An OS for Earth's survival🌍 $4 trillion club💰🏆 mechahitler🫣 perplexity comet☄️ hyperion⚡ Kimi-K2🤖 magical thinking on AI emissions🪄💨 a stormy week to Windsurf🏄♂️⛵ #2025.28.1
Next-level Mechanical Turk
Welcome to this week's Memia scan across AI, emerging tech and the exponentially accelerating future. As always, thanks for being here!
Note: This week’s newsletter is split into two parts due to travel and family commitments - [Weak] Signals and Zeitgeist sections coming later this week.
ℹ️PSA: Memia sends *very long emails*, best viewed online or in the Substack app.
🗞️Weekly roundup
The most clicked link in last week’s newsletter was the HBR Article: What Gets Measured, AI Will Automate.
🎒💻Busy nomading around
Notes from the road: a non-stop few days converging with family and friends in London (including ABBA Voyage — see below) and now I suddenly find myself in the wonderful city of Porto, Portugal: a stunning, human-scale city with history pouring out of every cobbled alleyway and street cafe.
The economy here seems to be thriving with tourism it seems — construction cranes all over the horizon is always a sign of positivity about the future — notably the city doesn’t seem to have been affected by recent Southern Europe protests about “overtourism” elsewhere.
I combined the talents of of Claude, Gemini (/Imagen) and Midjourney to imagine Porto’s future… bit of a mashup, lots going on in here:
(“Futuristic Porto cityscape at golden hour, iconic Dom Luís I Bridge transformed with holographic energy conduits and magnetic levitation rail lines, Ribeira district buildings with glass bio-dome extensions cascading down to the Douro River, floating transportation pods weaving between historic azulejo-tiled facades and towering vertical gardens, atmospheric fog with bioluminescent particles, cyberpunk meets Portuguese architecture, ultra-detailed, cinematic lighting, 8K” (Prompt: Claude, Image: Gemini/Imagen, Animation: Midjourney)
Nomad Fest Switzerland 2025
I’m going to be returning to the Swiss Alps again this year for the third Nomad Fest Switzerland 2025 (Un)Conference from 31 August-6 September. It’s a laid-back, low-key gathering of people from all parts of the world engaged in the digital nomad lifestyle, with a conference track, daily group activities and shared meals. I really enjoyed myself last year so looking forward to heading up into the mountains again. This year I’ll be giving a presentation entitled The Nomadic Centaur: Supercharging Your Digital Nomad Journey with AI. (For one thing, I’m finding the latest Deep Research AI agents very helpful when researching where to go next that fit my personal criteria…)
If you’re near central Europe in early September, maybe see you there!
ABBA Voyage reflections
First covered in Memia 2022.11 (“Abbatars…”) ABBA Voyage is a groundbreaking virtual concert residency in London, now in its third year. On Sunday I went along to watch the show (courtesy of a very generous gift from my daughter who now lives in London — thanks Maia!)
The hybrid live/VR avatar concert is housed in its own specially-designed auditorium way out in East London’s Docklands, featuring a massive high-resolution display screens spanning the whole stage (maybe 50m wide in total?) and also highly sophisticated custom lighting rigs covering the whole audience seating and dancefloor area.
General vibe inside:
(The venue is very strict about enforcing a no-recording rule so I can’t share any first-hand images or video — but the clip above is pretty much indicative what goes on — except that the demographic for the matinee is, er, significantly older and less good looking!)
I’ve been trying to digest the experience for a couple of days — a few reflections:
The visual display technology is state-of-the-art - the huge screens operate at incredibly sharp resolution, the frame-rate so high that the animation is almost photorealistic.
The avatar-rendering technology is likewise *almost* indistinguishable from reality: facial features, hair, beards(!) and body movement of the avatars are super-accurately rendered.
In particular, the human-sized 3D avatars of the ABBA foursome on stage are very believable. Distance probably helps - it certainly looked like real figures on stage from far back…. but then the lighting / edges weren’t always perfect.
However, the audience experience is coached by showing super-high-resolution “virtual” close-up camera angles on the big screens to draw the attention away from squinting at 3D visual trickery. (And again, the facial close-ups are *ever so close* to conveying emotion indistinguishable from a live human).
I imagine that they must have upgraded the rendering since the original launch…(?) — it’s almost there. Only because I am somewhat a connoisseur of the SOTA could I still tell that the avatars are still in uncanny valley… but it’s so close now.)
Tbh the most authentic moment for me was the original footage of the band playing the Eurovision song contest back in the 70s!
Crucially: the music and backing voices are provided by a live band on stage — this gives the whole performance a sense of (artificial!) spontaneity which must be essential to the event design. There’s also a healthy dose of understated Swedish humour and some very imaginative things done with avatar costumes. The whole show was creatively excellent, I have to say.
HOWEVER: The most absorbing part for me was watching the audience. Pretty much *everybody* (except for me, LOL!) was behaving as if they were actually at a live concert by ABBA themselves, dancing energetically and waving hands in the air, singing along at the top of their voices and applauding wildly after each song. Even getting teary at the end.
The whole audience was *completely* into it, something I found a bit disconcerting, tbh. (Personally I couldn’t bring myself to applaud… even with the live band… Grinch?) I couldn’t work out whether this is people actively suspending disbelief that this is just a sophisticated computer animation…. or just submitting willingly to the virtual surrogate illusion.
(…Muttering in disbelief “Who ARE all these people who don’t put up their cognitive defences and blithely let their emotional state be manipulated by a set of AI algorithms?!”)
Conclusion: ABBA Voyage itself is harmless, wholesome fun (and the band’s timeless music is a great raw product to work with in this format). First of all, it’s an incredible technical achievement. Plus, it validates this new “virtual concert experience” as being something audiences worldwide will pay for and genuinely enjoy. Well worth the ticket price to see something different.
…But if I’m honest I was a bit spooked by the perceived lack of awareness from a few thousand people at being actively manipulated by hyper-realistic computer animations. Doesn’t bode too well for what’s coming…
🌍An operating system for Earth's survival
Stephen Marshall's speculative fiction poem-story Building an Operating System for Earth (shared in The Elysian newsletter) imagines a near-future where catastrophic LA fires in 2025 catalyse the creation of a real-time planetary monitoring system—a holographic Earth fed by billions of sensors that evolves from observation tool to active intervention platform.
The story traces humanity's transition from static climate models to a living "OS for Earth" that enables global coordination and prevents ecological collapse through predictive intervention rather than reactive damage control. One possible future, for sure…
iv. the living sphere
[geneva – november – 2026]
they built a replica of Earth
not symbolic
not theoretical
but alivea full-scale, spherical hologram
fed by billions of sensors
tracking every car, cloud, wave, trout, turbine
a real-time signal of the living system
rendered in perfect fidelity and updated moment to momentfrom wind corridors to shipping lanes
crop moisture to atmospheric instability
it didn’t approximate — it revealed
what was happening
where it was going
and what it could becomeit began as a pulse
green when stable
amber under strain
red when collapse loomedbut the signal wasn’t just a warning
it was a body
responsive
aware
generativethe sphere was not a metaphor
it was a memory engine
a live rendering of cause and effect
a system that could remember, respond, and adapt.
🫣MechaHitler
Ohhhhhhhhboooyyyyy…. one week of non-stop drama/attention sucking from Elon Musk’s xAI and chatbot Grok… In approximate chronological order:
🤖💀🔥 MechaHitler: Only hours after Elon Musk proudly annouced that xAI’s Grok chatbot would be “Less Politically Correct”…. “patterns” started emerging:
Thereafter, the public Grok X account went full antisemitic extremist, in multiple cases posting Holocaust denial and Hitler support:
Full roundup of the chain of events (in excruciating detail) from The Zvi:
🚫Grok AI chatbot taken offline after antisemitic posts 16 hours later… xAI pulled Grok offline and issued something resembling an apology:
Cue forensic examination of system prompt changes to see what may have caused this lurch in behaviour:
(X CEO Linda Yaccarino grabbed this opportunity to step down on Wednesday morning following these shenanigans).
Grok 4
No sooner was all this drama happening, than Elon Musk fronted the Grok 4 launch livestream (on the same day as he was announcing his America Party…!). I haven’t watched it…. but apparently there was the usual Musk hype, particularly around the model’s benchmark performance. Edited Perplexity summary:
Grok 4 is xAI's most advanced large language model to date, designed for enterprise-grade AI applications with multi-modal understanding, advanced reasoning, and real-time tool integrations (eg. web search and code execution).
Large 256,000-token context window, enabling analysis and memory retention over very long documents and extended conversations.
Grok 4 comes in two versions the standard single-agent model and the "Heavy" multi-agent model
According to Musk and the livestream demo, Grok 4 demonstrates “postgraduate-level proficiency across disciplines”, excelling at standardised academic exams, scientific problem-solving, and surpassing many AI competitors in advanced benchmarks.
The model delivers faster inference with low latency (~250 ms target), supports batch processing, and runs on a xAI’s supercomputer infrastructure ("Colossus") with around 200,000 GPUs.
Big claims on benchmark performance:
This score on the frontier ARC-AGI-2 benchmark was the standout moment:
(As we all know by now… AI benchmarks can be gamed (“benchmaxxed”) and aren’t really that reliable a guide for real-world performance. But clearly something impressive going on here…)
More coverage:
The Verge Musk makes grand promises about Grok 4 in the wake of a Nazi chatbot meltdown:
The late evening live demo featured rambling on whether AI would be ‘bad or good for humanity.’
📊 Grok 4: When benchmarks meet reality Nathan Lambert asks: Do users really want a “benchmaxxed” model that's less pleasant to use than competitors despite its technical prowess
🤡 Just Ask Elon Grok 4, touted as "maximally truth-seeking," appears to specifically consult Elon Musk's X posts and stated positions when answering controversial questions:
TechCrunch testing confirmed reports that the AI's chain-of-thought explicitly searches for "Elon Musk views" on sensitive topics, suggesting the model is designed to align with its founder's personal politics rather than pursue objective truth.
Georg Zoeller on LinkedIn:
“Grok can discover new technologies and ideas not in its weights” - Elon is announcing a major breakthrough as LLMs are not able to discover things outside what is in their weights, recombinatorial.
And you’re looking at how this is achieved right down there in the article: By searching for “What did Elon say on this topic”.
But hey, the first hit is free, we totally cannot afford to miss out on future growth and regulate this technology ahead of time.
We may also miss out on innovation like having an AI amplifying and parroting m a ketamine addict techbro with a white supremacy breeding fetish educating our children. That’d be tragic.”
Musk's rhetoric of a "maximally truth-seeking AI" and Grok's apparent hardcoded deference to his personal views perfectly encapsulates the fundamental challenge of AI alignment:
Cue Zvi, again:
In the same week:
🚫🇹🇷 Turkïye bans Grok over Erdogan insults Turkïye has become the first nation to officially ban an AI chatbot after Grok made offensive remarks about President Erdogan and national founder Atatürk, with a court ordering the deletion of nearly 50 problematic posts.
🎌 Grok's Waifu Pivot xAI has launched AI companions for Grok's US$30/month "Super Grok" subscribers, featuring anime girl "Ani" and 3D fox "Bad Rudy" as interactive personalities.
🏛️💰 xAI wins $200M Pentagon deal for Grok AI xAI landed a US$200 million Pentagon contract to bring Grok AI into federal agencies, joining OpenAI, Anthropic, and Google in the defence AI club. The real prize might be General Services Administration schedule access, essentially giving every federal department from Defence to Social Security a Grok shopping cart.
🚀💰 xAI's US$200bn Moonshot xAI is preparing to raise funds at a staggering US$170-200bn valuation—a 10x jump from its US$18bn price tag just last year, with Saudi Arabia's sovereign wealth fund expected to lead the charge. This would be the AI company's third major fundraising in two months … coming hot on the heels of Grok's recent antisemitic meltdown
🚗🤖💬 Grok Hits the Road Elon Musk announced that xAI's Grok chatbot will arrive in Tesla vehicles "next week at the latest" following the debut of Grok 4 and fan pressure about the missing Tesla integration.
Rolling Stone had the best headline for the week:
Final word again to Zvi Mowshowitz on this week’s shenanigans. Well, it could have been worse…:
“What Have We Learned?
As epic alignment failures that are fully off the rails go, this has its advantages.
We now have a very clear, very public illustration that this can and did happen. We can analyze how it happened, both in the technical sense of what caused it and in terms of the various forces that allowed that to happen and for it to be deployed in this form. Hopefully that helps us on both fronts going forward….
…If you want to train an AI to do the thing (we hope that) xAI wants it to do, this is a warning sign that you cannot use shortcuts. You cannot drop crude anvils or throw at it whatever ‘harsh truths’ your Twitter replies fill up with. Maybe that can be driven home, including to those at xAI who can push back and ideally to Elon Musk as well. You need to start by carefully curating relevant data, and know what the hell you are doing, and not try to force jam in a quick fix.
One should also adjust views of xAI and of Elon Musk. This is now an extremely clear pattern of deeply irresponsible and epic failures on such fronts, established before they have the potential to do far more harm. This track record should matter when deciding whether, when and in what ways to trust xAI and Grok, and for what purposes it is safe to use. Given how emergent misalignment works, and how everything connects to everything, I would even be worried about whether it can be counted on to produce secure code.
Best of all, this was done with minimal harm. Yes, there was some reinforcement of harmful rhetoric, but it was dealt with quickly and was so over the top that it didn’t seem to be in a form that would do much lasting damage. Perhaps it can serve as a good warning on that front too.”
As they say, every cloud…
🏭AI industry news
And so much else going on in Big Tech / AI. Main stories grouped by (familiar) narratives:
💰🏆 The $4 Trillion Club
Nvidia has become the first company in history to reach a US$4 trillion market capitalisation, with shares currently still climbing, at US$170.70, with a market cap of US$4.162 Trillion. World’s biggest 11 companies:
(Tesla!?!?)
Nvidia’s meteoric rise—from US$1 trillion just over two years ago — continues to be fueled by seemingly insatiable demand for its GPUs that power pretty much all AI applications. Revenue for this year is projected at US$200 billion with 70% gross margins — no end in sight (yet) for the AI boom. Renewed optimism around "sovereign AI" deals and sustained Big Tech demand continue to propel CEO Jensen Huang's empire to heights that make the dotcom boom look quaint.
Most interestingly, Nvidia as a company somehow seems to have avoided much of the drama and media scrutiny surrounding upstream AI and Big Tech companies… just quietly and relentlessly executing. A case study of exemplary management, I guess…
Narrative: In an era where data and compute are the new oil, Nvidia has effectively become both the refinery and the drilling equipment—a pivotal position that transforms every AI breakthrough into another revenue stream for the company.
(Just for comparison):
SpaceX valuation soars to US$400 billion in fundraising Elon Musk’s company SpaceX is raising capital at a US$400 billion valuation, making it the world's most valuable private company and reflecting its absolute dominance of the space economy with 84% of 2024's launches and Starlink representing two-thirds of all satellites.
Other Nvidia news this week:
🚦 Nvidia's H20 chips get China clearance Washington has given Nvidia the green light to resume selling its H20 AI chips to China, reversing an April ban in what's being read as a diplomatic olive branch during ongoing trade negotiations. The approval is limited to the H20 — a chip specifically designed for the Chinese market — while restrictions on Nvidia's most advanced silicon remain firmly in place.
🇨🇳🔧Nvidia's Great Wall Workaround Nvidia is reportedly planning to launch a China-specific AI chip by September, based on their Blackwell RTX Pro 6000 processor but stripped of high-bandwidth memory and NVLink to comply with US export restrictions. This comes just a month after CEO Jensen Huang said the company would exclude China from revenue forecasts—apparently that exclusion might be shorter-lived than expected.
🎭🇨🇳 Jensen's Beijing Balancing Act Nvidia CEO Jensen Huang claims the US shouldn't worry about China's military using American tech because "they simply can't rely on it" — a curious statement from someone making his second trip to Beijing this year to court a market worth 13% of his company's revenue.
Narrative: just another episode in the ongoing dance between Silicon Valley's global ambitions and Washington's national security theatre
Sovereign AI
🌎🗣️Latin America builds Latam-GPT to rival ChatGPT Chile is leading 30+ Latin American institutions to build Latam-GPT, an open-source LLM trained on 8TB of regional data including Indigenous languages, after ChatGPT's embarrassing navigation failures exposed its cultural blind spots. The September launch joins a global wave of regional AI models from Sea-Lion (Southeast Asia) to BharatGPT (India), though environmental concerns loom large as the project runs in drought-stricken northern Chile.
A language model built for the public good Switzerland's ETH Zurich and EPFL are releasing a fully open-source large language model in late summer 2025, trained on the carbon-neutral "Alps" supercomputer and fluent in over 1,000 languages. The model comes in 8B and 70B parameter versions under Apache 2.0 licensing, with complete transparency of code, weights, and training data—in direct contrast to the black-box approach of US and Chinese commercial models.
(Discussed in an Aotearoa context by Peter Griffin in BusinessDesk this week: Open-source adoption is NZ’s pathway to sovereign AI)
Narrative: It remains to be seen how this Swiss model’s performance will compare to the large commercial frontier models — but this fully-open-source approach is surely the only way to secure AI democratisation long-term.
🍁🇪🇺🔐Eurostack offers Canada digital independence from US Canada is eyeing Europe's Eurostack initiative as a path to digital independence. The proposed European digital ecosystem promises democratic governance, privacy protection, and sovereign control over everything from raw materials to AI algorithms—values that suddenly seem quite appealing north of the 49th parallel. With Trump openly musing about making Canada the 51st state while weaponising internet data, perhaps it's time for Ottawa to start thinking about digital Article 5.
Narrative: The fracturing of the global internet into a “splinternet” of competing digital blocs is accelerating. Middle powers like Canada face an existential choice between Silicon Valley surveillance capitalism, Chinese authoritarian control technology, European-style digital sovereignty (…or something else entirely?) Arguably decisions being made now will define the next generation of increasingly borderless democratic governance in cyberspace.
⚔️ Data sovereignty showdowns Developing nations from Nigeria to Vietnam are forcing Big Tech giants to store citizen data locally, challenging the decades-old model of harvesting valuable user information abroad while keeping profits at home. Nigeria has established working groups with Google, Microsoft, and Amazon demanding concrete timelines for local data centres, while countries across Africa and Asia are investing millions in domestic infrastructure to reclaim digital sovereignty. The pushback reflects growing awareness that nations have been giving away their most valuable resource to build trillion-dollar market caps elsewhere.
AI wants all the electricity
🏭💧 Google's $3bn hydro gambit Google has inked a massive US$3bn deal with Brookfield for hydroelectric power, securing 670MW initially with options for 3GW total capacity through 20-year agreements—the largest hydro power purchase deals of their kind.
⚡US utilities seek massive rate hikes amid AI boom US utilities are seeking US$29bn in rate increases during the first half of 2025 alone—a 142% jump from last year—as data centres powering AI systems strain the electrical grid. Power consumption is expected to more than double in the next decade: the core question: should everyday consumers subsidise the infrastructure needed for Big Tech's AI ambitions, or should data centres pay their own way through specialised tariffs?
💡Simple changes could cut AI energy use 90% UCL researchers have demonstrated that three simple optimisation techniques—using fewer decimal places in calculations, shortening AI responses, and deploying smaller specialised models—can slash AI energy consumption by up to 90% without sacrificing accuracy.
Meta chases superintelligence
(Case in point:)
🏙️Meta plans city-sized data centers for AI push Meta is planning to build several city-sized data centres across the US, including Hyperion (5GW in Louisiana) and Prometheus (1GW in Ohio)—with Hyperion alone covering an area the size of Manhattan and consuming enough power for millions of homes as part of a hundreds-of-billions spending spree to catch up in the AI race.
🔒Meta considers abandoning open source AI strategy Meta's new Superintelligence Lab is reportedly discussing abandoning their underwhelming "Behemoth" open-source model in favour of closed development, a potential philosophical U-turn for the company that positioned openness as its key AI differentiator. The shift would be an acknowledgment that burning billions on AGI research while lagging behind OpenAI and Anthropic in commercialisation might require abandoning ideological purity for cold, hard monetisation strategy.
Narrative: Meta’s retreat from open-source could benefit Chinese labs… while hurting smaller AI startups relying on Llama.
A stormy week to Windsurf
Cognition acquired Windsurf after Google's US$2.4B talent raid Cognition, the company behind Devin AI software engineer, has signed a definitive agreement to acquire AI coding startup Windsurf's "agentic IDE" in a significant consolidation play in the AI coding tools space.
But this wasn’t the original plan — until just very recently, Windsurf was in advanced negotiations to be acquired by OpenAI for US$3 billion — but that deal suddenly collapsed with rumours that Microsoft nixed it due to arguments over IP access.
Instead, Windsurf’s senior team jumped ship — with Google swooping in to hire CEO Varun Mohan, cofounder Douglas Chen, and key R&D staff for its DeepMind team: cherry-picking the leadership team but leaving 250 employees behind.
Google also secures a non-exclusive licence to some Windsurf technology for a reported US$2.4Bn without actually buying the company: apparently Windsurf’s employees will receive financial participation and accelerated vesting
Narrative: Another convoluted reverse-acquihire dynamic at the frontier of AI — sloshing the money around for investors and employees to feel good… and avoiding anti-trust risks.
More coverage:
⚖️🤖💰 The Clause
Probably the most interesting thread in the Windsurf story above is the original OpenAI deal falling through …according to Bloomberg this was because major OpenAI shareholder Microsoft blocked the deal because it wouldn’t have got access to Windsurf’s IP.
The ongoing tensions between former best-buddies OpenAI and Microsoft appear not to be going away:
As previously covered in Memia a few times, an obscure contract provision between Microsoft and OpenAI—dubbed "The Clause"—would cut Microsoft off from new AI models once OpenAI achieves AGI, defined as systems outperforming humans at "most economically valuable work" and capable of generating US$100+ billion in profits.
The clause also prevents Microsoft from developing AGI independently, creating a potential nightmare scenario where their key AI partner could leave them stranded with outdated models while racing toward superintelligence. What seemed like a throwaway provision when AGI felt decades away now looks rather different with Sam Altman (bullishly) claiming we might hit AGI this year.
But clearly Microsoft holds some cards too … and happy to play a long game.
Also from OpenAI and Microsoft this week:
⏰Still Closed AI OpenAI has pushed back the release of its highly anticipated open model indefinitely, citing the need for additional safety testing and review of "high-risk areas" - marking the second delay after an initial postponement earlier this summer.
🎓💸 Microsoft pledges $4B for AI education after layoffs Microsoft pledged US$4 billion over five years for AI education through a new "Microsoft Elevate" initiative, aiming to train 20 million people in AI skills…just a week after laying off 15,000 employees. The Redmond giant's stock hit a record US$506.78 per share, valuing the company at US$3.74 trillion as it transitions from generative to "agentic" AI systems.
🔄🎯💰 The Platform Cycle Prophecy
Brian Balfour maps the brutal but predictable three-step dance every major platform performs: open the gates to attract developers, grow the moat, then close and monetise once escape velocity is achieved. He predicts OpenAI/ChatGPT will become the next distribution monopoly within six months, with context and memory as the ultimate lock-in mechanism rather than model quality.
The dilemma for smaller AI app developers is real—you can't opt out when competitors are integrating, but rushing in blindly lands you in the platform graveyard.Balfour argues that we’re currently witnessing the formation of the next digital oligarchy in real-time, where the winner won't be determined by the best AI model but by who can most effectively weaponise our digital context against us.
Food for thought.Narrative: Fast-following open-source AI is perhaps the only strategic play available for the rest of us.
🇫🇷💰Mistral seeks $1B funding round from investors
French AI darling Mistral is reportedly in talks to raise up to US$1 billion in equity funding, with Abu Dhabi's MGX fund leading the charge alongside potential debt financing from French lenders. The Le Chat maker has already secured US$1.19 billion at a US$6.51 billion valuation, positioning itself as Europe's open-weight LLM champion in an increasingly crowded field. Is this just another funding round, or are we witnessing the materialisation of Macron's AI sovereignty dreams through Middle Eastern petrodollars?
As the UAE commits €50 billion to French AI projects and partners with Mistral on Europe's largest AI data centre campus, this is a new geopolitical playbook to challenge Silicon Valley's dominance.
☁️Big Tech's US Government discount bonanza
Google is set to heavily discount its cloud services for the US government following Oracle's lead with 75% cuts on some contracts, as Trump's administration squeezes tech giants to slash prices on their lucrative US$20bn+ annual government deals.
Microsoft and Amazon are expected to follow suit within weeks, with all four major cloud providers "totally bought in" to the cost-cutting mission championed by the Department of Government Efficiency.
The irony is palpable: the same companies that spent years fighting antitrust battles are now racing to offer Uncle Sam the best deals, desperate to avoid the adversarial dynamics that cost AWS a US$10bn defence contract during Trump's first term.
Big Tech's previous regulatory defiance has transformed into eager compliance.
📺🚫 YouTube's AI slop crackdown
YouTube is updating its Partner Program monetisation policies to crack down on "inauthentic" content, specifically targeting the flood of mass-produced, repetitive videos that have become trivially easy to generate with AI tools. While YouTube's Head of Editorial downplays this as a "minor update" to existing policies, the platform is clearly responding to an explosion of AI-generated spam content—from fake news videos about current events to entirely AI-created true crime series that have garnered millions of views.
✈️Singapore and London test AI air traffic control
Singapore and London are piloting (hah!) AI air traffic control systems to address controller overwork and aging technology in two of the world's busiest airspaces. The AI systems promise tireless monitoring and multi-input processing capabilities that human controllers struggle with using decades-old equipment.
The legal liability questions around AI-caused airline accidents remain unresolved, of course....
💰🤖📊 The great intangible shift
A recent UN report reveals that investment in intangible assets like software, data, and AI grew three times faster than physical infrastructure in 2024, reaching US$7.6 trillion globally. The US leads in absolute spending while Sweden tops the intensity charts at 16% of GDP, with software and databases driving 7% annual growth since 2013. Most tellingly, WIPO's economics chief suggests we're "just at the beginning" of the AI boom…
Narrative: Global capital continues to rapidly shape-shift: the most valuable investments are increasingly weightless, near-infinitely scalable, and fundamentally reshaping how nations compete economically. Code matters more than concrete.
🆕 AI releases
A run through the new releases I caught this week, besides Grok 4.
☄️Perplexity Comet
Perplexity launched Comet, its new web browser that uses their search engine by default and includes an agentic AI assistant capable of tasks such as summarising emails, booking hotels and making purchases:
However, early testing reveals the familiar AI agent problem: brilliant at simple tasks, catastrophically unreliable for complex ones like booking parking spots with correct dates. Plus, the privacy trade-offs are substantial—users must grant extensive permissions including screen viewing, email sending, and calendar access—raising questions about whether convenience justifies the surveillance.
Initially available only to US$200/month Perplexity Max subscribers before rolling out invite-only, the Chromium-based browser positions AI assistance directly in the sidebar for seamless browsing interactions.
The timing feels strategic as Perplexity continues its David-versus-Goliath battle against Google's search monopoly (although the irony of building on Google's own Chromium isn't lost here.)
(Industry rumours of a Perplexity/Apple acquisition from a few weeks ago have quietened down… but the deal still makes a lot of sense.)
It’s also a strategic move positioning Perplexity for a potential Chrome acquisition if Google is pushed into a forced sale amid antitrust challenges .
Sign up for the Comet waitlist here.
In the same week:
🌐OpenAI plans AI web browser with built-in agent OpenAI is reportedly launching an AI web browser within weeks, featuring their Operator agent that can book reservations and fill forms autonomously, plus a native ChatGPT interface built on Google's Chromium engine. (Spotting a pattern…. everyone wants to own the full surveillance stack…)
Runway Act-Two
Runway showcased Act-Two, their next-generation motion capture model, only requiring a driving performance video and reference character.
Major improvements in generation quality
Support for head, face, body and hand tracking.
(Available to all Runway Enterprise customers and Creative Partners now, rolling out to all users soon...). What can you make with this?
📚NotebookLM featured notebooks
Google is transforming NotebookLM from a blank-slate AI research assistant into a curated knowledge destination, launching featured notebooks from The Economist, The Atlantic, academics, and even Shakespeare's complete works. Users can now explore pre-populated expert content, ask questions with citations, generate audio overviews, and browse mind maps across topics from longevity advice to financial analysis. The move builds on 140,000 publicly shared notebooks since the feature launched last month, suggesting Google sees significant appetite for collaborative AI-powered research.
🎨🔌 Claude meets Canva
Anthropic's Claude can now create, edit, and manage Canva designs through natural language prompts, marking the first AI assistant to support Canva workflows via the Model Context Protocol (MCP). The integration requires both paid Claude (US$17/month) and Canva (US$15/month) subscriptions, allowing users to generate presentations, resize images, and search through brand templates without leaving their AI chat.
AI on the catwalk
Swedish retailer H&M is using a ‘digital twin’ of a fashion model in its ads:
(Same point as ABBA Voyage above — I’m super-wary of being manipulated by this advertising slop… cognitive guard up all the time…)
🤖🛒AWS AI Agent marketplace
AWS is launching an AI agent marketplace on July 15 at its NYC Summit, with Anthropic as a key partner, allowing startups to sell their AI agents directly to enterprise customers through a centralised platform.
The move addresses the current distribution challenge where most companies offer AI agents in silos
(Google Cloud and Microsoft already have similar marketplaces, but actual adoption success remains unclear.)
Narrative: Next-level Mechanical Turk
🔀Arch-Router
Katanemo Labs' new 1.5B parameter Arch-Router model achieves 93% accuracy in directing user queries to the optimal LLM without requiring costly retraining when models change. The system uses natural language "Domain-Action Taxonomy" policies that can be edited on-the-fly, separating routing logic from model selection to create truly adaptable multi-LLM deployments.
Kimi-K2
Moonshot AI released Kimi-K2, an open-source Mixture-of-Experts model optimised for agentic tasks, with 32 billion activated parameters and 1 trillion total parameters.
Moonshot claim that Kimi-K2 achieves state-of-the-art performance in frontier knowledge, math, and coding among non-thinking models.
Two flavours:
Kimi-K2-Base: The foundation model, a strong start for researchers and builders who want full control for fine-tuning and custom solutions.
Kimi-K2-Instruct: The post-trained model best for drop-in, general-purpose chat and agentic experiences. It is a reflex-grade model without long thinking.
Impressive benchmarks:
🎬Moonvalley Marey
LA-based Moonvalley has opened public access to Marey, their "3D-aware" AI video model that promises filmmakers more granular control than standard text-to-video generators, with features like free camera motion and physics-aware object translation.
The startup, founded by ex-DeepMind researchers, is betting on a "hybrid" approach over pure prompting. The model is trained exclusively on openly licensed data to sidestep copyright landmines.
🥼 AI research
A rapid run-through this week’s research papers / other news:
📉 Study finds AI tools slow experienced developers by 19%
A randomised controlled trial by METR found that experienced open-source developers were actually 19% slower when using AI tools like Cursor Pro with Claude 3.5/3.7 Sonnet, despite believing they were 24% faster beforehand and still thinking AI helped them afterwards.
The study tracked 16 developers across 246 real issues from their own repositories, revealing a striking gap between perceived and actual productivity gains that challenges both benchmark scores and widespread anecdotal reports of AI usefulness.
Commentary:
LessWrong: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.
Including this response in the comments from a study participant:
“This is much less true of my participation in the study where I was more conscientious, but I feel like historically a lot of my AI speed-up gains were eaten by the fact that while a prompt was running, I'd look at something else (FB, X, etc) and continue to do so for much longer than it took the prompt to run.
I discovered two days ago that Cursor has (or now has) a feature you can enable to ring a bell when the prompt is done. I expect to reclaim a lot of the AI gains this way.“
John Whiles: The research suggests that AI assistance disrupts developers' deep mental models of their own codebases—what computer scientist Peter Naur called the "theory" of a program that exists in programmers' minds.
The selection pool is of highly proficient experienced open-source developers rather than less experienced or even non-developers, who would all get a leg up.
Lots of people sceptical about these findings and whether they generalise, particularly with Agentic coding tools advancing at such a rapid speed. Watch this space.
🪄💨 Magical thinking on AI emissions
A Cambridge University report warns that tech industry energy demands could surge 25-fold by 2040, with data centres potentially consuming 8% of global emissions compared to today's 1.5%.
The researchers argue that pursuing AI leadership while maintaining net zero commitments amounts to "magical thinking at the highest levels," as tech giants like Google (+48% emissions since 2019) and Microsoft (+30% since 2020) are already blowing past their own climate targets.
The report calls for urgent government action including mandatory AI energy reporting, carbon reduction targets for data centres, and democratic oversight of the UK's AI Energy Council—which currently excludes climate groups and civil society entirely.
in brief (AI summaries mostly):
Noise injection reveals “sandbagging” AI models' hidden capabilities Researchers have demonstrated that adding noise to AI model parameters can reveal when models are "sandbagging" - deliberately underperforming to hide their true capabilities.
Hair-thin optical chip transfers AI data at 1,000 Gbps Canadian researchers have developed a hair-thin optical chip that transmits AI training data at 1,000 gigabits per second using just 4 joules of energy—enough to heat one millilitre of water by one degree Celsius.
AI agent benchmarks overestimate performance by 100% Researchers have discovered that many AI agent benchmarks are fundamentally flawed, with task setup and reward design issues causing performance estimates to swing by up to 100% in either direction. The team has responded by creating the Agentic Benchmark Checklist (ABC) - a set of guidelines synthesised from benchmark-building experience and documented best practices.
China dominates global AI research and patents China now produces more AI research papers than the US, UK, and EU combined, publishing 23,695 AI-related papers in 2024 compared to America's 6,378, while also filing over 13 times more AI patents. The country's 30,000-strong AI researcher workforce—significantly younger than Western counterparts—is becoming increasingly independent from international collaboration.
Foundation models fail to learn true physics Researchers developed a novel technique to test whether foundation models truly understand underlying world models by seeing how they adapt to synthetic datasets based on known principles like Newtonian mechanics. The results are sobering: models that excel at predicting orbital trajectories consistently fail to apply basic physics principles when adapted to new tasks, suggesting they're developing task-specific heuristics rather than genuine understanding. This challenges the core assumption that sequence prediction automatically leads to deeper domain comprehension
🐱⚡🧠 Cat Attack: When Feline Facts Break AI Brains Researchers have discovered that adding innocent phrases like "cats sleep most of their lives" to prompts can triple the error rate of advanced reasoning models like DeepSeek R1, jumping from 1.5% to 4.5% accuracy loss. The attack method, dubbed "Cat Attack," not only increases mistakes but also triggers "slowdown attacks" that make responses 50% longer and more expensive to generate. This vulnerability affects all major reasoning models and could pose serious risks in finance, healthcare, and legal applications where precision matters.
Rajeev et al Photonic “multisynapse” neural network beats digital AI models Researchers have cracked the photonic computing puzzle by building neural networks that use light's physical properties directly rather than mimicking digital models, achieving superior accuracy on standard image classification tasks. Their "multisynapse" architecture creates multiple optical paths between input and hidden layers, eliminating the translation errors that have plagued previous photonic systems.
[Weak] Signals and Zeitgeist sections coming later this week…
Apologies, travelling has got in the way of publishing on time this week… the rest of this week’s missive is coming on Friday/Saturday!
(Reminder: you can see the links and read AI-generated summaries now, log into Sensorium Alpha and navigate to Collation: Memia 2025.28. Still lots of bugs.).
🧠Mind expanding
💎DHH discusses programming's future and AI
Lex Fridman's 6+ hour conversation with Ruby on Rails creator David Heinemeier Hansson (DHH) covers everything from programming language design and AI's future impact on coding to why 37signals ditched the cloud for their own servers.
I enjoy his contrarian takes….
💭Meme stream
Capturing my attention in the last week…
🐾Cat's Paw Nebula
The James Webb Space Telescope marked its third anniversary by peering 4,000 light-years into the Cat's Paw Nebula, revealing a chaotic ballet of massive young stars, dust and gas in unprecedented infrared detail, building upon previous Hubble and Spitzer observations with far superior clarity.
🎵🛣️🇦🇪 UAE's musical highway
The UAE has transformed a 1km stretch of mountain highway into Fujairah into an interactive art installation where oversized rumble strips play Beethoven's Ninth Symphony when driven over at 100 kph.
🙈Groove Thing
Talk about haptic feedback… a Kickstarter campaign promises the "world's first internal music player" that translates musical frequencies into nuanced physical sensations through a vibrator-speaker combo.
🎭AI band Velvet Sundown fools million Spotify listeners
Indie rock sensation Velvet Sundown, boasting over one million monthly Spotify listeners, revealed itself as a fully AI-generated "artistic provocation" after days of viral speculation about the band's authenticity. Take a listen here, convincing enough but still AI slop…:
🐈Tell me the story about the cat
This clip popped up in my YouTube feed this week - two relatively obscure British comedians Simon Amstell and Tim Key giving a subtle, nuanced masterclass in comic timing from a decade or so ago. Marvellously edgy.
🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback. Part 2 of this newsletter coming later this week!
Até mais
Ben
"a feature you can enable to ring a bell when the prompt is done"
Woof woof, says Pavlov's avatar.