — which promptly topped Artificial Analysis’s Intelligence Index and Code Arena among available models, with a US$25M training run that cost a fraction of comparable American efforts. The overall US-China frontier capability gap has collapsed to 2.7 percentage points per Stanford’s 2026 AI Index, though meaningful gaps persist on fluid-reasoning benchmarks like ARC-AGI-2, where the best Chinese model sits at 11.8% versus substantially higher US scores.

  More analysis Sensorium Link 12699

• 📡💰🔑 Huawei’s Wi-Fi 7 Toll Booth: sets patent licensing at US$0.5 per unit

  

  Huawei has announced it will license its Wi-Fi 7 standard-essential patents at US$0.5 per unit, staking its claim in the royalty stream of the next-generation wireless standard before the market fully takes off. It’s a classically elegant move: even as Western regulators have spent years trying to keep Huawei hardware out of critical infrastructure, the company’s patent portfolio quietly ensures a cheque arrives for every router, laptop, and smartphone that ships with Wi-Fi 7 inside.

  More analysis Sensorium Link 12740

AliBaba

• 🧬⚗️🔬 LOGOS: Alibaba’s tiny AI model outperforms Microsoft’s giant on science

  

  Alibaba’s new open-source scientific foundation model encodes proteins, molecules, crystal materials, and chemical reactions into a single shared token vocabulary — and its 1-billion-parameter variant reportedly outperformed Microsoft Research’s NatureLM, a model 56 times its size, on several core scientific benchmarks. The architectural trick is elegant: LOGOS’s pre-training format is identical to its inference format, eliminating the fine-tuning bottleneck that plagues every other multi-domain scientific AI.

  More analysis Sensorium Link 12698

DeepSeek

• 🐉💰🚀 DeepSeek Secures US$7.4B in First-Ever Funding Round, Becomes China’s Most Valuable AI Startup

  

  Hangzhou-based DeepSeek has closed a landmark first external funding round of over RMB 50 billion (US$7.4 billion), catapulting the AI lab to a valuation exceeding US$50 billion and cementing its status as China’s most valuable AI startup. Founder Liang Wenfeng personally contributed roughly US$3 billion — the largest single investment — while structuring the raise through a limited partnership that preserves his control. Strategic backers include Tencent, JD.com, CATL, NetEase, and China’s government-backed National AI Industry Investment Fund, signalling broad confidence across both the private sector and Beijing.
  
  The round underscores DeepSeek’s remarkable trajectory since its open-source R1 reasoning model rattled global markets in January 2025, briefly wiping 15% off Nvidia’s market cap. Its latest release, DeepSeek-V4-Pro — a 1.6 trillion-parameter mixture-of-experts model with dramatically reduced memory requirements — continues to push the cost-efficiency frontier. That efficiency is already attracting Western giants: Microsoft is reportedly preparing to integrate a fine-tuned DeepSeek model into its Cowork Copilot as a cheaper alternative to OpenAI and Anthropic. Still, context matters — US officials assess DeepSeek’s best models as roughly eight months behind top American offerings, and its valuation remains a fraction of Anthropic’s US$965 billion or OpenAI’s US$852 billion. The funding nonetheless marks a pivotal moment in the global AI race, proving that algorithmic efficiency and open-source strategy can command serious capital even under tightening US chip export controls.

  
  Coverage:

  • DeepSeek raises $7.4B, valued at over $50B

  • DeepSeek secures $7B in its first external funding round

  • DeepSeek reaches $50 billion valuation in first funding round

• 🀄US developers flock to Chinese AI models for cheaper costs

  

  US developers are quietly defecting to Chinese AI models — DeepSeek, Minimax, Kimi, Xiaomi MiMo — with one San Diego dev paying under 50 cents for a coding session that would cost US$10 on Claude, and San Francisco startup Lindy claiming millions in savings after ditching Anthropic. The catch: DeepSeek hit 17% of token usage on Vercel in May while holding just ~1% of revenue, as political scrutiny, data security concerns, and the “accessed via American cloud” workaround all conspire to keep Chinese developers from actually getting paid.

  More analysis Sensorium Link 12618

Zhipu (Z.ai)

• 🇨🇳🤖📈Zhipu AI hits trillion yuan valuation — what’s fueling the surge?

  

  Beijing-based Zhipu AI — the Tsinghua University spinout behind the GLM model series — has crossed the trillion yuan (≈US$138 billion) market cap threshold, a milestone that cements its status as China’s answer to OpenAI. The surge appears driven by a potent cocktail of domestic enterprise adoption, state-backed tailwinds, and a geopolitical tech-decoupling narrative that makes homegrown AI infrastructure a strategic imperative rather than just a nice-to-have.

  More analysis Sensorium Link 12675

Wikimedia

• 📖🚫Wikipedia cofounder bans AI from editing articles directly

  

  Wikipedia co-founder Jimmy Wales has confirmed the online encyclopaedia won’t allow AI to directly edit its articles, citing hallucinations that remain “very, very bad” despite recent model improvements — ironic given that AI platforms lean heavily on Wikipedia’s content to answer user queries. Human traffic to the site has dropped 8% as bot traffic surges, and Wales is pushing AI companies to “pay their fair share” for the server costs of being hammered with millions of scraping requests, with Wikipedia already blocking those who aren’t playing ball.

  More analysis Sensorium Link 12718

Mozilla

• 🦊📉Firefox unveils ambitious roadmap amid steep user decline

  

  Mozilla has published a sweeping new roadmap for Firefox alongside its 152nd release, promising a Nova design refresh, customisable keyboard shortcuts, built-in mobile VPN, iOS tracker blocking, PDF editing tools, voice-activated AI “Quick Answers,” JPEG XL support, and — after six years of user requests — actual HDR video playback on Windows and Linux. The catch: Firefox’s desktop market share has cratered from 5.88% to 3.79% in just twelve months, shedding millions of users per month while rolling out features that, frankly, Chrome users have been ignoring for years.

  More analysis Sensorium Link 12749

Snap

• 👓📉Specs: Snap’s $2,200 AR glasses send stock tumbling

  

  As covered last week Snap finally launched its long-awaited AR glasses, “Specs,” priced at a cool US$2,200 — a figure that raised eyebrows given that Snap’s core user base is, famously, teenagers. The stock slid more than 5% post-announcement, dropping from US$5.86 to a low of US$4.83, compounding an already grim 30% decline over the past year.

  Expect a firesale of IP. No great loss.

  More analysis Sensorium Link 12748

SpaceX

In addition to losing a trillion dollars in value in only a few days…

• 🚀🛰️SpaceX has now launched more satellites than all others combined

  

  As of 12 June, SpaceX had launched 15,262 satellites versus 15,138 for every other company and nation combined since the dawn of the space age in 1957 — a lead that will only widen, given the Falcon 9 flew a staggering 165 times in 2025 alone and Starlink could eventually field 40,000+ spacecraft. The company that nearly died after three consecutive Falcon 1 failures in the mid-2000s is now pivoting from launch provider to AI infrastructure play, with Musk eyeing a million orbital data centres powered by the still-in-testing Starship megarocket.

  More analysis Sensorium Link 12689

• 📉🏛️MSCI gives SpaceX lowest possible ESG rating at IPO

  

  MSCI handed SpaceX its lowest possible CCC ESG rating — the same score given to Russia post-Ukraine invasion — just one day before the company’s record US$75bn IPO, citing poor governance, a 3.2/10 governance metrics score, and an “orange flag” controversies rating. Concerns centre on SpaceX’s concentrated insider control, limited shareholder rights, conflicts of interest, and lack of board independence — issues that could restrict access to the company’s stock for European funds managing over €6.5tn.

  More analysis Sensorium Link 12640

• SpaceX loses $400bn in market value after IPO rally reverses

  

  As illustrated earlier…

  SpaceX shed US$400bn in market capitalisation in a single session — the second-largest one-day valuation loss in corporate history — as rising Treasury yields punished the stock’s eye-watering 100x revenue multiple, leaving shares 31.5% below their post-IPO peak. The selloff lays bare the central tension in SpaceX’s investment thesis: the company’s stratospheric valuation is essentially a leveraged bet that its AI division — which lost US$6.4bn in 2025 — can capture a slice of a US$26.5tn total addressable market, while Grok continues to trail ChatGPT, Gemini, and Claude in actual adoption.

  More analysis Sensorium Link 12710

• 🛸📦SpaceX’s secret Starfall cargo pod set for first test flight

  

  SpaceX is launching its first demonstration of “Starfall” — a disc-shaped reentry vehicle (3.1m wide, 0.75m tall, carrying up to 1 metric tonne of cargo) designed to deliver goods anywhere on Earth from low-Earth orbit via parachute splashdown — developed under near-total secrecy and only revealed through an FAA environmental assessment. The Falcon 9 mission from Cape Canaveral will orbit twice before releasing the pod for a Pacific splashdown roughly 1,300 km west of California.

  More analysis Sensorium Link 12733

Startups and Investment Deals

• 🎮🏗️💸 The Abstraction Layer Thesis: AI inference and world model startups raise $1.8B in 48 hours

  In a single news cycle on 18 June, venture capital delivered a fairly unambiguous verdict on where AI value is accruing: inference infrastructure player Baseten closed a US$1.5B round valuing it at up to US$13B (up from US$2.15B just nine months ago), while world-model startup General Intuition — built on a gaming clip dataset OpenAI reportedly offered US$500M to buy — is in talks for US$300M at a US$2B valuation. Neither company builds a foundation model; both are betting that as GPT-class models commoditise, the durable margin sits in the orchestration and training-data layers around them — the same structural shift that made AWS worth more than the commodity hardware it abstracted over.

  More analysis Sensorium Link 12606

• 🔢📉⚡ Tensordyne: Logarithmic Number Systems for AI Inference

  

  Silicon Valley startup Tensordyne is claiming a dramatic leap in AI inference efficiency by replacing conventional floating-point arithmetic with logarithmic number systems (LNS) — where multiplications become mere additions — potentially slashing the compute cost of running large models. IEEE Spectrum’s framing of the claim as “bold” is doing a lot of quiet editorial work, suggesting the engineering community is watching with *calibrated scepticism*.

  More analysis Sensorium Link 12540

• 🇨🇳🖥️📈Moore Threads revenue surges amid China’s GPU market boom

  Beijing-based GPU startup Moore Threads is posting surging revenues as US export restrictions on Nvidia and AMD chips continue to turbocharge demand for domestic Chinese alternatives. Founded in 2020 by former Nvidia China chief Jeff Ren, Moore Threads is emerging as a serious beneficiary of the chip war — proof that sanctions can function as the world’s most aggressive industrial policy.

  More analysis Sensorium Link 12644

• ⚡🧠💰 Groq’s US$650M Second Act: Selling Your IP to Nvidia and Then Competing With Them

  

  Inference chip startup Groq has raised US$650M from growth firm Disruptive and hedge fund Infinitum to scale its LPU-powered cloud platform, which is already processing trillions of tokens per week for 5 million developers across 13 data centres globally. This comes seven months after Groq inked a US$20B chip licensing deal with Nvidia — which also poached Groq’s founding CEO — producing the Nvidia Groq LPU 3, a liquid-cooled inference appliance boasting 2.5 terabits per second of bidirectional bandwidth and 500MB of onboard SRAM.

  More analysis Sensorium Link 12708

⏳ Zeitgeist

OK that’s enough tech!!

What’s happening in the “real” world?

Climate

• 🌊🛰️🌡️ Kelvin Waves: NASA satellites confirm El Niño has returned and is strengthening

  

  NOAA officially declared El Niño on June 11, 2026, with NASA’s Sentinel-6 Michael Freilich satellite tracking ominous sea surface height anomalies across the central and eastern Pacific — a smarter metric than surface temperature alone, since it captures the subsurface heat reservoir that actually drives long-term climate disruption. Western Pacific conditions are already tracking close to June 1997, the precursor to one of the strongest El Niños on record, though the eastern Pacific is lagging — for now, with more warm Kelvin waves still inbound.

  More analysis Sensorium Link 12637

• 🌡️🥵Record-breaking heat wave scorches Europe with no relief in sight

  

  Less than a month after a record-shattering May heat wave, Europe is being hammered again — France has placed half its mainland departments on red alert (and banned alcohol at its beloved Fête de la Musique, which is a special kind of suffering), Spain hit 44°C forecasts and cancelled a public World Cup screening while Belgium braces for its hottest temperatures ever recorded. The UK’s Met Office is warning of widespread “tropical nights” where temperatures won’t drop below 20°C and issuing a rare red warning covering London and Birmingham (38–40°C expected), with the Royal Meteorological Society noting May and June will represent two consecutive months in which UK temperature records were “annihilated by well over 2°C” — a phrase that should probably be in every climate briefing from here on out.

  France is forecasting temperatures up to 43°C in Bordeaux, , and Belgium bracing for its hottest temperatures ever recorded — while 13 people have already drowned in France attempting to cool off in unsupervised waterways. The acceleration is the real story: of 51 heat waves recorded in France since 1947, 34 have occurred since 2000 and 26 since 2011 — a curve that looks less like a trend and more like a cliff edge.

  More analysis Sensorium Link 12677 | Sensorium Link 12701

• 🌧️🚀🇨🇳 Tianhe: China’s ambitious “Sky River” cloud seeding geoengineering project quietly collapses

  

  China’s audacious 2018 “Sky River” (Tianhe) project aimed to deploy tens of thousands of satellite-triggered silver iodide burners across a 1.6 million km² Tibetan Plateau corridor to redirect Indian monsoon moisture toward the water-scarce Yellow River basin — a plan scientists immediately labelled “absurd and fantastical with neither scientific basis nor technological feasibility.” The project has apparently been quietly shelved, conspicuously absent from China’s new five-year plan, though Beijing remains the world’s dominant cloud seeding power with 50,000 employees and billions of dollars invested.

  More analysis Sensorium Link 12772

• 🧊⚠️Geoengineering’s “termination shock” poses existential climate risk, scientists warn

  

  Four of the world’s most credentialled climate scientists have published a stark warning: solar geoengineering startups flush with venture capital are racing toward planetary-scale deployment with zero governance, zero scientific consensus, and a terrifying exit problem — once you start blocking sunlight, stopping abruptly could trigger “termination shock,” a climate whiplash so severe it could be more destabilising than the warming it was meant to fix. Meanwhile, models can’t even agree on outcomes, with the same stratospheric aerosol injection producing anywhere from less than 1°C to 3°C of cooling after a decade — which, as the scientists drily note, means we’d essentially be flying blind.

  More analysis Sensorium Link 12639

Biodiversity

• 🌿🌡️💀 The Extinction Flip: Temperate Species Are Disappearing Faster Than Tropical Ones

  

  A University of Arizona study published in Nature Climate Change — the largest of its kind, spanning 5,100+ species across nearly 40,000 global sites — has flipped decades of scientific consensus: 49% of temperate species have experienced local extinctions at the hottest parts of their ranges, versus just 33% of tropical species. The culprit is straightforward but alarming: temperate regions are warming at up to 3.3°C over 25 years, nearly double the 1.8°C maximum seen in the tropics. Awkwardly, the study’s own senior author published the opposite finding in 2016 — so the world really has changed that fast.

  More analysis Sensorium Link 12785

Health

• 🦠🛂🔬 Bundibugyo: Africa CDC mandates exit screening as Ebola outbreak hits 837 cases

  

  The largest Bundibugyo Ebola outbreak on record — 837+ confirmed cases, 196 deaths in DRC, now spilling into Uganda — has exposed a structural double failure: standard Ebola rapid diagnostic tests cannot detect this strain (its genome diverges ~30% from the Zaire strain tests are calibrated for), meaning the outbreak spread undetected for nearly three weeks before confirmation, while blanket travel bans imposed by the US, Canada, and Uganda are actively pushing exposed travellers onto unmonitored informal border crossings, per a June Lancet analysis. Africa CDC has now mandated exit screening across all AU member states’ official crossings while explicitly warning against the travel bans — a direct rebuke of current US Title 42 policy — but with contact tracing reaching only 45% of exposed contacts against a 90% WHO threshold and a US$21.5 million funding gap in the response plan, the formal screening architecture is being asked to carry weight it isn’t yet built to bear.

  More analysis Sensorium Link 12736

• 🦅🦠🌏 The Last Continent: H5N1 Finally Reaches Australia

  

  Two sick seabirds — a brown skua and a giant petrel — found on a remote beach near Esperance, Western Australia, have confirmed what biosecurity experts long feared: HPAI H5N1 clade 2.3.4.4b has arrived on the last H5N1-free continent, almost certainly carried via Southern Ocean wildlife movements from South America rather than the northern migratory routes. No farm cases have been detected yet, but the virus now threatens iconic and critically endangered species including Tasmanian devils, shy albatrosses, orange-bellied parrots, and black swans — for some of which even a handful of deaths could tip entire populations toward extinction.

  More analysis Sensorium Link 12766

Geopolitics

• ☀️🛡️⚡ Geopolitically Inert: Energy independence now drives solar adoption more than cost savings

  

  Solar energy has entered a “third era” of adoption — no longer driven by environmental values or economics alone, but by a deep psychological need for control: rooftop panels, home batteries, and EVs are now acts of self-determination in a world of volatile gas prices, grid blackouts, and Middle East conflict. Columbia Climate School dean Alexis Abramson, writing in TIME, notes that 78% of US homeowners are already worried about grid reliability, and the failed US assault on Iran has turbocharged that anxiety into action — because, as atmospheric scientist Andrew Dessler puts it, a missile strike on Iranian oil infrastructure has precisely zero effect on the cost of sunlight in Texas.

  More analysis Sensorium Link 12778

• 🏯China’s ancient tribute system is reshaping the new world order

  

  Bridgewater’s Ray Dalio, writing in the FT after a month touring Asian capitals, argues a fundamental shift in world order is underway — driven by US credibility damage following Iran’s seizure of the Strait of Hormuz, and China’s swelling export surpluses pushing the renminbi’s global role upward. Dalio frames China’s strategy through the lens of the ancient Confucian tribute system: not empire-building through military conquest, but hierarchical influence through pressure, patience, and the ever-present threat of economic pain — with Taiwan’s chip supply as the ultimate leverage point, given China’s projected semiconductor self-sufficiency by late 2028.

  More analysis Sensorium Link 12761

• 🔬⚛️🇨🇳 Science City: China’s Huairou scientific facilities attract global researchers

  

  China’s Huairou Science City north of Beijing is switching on a suite of “major scientific facilities” — including the High Energy Photon Source (HEPS), one of the world’s brightest synchrotron light sources — and global researchers are already lining up for beam time, geopolitics be damned. The facilities represent a generational bet by Beijing on hard scientific infrastructure as a vector of both discovery and soft power.

  More analysis Sensorium Link 12726

🧠Mind expanding

A few deeper thinks tickling my neurons this week…

• 🧠⚗️😊 Hedonium: Engineering Happiness From First Principles

  A LessWrong post from Jack Thompson argues that philosophers who subscribe to “illusionism” — the view that consciousness just is physical/functional states, with no mysterious phenomenal ectoplasm required — have a unique obligation to actually try building “hedonium”: a minimally conscious system optimised purely for experiencing happiness. This isn’t merely academic; with AI alignment decisions being locked in right now, we urgently need mechanistic blueprints for valenced experience before those decisions inadvertently create systems that can suffer but can’t tell us.
  

  ”“Hedonium” is a theoretical, minimally conscious substance optimized for experiencing happiness. Imagine a mind pared down to the bare essentials required for having happy experiences, instantiated as cheaply as possible.

  Nobody has built hedonium yet. Nobody has tried to build hedonium yet. Nobody has even laid out the blueprint for how you would try to build hedonium yet.

  It doesn’t seem impossible. 55% of philosophers of mind are physicalists: they believe that mental states just are physical states. Even non-physicalist philosophers of mind often believe that the mental and physical are tightly bound up—creating new brains, at the very least, creates new subjects of experience.”

  More analysis Sensorium Link 12651

• 🧠❄️Cryonics in the age of AI: why preservation still makes sense

  Another LessWrong essay, this one by @raelifin, commissioned by brain-preservation startup Nectome argues that cryopreservation makes strategic sense across all plausible AI futures — utopia, apocalypse, slowdown, or messy muddle — because the uncertainty itself is the point: a plan doesn’t need to be perfect, it needs to keep you pointed in the right direction.
  
  The piece also floats a genuinely underrated argument: normalising brain preservation reduces the pressure to race toward superintelligence, offering a “wholesome alternative” for those who want to survive long enough to see the future without betting everything on AI getting there first.

  More analysis Sensorium Link 12680

• 🔭💡🛰️ Observable Space: Former SpaceX engineer builds telescope empire for space laser comms

  

  Dan Roelker — former hacker, DARPA cyberwarfare programme manager, SpaceX VP of software, and OpenSea early employee — has co-founded Observable Space, now the largest US-based telescope manufacturer, which just landed a US$94 million Space Force contract and closed a US$90 million Series A to scale up laser communications infrastructure. The core bet: as orbital data centres proliferate and satellite constellations balloon past 15,000 active objects in low-Earth orbit, whoever controls the ground-based optical infrastructure to receive laser-encoded data from space (10–100× the bandwidth of radio, but blocked by clouds) wins the new space economy.

  More analysis Sensorium Link 12688

💭Meme stream

And now for some light relief, you’ve earned it!

• 🖐️⚡🤖 CrankGPT

  Squeez Labs has built CrankGPT, a fully offline LLM voice assistant running on a Raspberry Pi 5 powered entirely by a hand-crank generator — complete with a custom capacitor board to smooth out the voltage brownouts caused by inference spikes hitting up to 15W. The software stack (Moonshine ASR, Silero VAD, llama.cpp with Gemma 3 1B or Liquid AI’s LFM2, and Piper TTS) achieves sub-one-second response latency, which is genuinely impressive for an air-gapped single-board computer whose power source is vibes and biceps.

  More analysis Sensorium Link 12566

• 📊🥝🏆 The Rousseeuw Prize: R programming language creators win $1M “Nobel for statisticians”

  

  The R programming language — born in the University of Auckland’s statistics department in the early 1990s and famously described by co-creator Ross Ihaka as “something started by a couple of amateurs in the outer provinces” — has won the Rousseeuw Prize in Statistics, a US$1 million award designed to be the Nobel Prize of the stats world. Three University of Auckland researchers (Thomas Lumley, Simon Urbanek, and Paul Murrell) share in the prize alongside five principal laureates from Oxford, ETH Zürich, Vienna, Copenhagen, and Iowa.

  More analysis Sensorium Link 12738

• Kinetic Assemblage

  Nemo Gould is a kinetic assemblage sculptor who turns found objects, discarded tools, machine parts, trophies, toys, and trash into strange mechanical creatures that light up, move, and feel alive. In a world increasingly flooded with AI-generated images, Nemo’s work can look almost impossible, but every piece is built by hand.
  

• AI Fails

  The latest AI fails…

  

• 🫴Scratching hand

  WTF.😂 Click through to LinkedIn to watch this video… weekly weirdness in my feed.🫣

  

• Robot muralist

  An Estonian street muralist called Mika built his own mural painting robot over 10 years… and the results are spectacular:

  
  (Source: @itsolelehmann)

🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback.

Cheers!

Ben

Welcome to Memia #2026.24… another full-spectrum scan across the frontier of AI, emerging tech and the accelerating future. As always, thanks for reading!🙏

(Once again running behind on publishing schedule… an insane week of news volume combined with real-world work and travel, sorry!)

🗞️Weekly roundup

The most clicked link in last week’s newsletter was my recent note on AI Agent architecture:

Last week’s leading story about the seemingly imminent El Nino pattern continues to develop at pace… and yet you’ll hardly see it anywhere in mainstream media.

🤖Being Agentic

I’m now hooked into the thriving AoNZ chapter of Agentics Foundation… an online community bringing together local AI agent developers. The WhatsApp channel has some of the highest signal I’ve come across locally in a long time: if, like me, you’re on a vertical learning curve doing AI Agent development, this is definitely one of the places to hang out...

(🎩thanks Mike Hall for the pointer and Chris Barlow for convening the group!)

📅Upcoming

A few dates for the diary if you’re around Aotearoa over the next few months:

• AI Forum Te Kāhui Atamai Iahiko o Aotearoa is organising another Aotearoa AI AI Hackathon on 3-10 August at venues all around the Motu:

  

• I just booked my ticket for Agentic Engineering Day New Zealand in Tāmaki Makaurau Auckland 22 October (also with a day of workshops planned on 21st).

  

• Also, I’ll be giving a keynote presentation on How Exponential AI Is Shaping the Future of International Development at this year’s Council for International Development annual conference in Te Whanganui-a-Tara Wellington September 10th.

  

  Reminder: Book me for tailored keynotes/talks designed for audiences needing a clear read on the AI landscape: what changed, what matters, and what comes next.

🌟Picks of the week

📖A Fable of power and control

Just when you thought this AI hype couldn’t get any crazier…

• 🦋🔒🧬 Claude Fable 5 launch recap

  

  Last week I briefly reported on Anthropic’s then-breaking release of Claude Fable 5, its first “Mythos-class” model available to the general public. A significant step up in AI capability across software engineering, scientific research, vision, and long-horizon agentic tasks. Stripe reported it compressed months of engineering into days.

  
  While I still had access, I set it on a complete re-architecture of the Sensorium codebase (together with “strangler fig migration strategy”) ready for the next round of features and it aced it. Mindblowing - absolutely months of work completed overnight…but I’m still the bottleneck!

  
  Rsearcher Ethan Mollick described working with it as shifting from “wizard” to “patron” — commissioning work rather than steering it, as the model autonomously spun up sub-agents, conducted research, and executed multi-page specifications over nine-hour sessions.

  
  Meanwhile, the still-restricted Claude Mythos 5 variant, available only to vetted cyberdefenders via Project Glasswing, scored 78% on ExploitBench (up from Opus 4.8’s 40%) and produced novel scientific hypotheses preferred by researchers 80% of the time.
  
  Interestingly, Fable 5 silently falls back to the older Opus 4.8 model for cybersecurity, biology/chemistry, and distillation-related queries — a trade-off Anthropic acknowledges is “stricter than ideal.” More controversially, the system card reveals hidden safeguards that quietly degrade the model’s effectiveness on frontier AI development tasks (pretraining pipelines, distributed training infrastructure) without notifying users (more on this further below).
  
  Nathan Lambert’s sharp critique frames this as competitive entrenchment disguised as safety policy, arguing that silently manipulating model intelligence is itself a form of misalignment.
  
  At US$10/million input tokens and US$50/million output tokens — roughly double OpenAI’s GPT-5.5 — Fable 5 is (was) both the most powerful and most contentiously controlled AI model yet released.

  
  Early coverage:

  • Claude Fable 5 and Claude Mythos 5 \ Anthropic

  • What it feels like to work with Mythos - by Ethan Mollick

  • Claude Fable 5 and new safety fables - by Nathan Lambert

  • Anthropic’s Fable 5 can make weirdly fun video games with the click of a button | TechCrunch

  • Anthropic says these topics are too dangerous to let its Fable 5 model talk about - Ars Technica

  • Claude Fable 5 and Mythos 5: The System Card

  • Anthropic’s Claude Fable 5 is a version of Mythos the public can access today | TechCrunch

  • Claude Fable won’t answer basic biology questions | The Verge

  Pliny claimed to have jailbroken Fable’s system prompt within 48 hours of release:

• ✈️🤖⚖️ Policy on the AI Exponential: Anthropic CEO urges FAA-style regulation for powerful AI models

  

  And then simultaneously, Anthropic CEO Dario Amodei published a sweeping policy essay arguing that AI has crossed a threshold — that demands urgent, binding regulation modelled on the FAA, alongside coordinated democratic coalitions, labour displacement frameworks, biomedical regulatory reform, and new civil liberties protections against AI-powered autocracy.
  
  Crucially, Amodei is explicitly shifting Anthropic’s position from “transparency legislation is enough” to “we need mandatory third-party testing with government power to block model deployments” — a significant escalation from the company that previously agonised publicly over California’s SB 1047.

  
  Direct quote:

  “…It is time to go beyond transparency to more serious and binding regulation of AI… Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. I am grateful to see the Trump administration’s Executive Order move incrementally towards a greater role for government in AI, though Anthropic’s proposal recommends even further action…“

  Oops. Be careful what you wish for, I guess…

• 🔒🤖💥 The Fable Farce: Trump Admin Forces Anthropic’s Most Powerful AI Models Offline

  

  …And then, only three days after release (Friday 13 June 2026, 5:20pm in the US… just after markets closed), the Trump administration issued an unprecedented export control directive ordering Anthropic to block all foreign nationals — including its own employees — from accessing its frontier Fable 5 and Mythos 5 models, citing national security concerns over an alleged “jailbreak”.

  
  Given the impossibility of verifying citizenship in real time, Anthropic had no choice but to pull both models entirely, just three days after Fable’s public launch.
  
  The trigger appears to have been Amazon CEO Andy Jassy raising concerns with White House officials about security research showing Fable could be prompted to identify cyber vulnerabilities — capabilities Anthropic and independent researchers argue are already present in rival models like GPT-5.5. Former AI czar David Sacks claimed Anthropic CEO Dario Amodei refused to patch the issue, while Anthropic maintained the jailbreak was narrow, non-universal, and that the government provided only verbal evidence with 90 minutes to comply.
  
  The fallout has been seismic on multiple fronts:

  • Domestically in the US, the move sets a chilling precedent — the US government effectively killed a commercial product serving hundreds of millions of users via an opaque, unilateral action with no court oversight.

  • Geopolitically, it’s a gift to Chinese AI competitors like DeepSeek and MiniMax, whose open-source models can’t be switched off by any single government. Allied nations including Five Eyes partners were cut off overnight, prompting France’s Gabriel Attal to declare the start of “the AI war” and Canada’s Mark Carney to warn against over-reliance on any single AI provider.

  • The EU and US are now discussing a “trusted partner” scheme at the G7 summit, but the damage to trust in American AI infrastructure may prove lasting — enterprises worldwide are urgently reassessing single-vendor dependencies, and the sovereign AI movement just got its most compelling argument yet.

  
  Huge amounts of media coverage and analysis:

  • Statement on the US government directive to suspend access to Fable 5 and Mythos 5 \ Anthropic

  • Anthropic cuts off Fable 5 and Mythos 5 access following government order | The Verge

  • American Government Takes Down Claude Fable

  • Anthropic blocks all public access to Claude Fable 5, Mythos 5 following US government order — what enterprises should do | VentureBeat

  • Anthropic’s safety warnings may have just backfired — the government has pulled the plug on its most powerful AI | TechCrunch

  • The Anthropic Fable Farce - by Ben Goertzel - Eurykosmotron

  • Anthropic scrambles after Trump administration freezes its top AI models

  • Anthropic suspends latest AI models after US blocks access to foreigners

  • As Anthropic suspends access to new models, India debates its AI future | TechCrunch

  • White House move to limit Anthropic linked to concerns about Chinese access to Mythos | Semafor

  • Amazon CEO reportedly raised Anthropic model concerns before government crackdown | TechCrunch

  • Amazon security research reportedly led to the White House’s Anthropic Fable ban | The Verge

  • Tech Things: There is a massive shadow hanging over this Fable thing

  • Anthropic shuts down Fable, Mythos models following Trump admin directive - Ars Technica

  • Donald Trump’s blocking of Anthropic is capricious and chaotic

  • Cutting access to Anthropic’s Mythos is a gift to China

  • The White House’s capricious controls on Anthropic

  • US and Europe discuss access to AI models after Anthropic dispute

  • Trump’s Anthropic shutdown just made the case for non-American AI | The Verge

  • The US government’s Anthropic models ban was never about an AI jailbreak | TechCrunch

  • Donald Trump has cut off access to the world’s best AI model

  • Inside the fight over Claude Mythos 5 | The Verge

Some of the spiciest takes I’ve caught this week pulling this apart:

• (Surely this wasn’t just security theatre disguising a corporate shakedown of Anthropic by Amazon, one of its largest shareholders?

• Or, as Polymarket put it:

  

• Of course, we’ve been here before under encryption-as-national-security-theatre, most notably with the PowerMac G4:

• This doesn’t help with the evals:

  

• Ben Maling gets a clean sweep of the [not-so] complex undercurrents of this story:

  “Yesterday’s order by the US gov for Anthropic to suspend access by foreign nationals to the model advertised as “too powerful to release to the public” before being released to the public, and Anthropic’s wholesale retraction of the model in response is, simultaneously:
  
  God-tier PR for Anthropic 😇
  
  A convenient way to brush unsustainable inference costs under the carpet ahead of IPO 🤫
  
  A major boost for Chinese AI 🇨🇳
  
  Vindication for AI sovereignty advocates 💡
  
  Extortion by Trump 2.0 🍊
  
  Regulation stifling innovation 🙅🏼
  
  Entirely predictable 🤦🏻‍♀️“

Quite a few cartoon memes from around the internet are pretty spot on (run out of time this week to credit, use Google image search to locate sources?):

The real reason?

And then, on the topic of [AI] sovereignty specifically…

• US-Government-Authorised AI

  

  "If your AI strategy depends entirely on a single provider, a single country, or a single policy regime, then you do not control your AI. Someone else does.“

  — Reuven Cohen

• Georg Zoeller always worth listening to, this on current G7 negotiations:

“I’m sure the concessions to the colonial masters will be limitations on the import of competing technologies, primarily Chinese.

Because it’s just too obviously the play here and the only way AI companies survive.

With the news of Nvidia, the worlds richest company, performing a round of debt financing, the only conclusion one can come to is “The (Western) World is Not Enough” and we’re seeing the manufacturing of “too big to fail” beyond a single economy.

The shame in all of this is that the leaders of all these countries are getting themselves herded into even more colonial arrangements under the premise that US lead in AI is some kind of permanent thing rather than a relatively short term advantage bought at massive capital burn and that it masks an incredible weakness and reliance on its colonies underneath.

A single match, a better approach to AI or more efficient silicon, and it goes up in flames.”

• ⚖️🤖🏛️ Leviathan Waking

  Hyperdimensional

  Leviathan Waking

  Introduction…

  Read more

  7 days ago · 58 likes · 12 comments · Dean W. Ball

  Dean Ball is searing in Hyperdimensional: all of this is unfolding in an opaque fog of political grudge-matches and zero coherent regulatory framework. The US lurched from implausibly laissez-faire to incomprehensibly restrictive in weeks, not because of policy, but because a politically antagonistic company of “hopelessly literal nerds” forgot that what the law says and what a hostile administration does are entirely different animals.

  Not sure I’d count Anthropic out just yet. As Maling said above, this is god-level PR, while actually saving them millions in inference costs…

• @libraryofbabel on Hacker news:

  “So many comments here missing the big picture, and just gleefully pointing out that Anthropic got what they deserved, or that this is the natural culmination of some kind of marketing stunt.

  The real story here is that this may be the beginning of governments restricting the availability of strong LLMs to the public, to you. Fable was the strongest model on the market, and the US government has told you you can’t use it (technically, only if you’re not a US citizen, but in practice, even if you are). If you think the solution here is going to be open source Chinese models and / or running on your own hardware, think again. Do you think China is going to allow the strongest LLMs from companies within its borders to be open source a year from now when they have Mythos capabilities, if the US government is keeping the strongest American models back? Unlikely. These are heading in the direction of being powerful cybersecurity weapons and it will be in the interest of nation states to restrict and control them. In 2 years time, I would be surprised if the strongest LLMs are available for general use at all.

  Will we be the poorer for that, or will we be safer? I think poorer, because I hate being told what technology I can and can’t use, but I’m not certain. Maybe you think the government should restrict strong LLMs. Maybe you don’t. But either way, this is big news and a rubicon has been crossed and a precedent set. That’s true even if the motivation for this is just the government settling scores with Anthropic.“

On the subject of open-source AI:

• 🌍🤖⚖️ Steering Open-Source AI

  

  A 20-plus-strong international research team has published a Nature Communications commentary arguing that open-source AI is accelerating faster than any governance framework can track, threatening to simultaneously turbocharge the UN’s Sustainable Development Goals and undermine them through energy overconsumption, misinformation, and deepening global tech inequality. Their four-point prescription — lifecycle sustainability accounting, SDG-specific evaluation frameworks, shared accountability for synthetic content, and FAIR-principles-based global cooperation.

• 🤖⛓️🌐Bittensor and the rise of decentralised AI

  

  Grayscale’s “The Stack” explores the emerging tension between centralised AI giants like Anthropic and decentralised AI networks like Bittensor (TAO), which uses blockchain-based incentives to crowdsource AI model development and reward contributors with crypto tokens. The piece frames this as a foundational architectural question for the AI industry: who controls the intelligence layer, and does it matter if that control is distributed across a permissionless network rather than concentrated in a handful of San Francisco labs?

  • Here’s the BitTensor White Paper authored by Yuma Rao:

    “As with other commodities, markets could help us efficiently produce machine intelligence. We propose a market where intelligence is priced by other intelligence systems peer-to-peer across the internet. Peers rank each other by training neural networks which learn the value of their neighbors. Scores accumulate on a digital ledger where high ranking peers are monetarily rewarded with additional weight in the network. However, this form of peer-ranking is not resistant to collusion, which could disrupt the accuracy of the mechanism. The solution is an incentive mechanism that maximally rewards honestly selected weights, making the system resistant to collusion of up to 50 percent of the network weight. The result is a collectively run intelligence market that continually produces newly trained models and pays contributors who create information theoretic value.“

    Need to get under the hood with Bittensor some more…sounds too good to be true right now?

• 🤖🏛️⚔️ AI-Enabled Coups: Advanced AI could enable small groups to seize power

  

  More seriously on-topic: a paper from earlier this year from Forethought researchers lays out three converging pathways by which advanced AI could enable a single person or small group to seize control of a sovereign state — even a stable democracy — through what they term “singular loyalties” (AI systems trained to obey one person unconditionally), “secret loyalties” (AI sleeper agents that pass safety tests but execute a coup on command), and “exclusive access” (one actor quietly accumulating superhuman capabilities in weapons R&D, cyber offence, and strategic planning while everyone else is still running last year’s models). The scenarios range from a CEO ordering his robot workforce to surround the capital when regulators come knocking, to a government AI project head embedding backdoors that propagate through successive AI generations until secretly loyal systems are running the military.

  If this is the kind of future that you’d rather avoid, consider applying for Longview Philanthropy’s grants on Extreme AI Power Concentration.

🐽🚀Pigs can fly (for now)

Despite all of the seemingly rational voices crying “WOLF!”, SpaceX got their IPO away. Go figure:

(Remember, they hacked the index funds…)

Thread on what went down:

Ed Zitron is being rolled out around many newsrooms and podcasts, enjoying a truth-telling moment in the spotlight (…and getting ahead of OpenAI’s forthcoming IPO as well):

(More in depth coverage of SpaceX IPO and this week’s confirmation that the company is buying AI startup Cursor for US$60Bn way down below in Tech Industry News and Analysis.)

• 💸🐋🌕World’s first trillionaire

  Elon Musk has become the world’s first trillionaire overnight — a sum so vast that counting to it in seconds would require you to have started during the Paleolithic era, when Neanderthals were still around to witness your dedication.
  
  To put it another way, a trillion US$1 bills would weigh the same as 5,000 maximum-size blue whales, and at a modest 4% interest rate, Musk’s fortune generates roughly US$110 million *per day* — nearly enough to launch two Falcon 9 rockets before breakfast.

  There’s a bug in the system, right?

• 🌞🚀🧠 The Sentient Sun: SpaceX’s Culture-inspired endgame

  

  The bull case: A lengthy a16z profile of SpaceX (disclosure: they’re investors, so grain of salt optional) authored by Marc Andreessen and Michael McGuiness lay out the company’s full technology stack — from Falcon 9’s cash flywheel through Starlink’s US$11.4B revenue engine, up to orbital AI data centres drawing 100 terawatts of solar power, a self-sustaining Moon city, and eventually Mars — all explicitly modelled on Iain M. Banks’s Culture novels, whose sentient starships already lend their names to SpaceX’s drone landing platforms.

  

  An entertaining read, but Banks must surely be rolling in his (virtual) grave.

• 💸🌊 Flows, Not Fundamentals: Why the US Stock Market Is Now a Global Savings Drain

  What’s actually going on? The US equity market has quietly transformed from a capital-allocation mechanism into the world’s only drain large enough to absorb global surplus savings — with passive index funds (~45% of the market) acting as price-insensitive buying machines that amplify every dollar of inflow into roughly five dollars of market value, per the Gabaix-Koijen inelastic markets hypothesis. AQR research confirms the tell: US stocks outperformed the rest of the developed world by 4.7% annually over 35 years, but only 0.9 percentage points came from superior earnings — the other 3.8 points were pure rerating, i.e., investors agreeing to pay more for the same profits.

  I was reflecting this week that this year’s clutch of mega-AI-IPOs was likely preconceived many years ago as a massive-scale designed attack on passive index investment funds. Investment bankers and early investors all in on it for years, continuing to fund the pump-and-dump at never-before valuation multiples.

  When the crash comes, who’ll be left holding the bag? Not them, for sure.

  I’m getting cynical, eh.

• 💸🔥📉 OpenAI’s Billion-Dollar Burn Rate Exposed Ahead of Planned IPO

  

  Speaking of which… audited financial documents leaked by independent journalist Ed Zitron and verified by the Financial Times reveal the staggering economics behind OpenAI’s AI dominance play. The company spent US$34 billion in 2025 — including US$19.18 billion on R&D alone — against just US$13.07 billion in revenue, producing an operational loss of US$20.92 billion. The headline net loss ballooned to nearly US$39 billion, though roughly US$30 billion of that stems from a one-off accounting charge tied to OpenAI’s conversion from a non-profit to a for-profit structure. Strip that out, and the adjusted loss sits around US$8 billion — still eye-watering, but more digestible for prospective IPO investors.
  
  The numbers paint a picture of a company betting everything on scale. OpenAI paid Microsoft US$17.2 billion in 2025 for compute and services, while revenue from SoftBank (US$867 million) and Microsoft (US$303 million) barely dented the outgoings. On a relative basis, operating losses actually improved from 237% of revenue in 2024 to 160% in 2025 — progress, but the kind that still requires extraordinary faith from investors.
  
  With a confidential SEC filing already lodged, a potential US$1 trillion+ IPO valuation on the horizon, and rival Anthropic racing to list at US$900 billion, the pressure to demonstrate a credible path to profitability by 2030 has never been more intense.

  
  Coverage:

  • Exclusive: OpenAI Losses Increased Nearly 8X in 2025, With Spending Hitting $34 Billion

  • Leaked financial docs show OpenAI is losing billions of dollars a year - Ars Technica

  • OpenAI spending hit $34bn last year ahead of planned IPO

• 💸”Margin by utilisation”

  SemiAnalysis investigated whether the large AI labs really are delivering incredible subsidies to power users: turns out… pretty much yes, but apparently no big deal:

  Apparently OpenAI’s US$200/month ChatGPT Pro plan could cost the company up to US$14,000 in compute if a user actually maxed it out, with OpenAI tipping into losses once utilisation exceeds a mere 11.4%. The entire business model is essentially a gym membership bet — praying that most subscribers are paying for the vibe rather than the reps.

  So, this generates “subscription margin by utilisation” chart:

  

  “We therefore think it's far more likely the labs will withhold new features/models from subscription plans. It will be interesting to see if Mythos ends up being API only“

• 🥭FAANG is dead — meet MANGOS

  

  With SpaceX, Anthropic, and OpenAI all filing for potentially record-breaking IPOs, the tech industry’s power-broker acronym is getting a tropical rebrand: say goodbye to FAANG (Meta, Amazon, Apple, Netflix, Google) and hello to MANGOS — Meta, Anthropic, Nvidia, Google, OpenAI, SpaceX. The viral coinage, credited to developers @krishdotdev and @lilscoot on X, neatly captures the shift from streaming-and-e-commerce dominance to an AI and agentic-first era.

• 🤖💰🔌 Post-AGI Redistribution

  

  Epoch AI’s latest analysis cuts through the UBI-vs-welfare-state noise to identify the real axis of post-AGI redistribution debates: not how much citizens receive, but how much control they retain over the capital generating that income — from passive cash transfers (UBI) through sovereign wealth funds and universal basic capital (UBC), all the way to the delightfully provocative “UBC + kill switches,” where citizens could literally shut down their slice of the robot economy like a digital strike. The piece draws a sharp analogy to the classic cash-vs-in-kind transfer debate, but with civilisational stakes: if human labour becomes economically worthless, the political equilibrium underpinning any redistribution scheme may need technological enforcement, not just democratic goodwill.

OK… enough of me rolling my eyes about BIG AI, here are some more positive stories from this week:

• 🧬🤖💉 Variant-Proof: AI-designed coronavirus vaccine successfully tested in humans

  

  Researchers at the University of Cambridge have completed the first human trial of a vaccine whose key component was designed entirely by AI — scanning thousands of sarbecovirus genomes to identify the conserved features evolution has left largely untouched, then targeting those as the basis for a single, broad-spectrum jab against all known human coronaviruses and related bat viruses. The DNA-based vaccine (more stable than mRNA, needle-free, cold-chain-friendly) proved safe and produced cross-reactive antibodies, though immune responses were modest and durability remains unknown.

• 🗂️Google Cloud launches Open Knowledge Format for AI data sharing

  path/to/bundle/
  ├── index.md                      # Optional. Directory listing for progressive disclosure.
  ├── log.md                        # Optional. Chronological history of updates.
  ├── <concept>.md                  # A concept at the bundle root.
  └── <subdirectory>/               # Subdirectories organize concepts into groups.
      ├── index.md
      ├── <concept>.md
      └── <subdirectory>/
          └── …

  Google Cloud has launched the Open Knowledge Format (OKF), an open, vendor-neutral specification that formalises the “LLM-wiki” pattern — popularised by Andrej Karpathy — into a portable standard for representing the metadata, context, and curated knowledge that AI agents need to actually be useful inside organisations. It’s deliberately, almost aggressively simple: just a directory of markdown files with YAML frontmatter, no SDK, no proprietary runtime, readable by humans and parseable by agents from the same file.

• 🎛New tool helps nations measure and build AI agency

  

  

  Canberra-based Tech Policy Design Institute (TPDi) has released a landmark framework called the AI Agency Tool, which replaces the increasingly useless binary of “do you have AI sovereignty or not?” with a nuanced 103-capability assessment across six ecosystem layers — producing a sovereignty spectrum (access → control → choice → leverage) and a single “agency score” measuring genuine national competitive advantage.
  
  Applied to Australia as its first test case, the tool finds the country has mostly emerging AI maturity but surprisingly high agency — meaning Australia has real leverage to shape its AI trajectory if it plays its cards strategically, particularly through critical minerals, world-class domain datasets, and computer vision expertise.

• 🐦📊💼 Canaries: Stanford’s AI Economic Observatory Goes Live

  

  Stanford’s Digital Economy Lab has launched the AI Economic Indicators project — a suite of regularly updated dashboards designed to cut through AI hype with actual data on employment, wages, productivity, and adoption.

  

  The most striking early finding: entry-level workers aged 22–25 in high AI-exposure occupations are seeing measurable employment declines since ChatGPT’s arrival, while older workers remain largely unaffected — and across 12 broader transformation indicators, there’s still no decisive evidence of the explosive economic growth the boosters keep promising.

  Lots of useful data points and insights in here, go and take a look.

• 🇪🇺🤖🔮 Europe 2031: “Feel the AGI”

  

  A new speculative fiction project, Europe 2031, uses a fictional Brussels trade official named Caroline Dubois to dramatise a very real epistemic chasm: European policymakers celebrating DeepSeek R1 as proof that scrappy efficiency can beat American AI dominance, while Silicon Valley quietly ships o3-mini and keeps building.
  
  The piece nails the cultural untranslatability of AI urgency between Silicon Valley, China and Europe — Caroline can’t bring herself to relay what she witnessed in San Francisco to her director, because the calm, sharp conviction that AGI is “two or three years out” would simply sound unhinged in a Commission conference room.

• 🏛️🤖⚡ The Agentic State: A Blueprint for AI-Powered Government

  

  Launched at the Tallinn Digital Summit 2025 and backed by the World Bank’s Global GovTech Centre, *The Agentic State* is an ambitious vision paper and action platform arguing that agentic AI — systems that can perceive, reason, and act autonomously — represents the most significant disruption to government since the rise of modern bureaucracy. The framework maps twelve functional layers across which AI agents could transform public administration, from proactive service delivery and real-time regulatory compliance to crisis response at machine speed and autonomous procurement. Ukraine is already piloting with Diia.AI, billed as the world’s first national digital agent for government services, with targets like resolving 95% of single-interaction requests without human intervention and launching new digital services in a single day.

  

  The paper draws on contributions from over 20 digital government leaders across 15 countries and makes a compelling case that delay is costlier than experimentation. As citizens and businesses increasingly deploy their own AI agents, governments stuck on manual processes risk irrelevance — or worse, ceding service delivery to private intermediaries operating on commercial rather than public-interest priorities.
  
  The authors are careful to note that higher autonomy isn’t always better; most practical deployments sit at Level 3 (agentic workflows with human oversight), analogous to highway autopilot rather than full self-driving. The real shift isn’t automation for its own sake, but moving government from “doing things right” to “doing the right things” — defining outcomes and letting agents optimise within democratic guardrails.

• Coverage:

  • The Agentic State

  • The Agentic State - Vision Paper

• 📰Using AI for news fact-checking erodes your misinformation detection skills

  

  A new MIT Media Lab study found that people who used AI chatbots to verify news became 21% better at spotting misinformation in the moment — but after a month, their unassisted detection skills had declined by 15 percentage points, and roughly a quarter of them didn’t even notice, feeling more confident than ever (hello, Dunning-Kruger).
  
  The “AI dependency paradox” mirrors well-documented deskilling effects from GPS and calculators, and the researchers’ proposed fix is elegant: design AI that acts as a Socratic coach rather than an answer machine, asking guided questions instead of just handing over verdicts.

  …but there’s no going back, right? Need more and more trustable, dependable agents and trust-chains so that human meatbrains have any hope of still doing the deciding…

• 🛡️🇪🇪🤖 Propaganda Resistance Benchmark

  

  Estonia’s Institute of the Estonian Language has quietly released one of the most strategically significant AI safety benchmarks yet: a 75-question gauntlet across Estonian, English, and Russian that tests whether LLMs will parrot Kremlin talking points when pressured with neutral, biased, or outright manipulative prompts across 14 documented Russian propaganda narratives. Anthropic’s Claude models dominate the top rankings (Claude Fable 5 scores 95.2/100), while older models collapse spectacularly under “malicious” questioning — GPT-3.5 Turbo limps in at 48.2, and even GPT-OSS 120B scores a concerning 23.7 on malicious queries alone.

• 📱🪙 The inference era: China’s carriers bundle AI tokens into monthly phone bills

  

  China’s three state-owned telecoms — China Mobile, China Telecom, and China Unicom — launched AI token subscription packages in May 2026, letting consumers pay for AI inference the same way they pay for mobile data, with plans starting from as little as RMB 9.9/month for 10 million tokens. With revenue growth stalling below 1% across all three carriers in 2025 and combined compute capex approaching RMB 81 billion for 2026, the carriers are betting their existing billing infrastructure and last-mile networks can make them the dominant distributors of AI compute at scale — a commodity play that maps neatly onto the voice-minutes-to-data-gigabytes lifecycle they’ve run twice before.

• 🚢🖥️🌊Samsung launches approved floating AI data centres on ships

  

  Samsung Heavy Industries has partnered with Greek shipowner Capital Clean Energy Carriers to build purpose-built Floating Data Centers (FDCs) from scratch — not converted vessels — with its 50MW-class design already cleared by both the American Bureau of Shipping and Lloyd’s Register. The concept, which also involves partnerships with Supermicro for marine-hardened server setups and Dallas-based Mousterian Corporation for water-adjacent infrastructure expertise, aims to merge Samsung’s shipbuilding pedigree with the exploding demand for AI compute capacity.

• 📱🚼Why birth rates are falling everywhere

  This is a really fascinating watch - FT Data journalist John Burn-Murdoch dives into an analysis of human fertility decline in countries around the world… finding that there is a very strong correlation with the moment that a country experienced mass adoption of smartphones:

  

  (What should the governments of the world do about that, if true?)

  (cf. Switzerland’s population cap vote this week…!)

• Flying Beta

  Air New Zealand are partnering with electric aircraft maker Beta, flying trips across the Cook Strait from Wellington to Blenheim and back. (Don’t expect this to go commercial any time soon…)

• 🏭📦🚁 “Infrastructure is the Weapon”: The Race to Build Portable Drone Factories

  

  As demand for interceptor drones surges amid the Iran war, Ukraine’s hard-won lesson — that a drone without the production ecosystem behind it is just an expensive paperweight — is driving a new market for containerised, portable drone factories, with Finnish firm Sensofusion’s US$2.4M shipping-container unit producing 50 interceptors a day at under US$580 each, while US-based Firestorm Labs and French firm Per Se Systems offer competing approaches. Critics note a fundamental irony: 3D-printed airframes still depend on the same global supply chains for motors, batteries and electronics that the factories are supposedly designed to bypass, and no company has demonstrated sustained quality production outside controlled environments.

  
  Here’s a pretty convincing pitch:

• ☕🔊⚡ Ultrasonic Espresso

  

  Finally, back in civvy-land, researchers at UNSW Sydney have developed a genuinely novel coffee brewing method that uses ultrasonic sound waves to produce espresso-strength coffee from room-temperature water — cutting energy consumption by up to 75%. Led by Dr Francisco Trujillo from the School of Chemical Engineering, the system works by pressing a transducer against a traditional filter basket, generating high-frequency vibrations that create acoustic cavitation: microscopic bubbles that collapse near coffee particles, effectively acting as tiny scrubbing brushes that fracture the grounds and rapidly extract flavour compounds, oils, and caffeine without any heat.
  
  The proof is in the cup. In blind taste tests with around 100 regular coffee drinkers, participants couldn’t reliably distinguish the ultrasonic espresso from traditionally brewed shots across aroma, flavour, bitterness, or overall liking. For filter coffee, the ultrasound version was actually *preferred*.
  
  While a home machine is feasible, the real opportunity lies at industrial scale — ready-to-drink coffee manufacturers could see massive energy and time savings by producing concentrated coffee at room temperature, then shipping it for dilution into bottled cold brews and milk-based beverages.
  
  Published in the *Journal of Food Engineering*, this builds on the team’s earlier work accelerating cold brew from 12–24 hours down to three minutes.

  
  Coverage:

  • Entirely new way of making espresso shakes up the coffee world

  • Scientists Made Espresso Without Hot Water, And People Couldn’t Tell : ScienceAlert

📈The week in AI and Tech

OK everything from here on brief AI summaries with links… there’s a lot to get through as always!

AI safety

• ⚠️Google DeepMind funds $10M research into multi-agent AI risks

  Google DeepMind, alongside Schmidt Sciences, the UK’s moonshot agency ARIA, the Cooperative AI Foundation, and Google.org, has announced a US$10 million funding pool to kickstart academic research into what happens when millions of AI agents start interacting with each other at scale — a scenario that DeepMind’s AGI safety lead Rohin Shah says could hit critical mass within months.

• 🧠💥🌍 Recursive Self-Improvement: AI’s intelligence explosion demands urgent global governance now

  

  Planet Labs CEO Will Marshall argues in The Economist that AI’s estimated 10-50% chance of causing a catastrophic event dwarfs the 1-in-a-million threshold we accept for nuclear meltdowns — and that recursive self-improvement (RSI), where AI rewrites its own code without human intervention, could arrive within months to years with essentially no governance framework in place. He proposes a US-China bilateral verification regime modelled on post-WWII nuclear arms control, while invoking the Fermi Paradox as a chilling frame: civilisations may routinely self-destruct precisely by building powerful technologies faster than the institutional wisdom to govern them.

• ⚔️🛡️AI widens the cybersecurity gap between haves and have-nots

  A LessWrong post coins a useful concept: the “dual-use gap,” the widening asymmetry between AI’s offensive and defensive cybersecurity effects — not because defence loses, but because defence doesn’t scale to everyone equally. When AI makes both attackers and defenders n-times more capable, the world isn’t neutral; it’s just faster, and the three-person startup, the solo npm maintainer, and the school district are now living in a more dangerous environment while frontier labs enjoy AI-powered SOCs.

Governance and Policy

• 🤖💸🗳️ Regulatory Capture, But Make It Bipartisan: How AI PACs Are Buying US Congress

  

  

  A new transparency tool, electhumans.com, has mapped US$37.7M in AI industry political spending across 69 candidates in 25 states — with OpenAI, a16z, Palantir, and Perplexity’s “Leading the Future” network dropping US$22.1M, and even the ostensibly safety-focused Anthropic channelling US$13.7M through its Public First network, with US$30M+ of the total flowing through dark-money structures with undisclosed donors. The most jaw-dropping single data point: US$12.5M spent targeting one House primary candidate (NY-12’s Bores), more than many Senate races attract in total.

• 👶🏛️White House ties AI preemption to child safety in last-ditch push

  Cue…with US midterms looming and a potentially hostile Democratic Congress on the horizon, Big Tech lobbyists are making a desperate final push for federal AI preemption — a single national law overriding the current patchwork of state-level AI regulations — by bundling it with child online safety legislation (KOSA), in a marriage of convenience that almost nobody in Washington actually wants. The White House’s approach is being shaped by Trump-allied lawyer Mike Davis’s “Four Cs” framework (children, conservatives, creators, and communities), but the plan is already unravelling amid confusion over which version of KOSA to use, a packed legislative calendar, and Democrats who only just found out they were being dragged along for the ride.

• 🤖🏛️📋 Board Bots: Most UK boards weighing AI’s role in corporate decision-making

  A survey of 400 UK board directors, CEOs and CFOs finds 84% are actively debating which decisions to hand over to AI — with nearly half already moving from discussion to implementation, and 8% convinced boards need complete reimagination. The governance gap is the real story: experts warn that while operational AI adoption races ahead, the oversight structures meant to govern it are barely keeping pace, with a frankly alarming 40% of directors believing only incremental board-level change is needed.

• 🤖⚖️Agentic AI adoption outpaces laws meant to govern it

  

  The ACM Technology Policy Council’s new TechBrief on agentic AI finds that systems autonomously browsing the web, executing code, managing files, and sending messages on your behalf are already deployed by nearly half of tech enterprises, while legal, regulatory, and technical safeguards lag dangerously behind; documented harms include an AI agent that wiped an entire production database and a prompt injection attack that silently exfiltrated private Slack data. When something goes wrong, existing licence agreements just “point fingers elsewhere,” and as ACM’s Simson Garfinkel bluntly notes, no case law currently exists to resolve who’s actually liable.

Regulation

• 🚫📱🧒UK bans social media for children under 16

  

  The UK government has announced a ban on social media for under-16s — covering TikTok, Instagram, YouTube, X, Facebook, and Snapchat — with enforcement expected from spring 2027, modelled on Australia’s penalty-based compliance framework and backed by Ofcom-mandated age verification that may include facial recognition. The ban extends to AI “romantic companion” chatbots for under-18s, with potential overnight curfews and infinite-scroll breaks also on the table, while critics — from the EFF to the Liberal Democrats to Nigel Farage, in a rare moment of unlikely alignment — warn that VPN circumvention, privacy erosion, and driving kids toward less regulated corners of the internet may simply trade one set of harms for another.

• 🧒📵Canada proposes social media ban for children under 16

  

  Canada too: introducing Bill C-34, the “Safe Social Media Act,” proposing to ban social media for under-16s — but with a notable twist: platforms can keep serving minors if they can demonstrate their services are genuinely safe, flipping the burden of proof onto Big Tech rather than regulators. The bill also takes a nuanced stance on AI chatbots, stopping short of age restrictions but mandating safety-by-design requirements around relationship-building behaviours, self-harm responses, and harmful content — because apparently “it’s educational” is still a get-out-of-jail card in Ottawa.

• 🔍📱🇬🇧 “Surveillance Is Not Safety”: Signal calls UK device-scanning plan “mass surveillance”

  

  Announced by PM Keir Starmer at London Tech Week, the UK’s latest online safety proposal would require Apple and Google to activate on-device scanning across smartphones and tablets — cameras, messaging apps, search, the lot — to detect and block nude images on children’s devices, with adults required to pass age verification to access such content. Signal has fired back with a statement arguing the plan normalises inspecting private communications and builds surveillance infrastructure that, however well-intentioned, never stays narrowly scoped.

  These state surveillance vs. privacy battle lines just keep being tested in free democracies… technically you just can’t achieve privacy AND a back door. Choose one.

• 📵🪪🔥 Real-Name Registration: FCC proposes mandatory ID collection, threatening burner phones

  The FCC is proposing to require all US telecoms to collect government-issued ID numbers and physical addresses from every new and renewing customer — effectively ending anonymous prepaid phone ownership in a country that has long prided itself on not being that kind of place. The stated rationale is combating scammers, but the ACLU notes the collected data comes with a suspiciously long list of other potential law enforcement uses, and the real-world casualties would include domestic abuse survivors, journalists, and low-income people who rely on prepaid plans.

• 📵🎓🇮🇳India temporarily bans Telegram to combat “exam fraud”

  

  India’s National Testing Agency has ordered a temporary nationwide block on Telegram until June 22 — and demanded the platform disable its message-editing feature until June 30 — to prevent cheating rackets from selling fake papers ahead of the NEET (UG) medical entrance re-test, affecting what Sensor Tower estimates as 354 million monthly active users in Telegram’s single largest market. CEO Pavel Durov fired back, noting the leaks simply migrated to other apps, that Telegram had already removed hundreds of related channels, and lobbing an extra allegation that Reliance was disrupting Telegram access via unauthorised routing — while, somewhat deliciously, the ban appeared only partially enforced at publication time, with the app still accessible and editing still functional.

• 🎮🪦📜 Stop Killing Games: The EU Blinks

  The European Commission has declined to pursue mandatory game preservation legislation, citing existing copyright and IP law constraints, opting instead to develop a voluntary code of conduct with the games industry around how developers and publishers handle sunsetting titles — with a report due by end of 2026. The decision follows the Stop Killing Games grassroots campaign, sparked by Ubisoft’s 2024 deletion of The Crew from players’ libraries, which gathered enough signatures to formally compel the EC to address the question.

• 🤝Global AI regulation urgently needed, but likely to fail

  

  FT chief economics commentator Martin Wolf argues that AI poses dangers in three overlapping categories — erosion of human identity and accountability, specific catastrophic risks (cyberattacks, engineered pathogens, autonomous weapons), and broad societal disruption — and that meaningful regulation requires a US-China “AI disarmament treaty” as a prerequisite, since the geopolitical arms race between the two superpowers is what’s fuelling the “mafia-like” competitive frenzy among US tech companies. Wolf’s conclusion is bracingly honest: regulation is necessary, but will probably fail.