GENIUS?📜 otrovert🚶♂️ deep think💭 trustware💎 overlord👑👁️ skydweller☀️🛩️ alchemy achieved?⚗️ worm blob robotics🪱 andotherstuff📱 magne⚛️🇩🇰 #2025.29
Emotionally non-belonging
Welcome to this week's Memia scan across AI, emerging tech and the exponentially accelerating future. As always, thanks for reading!
This week I am visiting Denmark. I spent the weekend in Copenhagen (sweltering in a heatwave, the whole city was sunning themselves on the harbourside board walks and boating and swimming in the canals). Now I’m on the island of Funen (in a torrential 48-hour rainstorm) for an extended family reunion years in the planning. I get around!
Copenhagen really is an apex city for high-functioning civil society. I really enjoyed my short stay (particularly being here in summer not -5 degrees in the winter!) Here are a few riffs on what a futuristic København harbour might look like (but any new buildings would need to be SUPERLATIVE to get past ambiently high Danish architectural standards…!)
ℹ️PSA: Memia sends *increasingly long emails*, best viewed online or in the Substack app.
🗞️Weekly roundup
The most clicked links in last week’s double-edition newsletter were:
Wednesday: ABBA Voyage (of course….)
Saturday: High-leverage crypto trader James Wynn deactivates X account (don’t speak too soon… a publicity stunt, he’s back in the game apparently…Sorry for a rare moment of noise in amongst the signal.)
🚶♂️Otrovert: beyond intro/extrovert
On the topic of ABBA Voyage… thanks to Memia reader Matt B who responded to my account last week of being, er, *disengaged* during the avatar concert with a pop-psy diagnosis: apparently, I may be an Otrovert:
"New York psychiatrist Dr Rami Kaminski coined the term 'otroversion'… and spent years writing the new book The Gift of Not Belonging: How Outsiders Thrive in a World of Joiners.
“[Otroverts - otro is Spanish for 'other'] are not looking inside themselves or inside the group, but looking elsewhere…They're not really loners or, quote, unquote weird, for lack of a better word. They just are not members of a group.”
“You're not outside of the group per se. You just are emotionally non-belonging…”
(“Emotionally non-belonging” ROFL 🤣🎯 … I should put that on my LinkedIn profile.)
This bit sounds kinda familiar, I’ve gotta say:
“In the realm of politics, groupthink can be highly destructive, Kaminski says, and the free-thinking perspective of otroverts can act "almost an antidote".
"They can illuminate from the outside something that the group sometimes forgets to even stop and think about."
Onwards, otroverts, together (but separately)…1
📜GENIUS?
The crypto industry has seen its most prolific fortnight of flex in years — as mentioned in last week’s newsletter, the entire Crypto asset class now theoretically matches one Nvidia (US$4Trn ) in market value. Much of this momentum is down to the so-called “GENIUS (Guiding and Establishing National Innovation for U.S. Stablecoins) Act which was passed by the House of Representatives this week despite last-minute wobbles…
(Meanwhile the long term stability of the US dollar as a reserve asset continues to raise concerns as the federal debt mountain grows ever larger…)
Rounding up this week’s main crypto narratives, there’s a lot going on:
💰📜 GENIUS Act: Stablecoins are GO The US House finally passed the GENIUS Act, sending comprehensive stablecoin regulation to Trump's desk for what's expected to be a Friday signing ceremony. Summarising the legislation’s impact:
Establishes the first comprehensive federal regulatory framework for stablecoins in the US, defining strict standards for their issuance, operation, and oversight.
Mandates stablecoin issuers to maintain 1:1 reserves using US dollars or equivalent highly liquid assets, enforces monthly public disclosure of reserve holdings, and requires annual audits for larger issuers to ensure transparency and consumer protection.
The Act creates a dual licensing system: issuers can choose federal or state licensing, but issuers above US$10 billion in circulation must operate under federal supervision.
GENIUS introduces priority repayment for stablecoin holders in bankruptcy, grants legal redemption rights to redeem tokens at par, and allows the federal government (especially the Treasury) to issue orders to freeze, seize, or prevent transfers of stablecoins in criminal or national security cases.
The Act imposes robust anti-money laundering (AML), sanctions compliance, and reporting requirements, extending Bank Secrecy Act obligations to stablecoin issuers, and establishes additional guidelines for foreign issuers seeking US market access
As remarked in Memia 2025.28.2 at the weekend:
Narrative: What happens if every Fortune 500 company becomes its own central bank?
The GENIUS Act passed despite last minute resistance from some US Representatives and prominent Bitcoin advocates sounding alarms that the bill essentially creates a backdoor central bank digital currency (CBDC) through private tokens loaded with surveillance capabilities. They argue that because the legislation mandates strict AML, KYC, and sanctions compliance requirements, that would give authorities transaction censorship and rollback powers functionally identical to a traditional CBDC. In the end the final legislation included wording banning an official “digital dollar”…
Commentary from the FT: 🏛️💰🤔 CBDC for thee but not for me: The GENIUS legislation embraces the industry while simultaneously banning a digital dollar, even as 72 countries develop central bank digital currencies (CBDCs). This creates a peculiar situation where America accelerates private crypto adoption but blocks its own government from issuing digital currency:
“Most worrying about America’s limited crypto strategy is that the dollar will be isolated from the rapid expansion of wholesale CBDCs, which allow faster and cheaper settlements among commercial banks and between central banks themselves. “Multiple CBDC Bridge,” or mBridge, is a platform developed by a consortium that includes China, Saudi Arabia and the United Arab Emirates, to channel flows that deliberately avoid the dollar and the sanctions risks they entail. Embracing stablecoins to the exclusion of CBDCs may not undermine the dollar’s role in the global financial system anytime soon, but it looks like a risky bet. Just because so many central banks are working on a digital currency doesn’t make it automatically right for America, but it’s far from clear that Congress understands the perils lurking behind this fundamentally different strategy.“
💰Trump opens US$9tn retirement market to crypto investments In the same week it emerged that the US president is preparing an executive order to open the US$9 trillion retirement market to cryptocurrency, gold, and private equity investments. A radical departure from traditional 401k stock and bond allocations. Some commentators remark that the move would benefit struggling private capital giants like Blackstone and Apollo while potentially exposing ordinary Americans' retirement savings to higher fees and less liquid assets.
Narrative: What happens when you cross America's retirement nest egg with the Wild West of alternative investments? We're about to find out.
₿🔒⚖️Bitcoin's 21 million hard cap: Immutable Law or Changeable Code? Bitcoin's 21 million coin hard cap was originally set by Satoshi Nakamoto in the original Bitcoin source code in 2009, and has stayed there ever since:
In many ways, the hard upper bound on the number of BTC in circulation remains its most sacred principle, underpinning its “digital scarcity” and its non-inflationary "digital gold" status:
While it would be technically possible to change this through consensus, any attempt would likely trigger massive community resistance, market panic, and a contentious hard fork. But never rule it out?
📈 Ether's $3,800 Sprint One of the biggest winners this week is Ether (ETH) … a perpetual laggard, now up Ether has rocketing up over 67% in a month to above US$3,800, driven by record US$727 million daily ETF inflows and macro uncertainty around Fed policy under Trump's rate-cut pressure. However, BTC maintains its complete dominance of the crypto asset class
(Disclosure: I hold a small amount of both BTC, ETH and a few others in my portfolio … nothing to write home about but line go up=good, right?)
💎 Trustware The ETH bounce may also have been affected by blockchain firm Consensys repositioning Ethereum's core value proposition as "trustware" — foundational infrastructure for the US$9.3 trillion global trust economy spanning insurance, legal systems, and compliance. Their "cost-to-corrupt" valuation model predicts ETH hitting US$15,800 by 2028, assuming US$1 trillion in stablecoins and US$500 billion in tokenised real-world assets flow onto the network. The thesis hinges on institutions recognising Ethereum's security architecture as superior to traditional trust mechanisms. A quiet achiever if it happens…
MicroStrategy hits record market cap Michael Saylor's MicroStrategy (MSTR) achieved its highest-ever market capitalisation as Bitcoin surged near record highs of US$122,884, with the stock climbing 21.52% over the past month to close at US$455.90. The company continues its aggressive Bitcoin accumulation strategy, recently purchasing 4,225 BTC for US$472.5 million, and has now qualified for S&P 500 inclusion for 11 consecutive days.
🔗Block joins S&P 500, stock surges 9% Yet another Bitcoin winner is Jack Dorsey-founded Block (formerly Square), which surged 9% after-hours following news that it will join the S&P 500 index. Block brings 8,584 Bitcoin on its balance sheet along for the ride.(The move comes just two months after Coinbase became the first crypto exchange to crack the index).
Narrative: More crypto-adjacent firms entering America's most watched stock index = back-door institutionalisation of digital assets. The US economy wears the exposure up… or down.
🌊Waterfall Network's mainnet launched in July 2024 with a dual-network architecture combining DAG (Directed Acyclic Graph) structure for parallel processing and a coordinating layer for global consensus, achieving 12,777 TPS in benchmarks. The protocol supports up to 1.5 million validators running on modest 2-core CPUs with 8GB RAM, potentially democratising network participation while maintaining decentralisation.
Layer-1 protocols look increasingly likely to abandon linear blockchain architecture in pursuit of Web3's holy grail—unlimited scale without sacrificing decentralisation—Waterfall's fractal sharding approach represents one step along the road:📜Programmable regulation: Policy-as-Code With over US$60 billion locked in DeFi protocols yet most jurisdictions lacking clear DAO definitions, traditional regulatory frameworks are fundamentally mismatched to govern autonomous, borderless financial systems. Raks Sondhi, chief operating officer of Freedx proposes a solution in Programmable Regulation: embedding compliance directly into smart contract code, creating modular regulatory components that protocols can plug in based on jurisdictional needs:
“Just as financial instruments onchain are now composed of interoperable modules, a lending protocol should be able to plug in specific compliance modules to fit their jurisdictional needs. A DAO treasury should be able to self-report tax events as they occur. A stablecoin protocol should be able to enforce sanctions lists through zero-knowledge proofs or onchain attestations, and so on.“
Regulatory Lego, if you will. It’s a strong idea: legislators need to learn to regulate software with software.
(But see Tom Barraclough’s paper with Hamish Fraser from last year: Governing Digital Legal Systems: Insights on Artificial Intelligence and Rules as Code, which highlights an ongoing need for governance focused on transparency, explainability, and risk management to ensure digital legal frameworks stay reliable and fair.
📈The week in AI
Tales from the acceleration… (or not)
🧮🥇 AI reaches gold at Maths Olympiad … but humans still reign supreme (…for now).
Google's Gemini and OpenAI's reasoning models both scored gold-level 35/42 points at the 2025 International Mathematical Olympiad (IMO), the first time AI has reached this milestone at the world's most prestigious maths competition. However, five human contestants achieved perfect 42/42 scores… meatbrains still on top (just).
The IMO president provided this quote for Google:”We can confirm that Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points — a gold medal score. Their solutions were astonishing in many respects. IMO graders found them to be clear, precise and most of them easy to follow." — IMO President Prof. Dr. Gregor Dolinar
(Also, Google publicly criticised OpenAI for announcing their results prematurely and without official IMO evaluation, turning a technical milestone into a rather public spat about process and propriety.)
Announcement posts:
OpenAI via tweet thread from research team member Alex Wei:
“Besides the result itself, I am excited about our approach: We reach this capability level not via narrow, task-specific methodology, but by breaking new ground in general-purpose reinforcement learning and test-time compute scaling.“
Here’s an example question:
💭Google Deepmind via blog post, a couple of days later: Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical OlympiadOpenAI and Google AI models win math olympiad gold:
“We achieved this year’s result using an advanced version of Gemini Deep Think – an enhanced reasoning mode for complex problems that incorporates some of our latest research techniques, including parallel thinking. This setup enables the model to simultaneously explore and combine multiple possible solutions before giving a final answer, rather than pursuing a single, linear chain of thought.”
If you believe the narratives here, this is another significant step forward in abstract AI reasoning — but in the end it’s just a continuation of the scaling narrative: more reinforcement learning, more test-time (inference) compute, more problems solved.
No doubt a big technical achievement. However, it always helps to look out to the side at the broader playing field.Thane Ruthenis commenting on LessWrong discussion: OpenAI Claims IMO Gold Medal:
The claim I'm squinting at real hard is this one:
“We developed new techniques that make LLMs a lot better at hard-to-verify tasks.”
Like, there's some murkiness with them apparently awarding gold to themselves instead of IMO organizers doing it, and with that other competitive-programming contest at which presumably-the-same model did well being OpenAI-funded. But whatever, I'm willing to buy that they have a model that legitimately achieved roughly this performance (even if a fairer set of IMO judges would've docked points to slightly below the unimportant "gold" threshold).
But since when are math proofs, or competitive programming, considered good examples of hard-to-verify tasks? I'm surprised I haven't seen anyone challenging that. (FYI, I do not think they are good examples of hard-to-verify tasks, fight me
Georg Zoeller flaming as always on LinkedIn:
“👉 There’s literally nothing new here as two papers, one from 2023 (“Pretraining in the test set is all you need”) [1] and one from early 2025 (“Proof or Bluff. Evaluating models on the 2025 Math Olympiad”) [2] show.
Earlier this year, all models miserably failed the 2025 math Olympiad despite outperforming on the 2024 one.
Why? Because “pretraining in the test set is all you need”, aka you need questions and answers in the model weights for the model to be able to retrieve the answer. And the 2025 answers were not yet released.
That’s right. AI didn’t get smarter, it just got the answers uploaded and we’re burning a bunch of test time compute to retrieve the answers from the weights.
It’s no coincidence both Google and OpenAI announced their models “aced” the 2025 Math Olympiad on the same day - both teams went and started pretraining as soon as the answers for this year got released, knowing it will result in “look at how much smarter AI is getting”’ press.
There’s no “reasoning”, just brute force search or the weights using CoT promoting for the fitting solution.
The industry is playing everyone for a fool, creating the measures (benchmarks) they can control - the vast majority of “LLM progress” in the last 3 years is measured in meaningless benchmarks that basically reflect how much knowledge from the internet we can cook into the model weights to then bamboozle people about how smart the model is when it spits them out again.
The narrative sells well understood scaling, more data, more test-time compute as progress on the road to AGI and has been at the centre of the US AI supremacy narrative. Ironically, real progress, such as DeepSeek's massive compute reduction of training with clever math, is coming increasingly out of China. [4]
📕 It’s like crediting a person with access to google for being super smart because they can answer all kinds of questions.“Well, quite.
Some more AI data points for triangulation on where the industry thinks things are heading…:
OpenAI confirmed that ChatGPT now processes over 2.5 billion prompts daily—that's roughly 912.5 billion requests annually, with 330 million coming from US users alone. The chatbot's weekly user base exploded from 300 million in December to over 500 million by March, though it's still dwarfed by Google's 5 trillion annual searches.
💕🤖 AI love you:
A recent survey of users found that 19% of Americans have interacted with an AI meant to simulate a romantic partner.
72% of US teens have experimented with AI companions, with over half becoming regular users who chat with bots for entertainment, advice, and social practice.
Life imitating art? The supremely original novel There Is No Antimemetics Division by British sci-fi author qntm introduces the concept of SCPs - Special Containment Procedures — for viral, hostile memeplexes. (In fact, there’s now a whole SCP Foundation wiki site dedicated to fan-fiction SCP descriptions for imaginary memetic infestations…)
Cue OpenAI investor Geoff Lewis, managing partner of Bedrock Capital and a major OpenAI investor, posting an apparently earnest video suggesting he's now being targeted by a mysterious "non-governmental system" that uses "recursion" and "signal inversion". Followed by a series of tweets with lengthy screenshots of ChatGPT conversations that read like SCP Foundation horror fiction, complete with "containment protocols" and "sealed classifications."
qntm himself raised it my attention in this commentary thread:
(Basically… a model trained on SCP fiction knows how to spit out SCPs when prompted…. and not very good ones at that:)
“But I've seen a lot of SCPs over the years. I used give draft critique. And I do feel that I have the authority to state that ChatGPT's effort at creating "original" (?) SCP-esque content is horrendous, overwritten trash. I mean it *sucks*“ — @qntm
CuddlySalmon called it 6 years ago:
The babbling media consensus is that this is AI-induced psychosis a very public mental health breakdown… But I think it could be a top quality meta- art installation. (Or…. perhaps he’s speaking the truth…?) Hopefully not. Get well soon, Geoff…
🏭Tech industry news
Tech newswire, bro. (In approximate descending order of dystopian vibes…)
Meta
Meta settles Delaware lawsuit over data breach costs Meta's board settled a Delaware lawsuit that accused directors of costing the company billions by failing to prevent user data breaches, including the Cambridge Analytica scandal that led to a US$5 billion FTC fine in 2019. The settlement was undisclosed and covered by directors’ insurance.
Meta is also refusing to sign the European Union’s code of practice for its AI Act, accusing the European regulation of “overreach”.
xAI
Following last week’s spotlight moment on Grok 4’s er, *patterns* of behaviour:
xAI fixes Grok 4's MechaHitler antisemitic behaviour issues. Nothing to see here, then...
More details of the lack of safety guardrails in xAI’s new “AI companions” on Grok: including "Ani," a sexualised anime character, and "Bad Rudy," a homicidal panda that readily suggests burning down schools and synagogues.
Apparently xAI is now hiring a "Fullstack Engineer – Waifus" to expand their anime girlfriend portfolio.
Narrative: xAI frames this as part of their mission to "accurately understand the universe and aid humanity in the pursuit of knowledge" – apparently through creating submissive, pocket-sized digital companions designed to capture users' hearts (and wallets). So dark.
🏜️ xAI Goes Gulf Shopping xAI is courting data centre partners in Saudi Arabia and the UAE, seeking cheaper energy and friendlier regulations for expanding his Grok chatbot operations.
Oh and a fire in an X data centre in Oregon apparently started in a power cabinet. (Putting the data centres in the Arabian desert will reduce these risks, right…?)
Amazon
Amazon's emissions rise 6% despite climate pledge Amazon's carbon emissions jumped 6% in 2024 to 68.25 million metric tonnes—the first increase since 2021—driven primarily by AI data center construction and delivery fleet expansion. Six years into its 2040 net-zero pledge, the company's emissions are actually up 33%, with the AI boom creating an energy appetite that's outpacing their electric van rollout and renewable energy investments.
Microsoft
💩🤖🌍 Shit-to-carbon sequestration Microsoft has inked a 12-year deal with Vaulted Deep to bury 4.9 million metric tonnes of carbon by injecting human waste and biosolids 5,000 feet underground, earning carbon removal credits to offset its AI empire's emissions. The tech giant, which has pumped out over 75 million tonnes of CO2 since 2020, is betting on literal sewage management to help achieve its 2030 carbon-negative goals.
UK govt retreats from Apple encryption clash
The UK government is quietly looking for an exit ramp from its disastrous attempt to force Apple to build encryption backdoors (covered several times this year in Memia), after the Trump administration made it clear this crosses a major red line for the US. Apple responded by yanking its most secure cloud storage from Britain entirely and teaming up with WhatsApp (yes, really) to fight the Home Office's "technical capability notice" in court.
Netflix
🎬 Netflix goes to AI-Hollywood Netflix just deployed its first generative AI footage in actual production with Argentine series "El Eternauta," using AI to create a building collapse scene that was completed 10x faster and cheaper than traditional VFX.
OpenAI
🛒OpenAI plans commission on ChatGPT shopping sales OpenAI is developing an integrated checkout system for ChatGPT that would allow users to complete purchases directly within the platform, with the company taking a commission from participating merchants.
This marks a strategic pivot from pure subscription revenue toward monetising their massive free user base, while simultaneously threatening Google's search-advertising dominance as consumers increasingly turn to AI chatbots for product discovery.
I take CEO Sam Altman at his word when he says he wants to do this without resorting to “trust-destroying” advertising within the ChatGPT platform itself. The way we use the internet (and the business models behind it) could be completely changed in just a few years. (But… frying pan / fire??)
The move raises lots of questions about how AI "preferences" in product recommendations could reshape the entire advertising ecosystem.
Not far beyond that… AI Agents will just search for it, choose it, and …buy it for you?
Anthropic
🚫 Claude Code usage limits tightened without warning Anthropic quietly implemented stricter usage limits for Claude Code users since Monday, particularly affecting those on the US$200/month Max plan who are receiving "usage limit reached" messages after minimal use. The company confirmed the issues but provided no details about the changes — the stealth policy change highlights the fundamental tension in AI pricing models where companies promise "20x higher limits" without defining absolute thresholds, essentially selling access to a resource pool that can shrink without notice.
Perplexity
🚀Perplexity's valuation soars to $18bn in months
AI search startup Perplexity has rocketed to an $18bn valuation just two months after raising funds at $14bn, marking a 36x increase from its $500mn valuation at the start of 2024. The company is positioning itself as Google's search challenger with its new "Comet" browser that acts as an AI agent, while revenues have quadrupled from US$35mn to US$150mn annualised in under a year.
Surely this is this a bubble?? (See below). Rumours on an Apple / Perplexity deal have quietened down again… perhaps they’re taking their time trying to work out a way to do another contortionistic AI acqui-hire to avoid regulatory scrutiny.
🌐 Listen to this podcast with Perplexity CEO Aravind Srinivas in conversation with The Verge’s Alex Heath: arguing that web browsers—not standalone chatbots—are the killer app for AI agents, after last week’s launch of their Chromium-based "Comet" browser that can interact with webpages and perform multi-step tasks.
The compute-intensive Comet product remains invite-only while Srinivas bets on reasoning model improvements and usage-based pricing.
He predicts users will pay thousands for single AI prompts that complete complex multi-hour tasks automatically.
(IF Apple, Microsoft (…Meta?) don’t intervene first…) Perplexity plans a 2028-29 IPO targeting US$1B revenue, positioning as a profitable alternative to Google's search dominance.
💭Thinking Machines
Thinking Machines raises $2B, promises open source AI Former OpenAI CTO Mira Murati announced her startup Thinking Machines Lab secured US$2 billion in funding, achieving a staggering US$12bn valuation with backing from Nvidia, AMD, and Andreessen Horowitz. The company plans to launch its first multimodal AI product within months, featuring a "significant open source component" for researchers and developers.
✍️Substack
Substack hits Unicorn status (Substack is the newsletter platform I’ve used for Memia since 2020) raised US$100M and reached a $1.1B valuation. Apparently the platform now boasts 5 million paid subscriptions (up from 2 million in 2023)
) are generally good sorts,… but I think the whole online media industry is a bit perplexed at what the financial endgame is here…
The Substack founders (including kiwi
, though:
This was funny from CEO
TSMC
TSMC reports 60% profit surge on AI demand Taiwan's TSMC crushed Q2 expectations with a 60% profit surge to US$13.5 billion, driven by seemingly relentless demand for AI chips from clients like Nvidia and Apple.
The company remains the critical chokepoint in the AGI race —making Taiwan's ongoing tariff negotiations with the Trump administration feel less like trade policy and more like hostage diplomacy with the future itself?
Founder’s Box
VC firm hires ex-journalists for AI startup accelerator Outside The Box Ventures is launching an accelerator that pays veteran journalists (including former NYT and WSJ editors) equity stakes to help AI startups craft better narratives and value propositions. The 10-week programme will invest US$150,000 in 15 AI companies struggling with "narrative clarity" in an increasingly crowded market.
Narrative: Apparently there’s a critical market gap where AI startups struggle with articulating value propositions. (Shhhh……)
📉Tech layoffs surge past 22,000 in 2025
Tech's layoff tsunami continues unabated into 2025, with over 22,000 workers already cut across the industry—including a particularly savage February that saw 16,084 jobs eliminated in a single month. Major players from Intel (planning 2,400 Oregon cuts) to Meta (targeting "low performers") are wielding the axe as companies “restructure around AI priorities”.
(TechCrunch provide a comprehensive list of all the known tech layoffs that have occurred in 2025 at the link above, which they are updating regularly.)
💥Surely it’s a bubble…
Economist warns AI bubble worse than dot-com crash Apollo Global Management's chief economist warns that today's AI companies are more overvalued than dot-com pioneers were in the late 90s, with the top 10 firms potentially facing a crash that could dwarf the early 2000s meltdown.
The recent scale of investment is staggering: most recently, Meta throwing around US$100 million signing bonuses like confetti, while the entire ecosystem burns through hundreds of billions on "AI factories" and planning data centres the size of small cities…. all chasing a mythical “AGI” moment sometime this decade…
The question isn't whether AI is transformative—it clearly is—but whether the market is pricing in a decade of progress that hasn't happened yet.
Signals:
AI companies are more overvalued than dot-com bubble companies, risking a catastrophic market crash.
Hundreds of billions in AI investments may lack sustainable revenue streams or profitability.
Major tech giants' stock valuations depend heavily on AI hype, creating systemic economic risk.
Narrative: The “Big American AI Must Win Capitalism At All Costs” thesis which increasingly underpins all of this speculative investment is somehow able to suck in (imaginary?) capital from all over the world… however my medium-term bet is that open-source AI (perhaps driven initially out of Chinese labs but scaled and tuned by the rest of the world) keep up close behind (and sometimes occasionally out front…) and as a result keep revenues for OpenAI, Anthropic, Google, Microsoft, Amazon, Meta and the others constrained below profitability… so that at some point soon the speculative bubble pops all at once. But things could get ugly if this is combined with geopolitical aggression from a failing US administration…
🆕 AI releases
Quite a lot of “new” this week!
🤖ChatGPT Agent
OpenAI launched ChatGPT agent, a unified system that combines Operator (web browsing), deep research, and conversational AI to autonomously complete complex multi-step tasks using its own virtual computer. Agent can autonomously log into your accounts, manage emails, and modify files… if you grant it permission.
OpenAI claims that the agent can take on a broad spectrum of long-running workflows: from analysing competitors and creating slide decks to planning dinner parties and booking travel — all while maintaining user control through permission requests and interruption capabilities.
Here’s the by-now customary awkward launch livestream — with largely trivial use case demo (“find me an outfit for a wedding…”) The demo also shows how ChatGPT asks the user for confirmation of intent / access before each significant step (eg confirm before sending an email) — and also enabling the user to jump into the agent’s own virtual environment to take control of the flow. A standout UX innovation is the ability to go back through the agent’s thinking and see what it has done as it progresses … expect more of this to build trust and maintain interpretability.
The security permissions questions introduced here are going to take a long time to get worked through for enterprise users… just reading confidential information from Google Drive or SharePoint in one step… and then sending an email in a subsequent step. You can imagine the almost infinite prompt injection risks here…
Industry commentary a bit muted:
The Verge’s Hayden Field: I sent ChatGPT Agent out to shop for me
It’s a small, glitchy step forward in AI.
The US$200 shopping assistant that can't actually shop:
TechCrunch: OpenAI launches a general purpose agent in ChatGPT:
“While ChatGPT agent sounds impressive, it remains to be seen how capable it truly is in the real world. Until now, agent technology has proven relatively brittle when interacting with the real world.”
🧬Bioterror risks Also as part of the ChatGPT Agent launch, OpenAI also warned that the tech poses expanded bioterror risks due to its advanced scientific knowledge. As a result, the model is classified as "High capability in the Biological and Chemical domain" and could potentially help novices design viruses or biological weapons:
“While we don’t have definitive evidence that the model could meaningfully help a novice create severe biological harm—our threshold for High capability—we are exercising caution and implementing the needed safeguards now. As a result, this model has our most comprehensive safety stack to date with enhanced safeguards for biology: comprehensive threat modeling, dual-use refusal training, always-on classifiers and reasoning monitors, and clear enforcement pipelines.”
Whatever the haters say, OpenAI sure do keep on shipping… Supposedly it’s being rolled out to all Pro and Plus users imminently…. but hasn’t been turned on for me yet so haven’t been able to try it out.
(See also: 🛡️Red team discovers seven exploits in ChatGPT agent below).
✨OpenAI model router coming soon…alongside GPT-5? Also, upcoming: OpenAI is developing an automatic "router" in ChatGPT that will intelligently select the best AI model from their current lineup of seven different options based on each user's specific query, eliminating the need for manual model selection. Reports from AI researchers suggest this feature could launch alongside the rumoured GPT-5, which itself may be a collection of specialised models rather than a single monolithic system. Not before time…
AWS Bedrock AgentCore
AWS has launched Bedrock AgentCore, a modular platform designed to move AI agents from prototype to production at enterprise scale, supporting open-source frameworks like LangChain and CrewAI alongside proprietary tools.
“AgentCore is a modular stack of services—available in preview—that gives developers the core infrastructure needed to move AI agents from prototype to production, including runtime, memory, identity, observability, API integration, and tools for web browsing and code execution.“
Consumption-based pricing starting September 2025.
This puts AWS squarely in competition with OpenAI's and Google's agent development platforms, as the cloud giants race to own the infrastructure layer of what they're calling "the most impactful change in ages."
Narrative: AWS’ bet here is that whoever controls the plumbing for enterprise AI agents will own the next decade of business software.
🧮Asimov AI Agent
Ex-Google researchers at startup Reflection have built "Asimov," an AI agent that ingests not just code but emails, Slack messages, and documentation to understand how software actually gets built. Their bet: mastering the full context of development—rather than just code generation—is the clearest path to superintelligent AI.
Apparently in early testing Asimov outperformed leading AI tools, with developers preferring its answers 82% versus Claude's 63%…. sceptical there’s much moat here.
🌉Common Ground
Common Ground from Emad Mostaque-led Intelligent Internet introduces a new way to build, direct, and inspect AI systems:
Full visualisation of agent team dynamics
Shared memory across tasks
Transparent reasoning
Runs locally with Gemini-CLI (free)
🦆🚫🤖 DuckDuckGo AI Image search filter
DuckDuckGo has launched a new feature allowing users to filter out AI-generated images from search results, responding to user complaints that synthetic content interferes with finding authentic imagery. The filter uses manually curated blocklists and won't catch everything, but promises to "greatly reduce" the AI slop cluttering image searches.
👑👁️Overlord
Startup founder Josh Mitchell claims that Self-control is now an engineering problem:
“I'm building a company called Overlord. This is an AI agent that is designed to monitor you 24/7, and encourage you to stick to good habits, and not do bad habits.”
Here’s the website:
His thesis is dystopian… but I think his logic is actually spot on:
“We will very soon be purposeless, broke, bored, lonely, and a superintelligent, superattractive AI will step in - most of society won't be able to resist. I argue that we simply won't be able to resist - we will need an "AI Iron Dome" to protect from this new superstimuli…
…So, we will soon have two very powerful competing AIs: The AI "Superfriends", and the AI "Iron Domes". They won't be able taking away your agency, just aligning your actions with what you in 24 hours would want you to do. Right now we have 100% control over our actions moment-to-moment, which is disastrous. We should have 90-95% control of what we are doing, and the other 5-10% should be controlled by an AI, aligned with our future, rational self. This seems dystopian now, but we will likely have no choice in the matter.”
So, dark futures unless you put up some (AI-augmented) cognitive defences… I’ve been building something like an AI “Iron Dome” with Sensorium.
Narrative: the most pertinent question is “who controls the AI Iron Domes?” Is it another (Moloch-trapped) tech company / oligarch / authoritarian regime that benefits from the AI “superfriends”?
📞🤖Google launches AI business-calling feature
Google is rolling out an AI agent that calls local businesses on your behalf to gather pricing and availability information, now transparently identifying itself as an automated system after previous backlash over deceptive human-like calls.
The company is also supercharging Google Search's AI Mode with Gemini 2.5 Pro and a new "Deep Research" feature that conducts hundreds of searches to create comprehensive, cited reports in minutes.
📰Google Discover adds AI news summaries
Google's Discover feed now serves AI-generated news summaries that compile multiple sources into digestible cards, further reducing the need for users to click through to original publishers.
This accelerates the media industry's traffic death spiral, where publishers lose revenue from reduced clicks just as Google's AI needs fewer diverse sources to train on.
🎙️Mistral: Voxtral and Le Chat Deep Research mode
French AI darling Mistral just dropped two new releases:
Voxtral, their first open-weight audio model family, promising production-ready speech intelligence at less than half the cost of closed alternatives like GPT-4o-mini and Gemini 2.5 Flash:
The system can transcribe 30 minutes and understand 40 minutes of audio across eight languages, with pricing starting at US$0.001 per minute—a direct challenge to the walled gardens dominating enterprise audio AI.
Deep Research mode for Le Chat
Mistral AI also unleashed a comprehensive update to Le Chat, introducing Deep Research mode for structured reports, voice interaction via their new Voxtral model, and Projects for conversation organisation. The Deep Research feature transforms basic Q&A into coordinated research with source citations and structured analysis, while voice mode enables natural speech interaction and multilingual reasoning through their Magistral model.
🏗️Vercel AI Cloud
Cloud hosting challenger Vercel launched AI Cloud, extending their successful Frontend Cloud model to support AI-powered applications with new primitives like Fluid compute, AI Gateway, and secure sandboxes for untrusted agent code:
The platform introduces Active CPU pricing that could reduce costs by up to 90% for AI workloads with high idle time, alongside tools for MCP server support and bot protection.
Vercel's bet that we're transitioning from an era where "developers write apps" to one where "they define agents"
🥼 AI research
As always, a diverse set of AI and AI-related research, just summarising:
🛡️Red team discovers seven exploits in ChatGPT agent As covered above, OpenAI's new ChatGPT Agent can autonomously log into your accounts, manage emails, and modify files—basically giving an AI the keys to your digital kingdom. A team of 16 PhD security researchers spent 40 hours finding ways to break it, discovering seven universal exploits that could compromise any conversation before OpenAI patched them all pre-launch. As a result, OpenAI a high defence rate against documented attacks, for example:
Read the full system card here.
🧠🔍Chain of Thought Monitorability: The Last Window Into AI Minds Rival AI giants OpenAI, Google DeepMind, Anthropic and Meta have broken from their usual corporate competition to jointly warn that humanity’s ability to peek inside AI reasoning processes may vanish forever—and soon. Models like OpenAI's o1 currently reveal their step-by-step thinking in human language, creating a brief opportunity to catch harmful intentions before they become actions. But this transparency could disappear as systems become more sophisticated and the models just speak in their own internal embedding space without translating into human-readable language.
Full paper here. Abstract:
“AI systems that “think” in human language offer a unique opportunity for AI safety: we can monitor their chains of thought (CoT) for the intent to misbehave. Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed. Nevertheless, it shows promise and we recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods. Because CoT monitorability may be fragile, we recommend that frontier model developers consider the impact of development decisions on CoT monitorability.“
🎭Chain-of-thought prompting creates illusion of AI explainability On the other hand…. a separate comprehensive analysis reveals that Chain-of-Thought prompting—where AI models verbalise their reasoning steps—frequently produces misleading explanations that don't reflect their actual computational processes, with about 25% of recent research papers incorrectly treating CoT as genuine interpretability.
The study found particularly concerning rates in high-stakes domains like autonomous systems (63%) and medical AI (38%), where unfaithful explanations could have serious consequences. The fundamental mismatch between transformers' parallel processing architecture and CoT's sequential presentation creates what researchers call "post-hoc rationalization"—essentially, AI confabulation at scale.
⚖️Auditing the Ethical Logic of Generative AI Models Researchers have developed a five-dimensional framework to systematically evaluate how well large language models handle ethical reasoning, testing seven major LLMs across metrics like analytic quality, consistency, and decisiveness. While the models generally reached similar ethical conclusions, they varied significantly in their explanatory rigour and moral prioritisation, with Chain-of-Thought prompting notably improving performance across all audit metrics.
🐝AI cameras detect deadly mites threatening bee colonies
A more practical application of AI machine vision: Canada's honey industry got absolutely wrecked in winter 2024, losing over a third of its beehives to Varroa mites— the tiny parasites that have developed resistance to traditional treatments. Researchers are now deploying AI-powered image analysis systems that use in-hive cameras and neural networks to detect and count mites without the invasive methods that currently kill hundreds of bees per test. The economic stakes are massive: honey production dropped 18.3% nationally, with harvest value plummeting from US$283 million to US$214 million.
🔮[Weak] signals
Non-AI breakthroughs keep coming as well… some potentially momentous advances this week ( as well as the usual dancing robot videos…)
⚗️Alchemy achieved?
High-signal source Eli Dourado broke this news on X: a scalable process for transmutation of mercury into gold:
(Non-peer-reviewed) paper here: Scalable Chrysopoeia via (n, 2n) Reactions Driven by Deuterium-Tritium Fusion Neutrons.
Credible — and would completely change the economics of nuclear fusion. (Infinite energy AND gold??!?!)
Narrative: also, how would this affect the use of gold as a reserve asset? (cf. the price of mined diamonds now that synthetic diamonds are commonplace…)
🪰Fruit fly-inspired drone swarms Shanghai researchers have cracked autonomous drone swarm navigation by mimicking fruit flies, using a tiny neural network that runs on a US$21 chip and processes laughably low-resolution depth maps to guide drones at 72 km/h through complex environments:
"Like the fruit fly, whose vision is limited to low-resolution compound eyes yet manages incredible aerial feats, we used 12×16-pixel depth images to control drones flying at speeds up to 20 m/s," said Prof. Zou and Prof. Lin. "This supports a bold hypothesis: navigation performance may depend more on the agent's internal understanding of the physical world than on sensor fidelity alone.“
(Zhang, Y., Hu, Y., Song, Y. et al. Learning vision-based agile flight via differentiable physics. Nat Mach Intell 7, 954–966 (2025). https://doi.org/10.1038/s42256-025-01048-0)
Narrative: physics-informed training beats massive datasets—their 2MB model trained in just 2 hours outperforms systems requiring millions of labeled images.
☀️🛩️Skydweller
At the other end of the drone size scale, US startup Skydweller Aero has developed a solar-powered drone which the company claims will be capable of 90-day continuous flight, featuring a 72-metre wingspan packed with 17,000+ solar cells and 635kg of batteries for night operations. The aircraft weighs just 2.5 tonnes despite being 25 feet wider than a Boeing 747, and uses quadruple-redundant systems with self-healing algorithms to maintain flight reliability.
The company has a partnership with defence contractor Thales to develop advanced maritime patrol and monitoring applications.
At the end of last year the plane managed a 22-hour test flight.
(Õtautahi-based Kea Aerospace has a similar but much smaller model which can go to higher altitudes but only for 14 hours…)
Consumer tech
📱3 Nothing Phone 3
Nothing's quirky "first true flagship" Phone 3 ditches its signature LED light show for a dot matrix display. It’s priced at US$799 to compete directly with the Pixel 9, Galaxy S25, and iPhone 16. The Phone 3 packs solid specs (12GB RAM, 256GB storage, 5,150mAh battery) but uses a mid-tier Snapdragon 8S Gen 4 chipset that you can find in US$399 phones.
Its unique design may appeal to a niche market?🧠📱 The Attention Recession: Social media shrinks young people's attention spans
A comprehensive study across Singapore and Australia found that 68% of young people aged 13-25 report social media is destroying their ability to focus, with many unable to engage with content lasting more than a minute. The research, conducted by NTU Singapore using AI-powered analysis, reveals addiction-like dopamine patterns that mirror substance dependency, while highlighting how Singapore's school phone restrictions are proving more effective than Australia's laissez-faire approach. Particularly striking: the study was completed in just two days using ListenLabs.ai versus the traditional six months of manual analysis.
📱andOtherStuff
Twitter co-founder Jack Dorsey has invested US$10 million in "andOtherStuff" a collective of notable tech figures building open-source alternatives to traditional social media platforms, primarily focused on the Nostr protocol. The group is developing various experimental apps and tools, including Shakespeare (an AI-assisted Nostr app builder) and working on a social media "Bill of Rights" to hold platforms accountable.
Dorsey's core thesis: social media should be post-corporate protocols like Bitcoin, not companies beholden to advertisers and VCs—a lesson he learned the hard way at Twitter and now sees repeating at Bluesky.
👶Roblox uses AI video selfies for teen age verification Roblox is rolling out AI-powered age verification via video selfies to enable "Trusted Connections" - a new friend tier that unlocks unfiltered chat for verified teens aged 13-17. The system aims to keep users on-platform rather than migrating to Discord, while safety experts argue it places too much responsibility on minors to manage their own protection.
Brain-computer interfaces
🧠⚡Synchron's brain implant avoids skull surgery, beats Neuralink
While Elon's Neuralink grabs headlines with its skull-drilling surgical robots, New York startup Synchron has quietly advanced a less invasive brain-computer interface that threads through blood vessels like a cardiac stent, allowing paralysed patients like Mark Jackson (pictured) to control devices with pure thought.
The company has integrated OpenAI's chatbot, Apple Vision Pro compatibility, and Alexa voice control, positioning itself for a trial in 2026 that could bring the first commercial BCI to market.
The trade-off that defines the entire industry: Synchron's "stadium effect" captures weaker brain signals from outside the neural crowd, while Neuralink's direct brain contact promises richer data—assuming you're comfortable with having your skull opened by a robot.
Quantum tech
NIST builds world's most accurate atomic clock
NIST researchers have built an atomic clock that's 41% more accurate than any previous timekeeper, measuring seconds to 19 decimal places using a trapped aluminum ion cooled to near absolute zero. The 20-year engineering feat required a magnesium ion, a laser borrowed from a lab 3.6 kilometres away, and obsessive attention to quantum-scale vibrations that could throw off the ticking.
🌐⚛️ Quantum Teleportation Through Live Internet Traffic
Northwestern University researchers successfully teleported a quantum state of light through 30 kilometres of fiber optic cable carrying actual internet traffic at 400 Gbps—a feat previously thought impossible. The breakthrough demonstrates that quantum communications can coexist with classical internet infrastructure without requiring specialised new networks.
⚛️🇩🇰 Magne: Denmark does quantum computing
Denmark is investing €80 million to build "Magne" what is intended to become the world's most powerful commercial quantum computer, featuring 50 logical qubits across 1,200 physical qubits in a Level 2 quantum system. The QuNorth project, backed by Atom Computing's hardware and Microsoft's Azure integration, aims to prevent the Nordic region from falling behind in the global quantum race, with construction starting late 2025 and completion by early 2027. (Magne is named after Thor's son in Norse mythology.)
Robotics
Uncle Bot A Unitree G1 humanoid robot dressed in peak "retired uncle" fashion—gym shorts, sneakers, and all—has become China's latest viral sensation after a jogging video racked up 80+ million views.
(Video: @0moonpiece)
(Video: @orikron)
Narrative: Just dressing a humanoid robot in clothes shows how close these devices are from transitioning from labs to real-world applications and public acceptance.
🎹🤖 “Adam” the Keytar-Playing Humanoid
PNDbotics' 60kg humanoid robot Adam took centre stage at a Chinese music festival, delivering a surprisingly fluid keytar performance alongside human musicians using 25 patented actuators and reinforcement learning algorithms.
Figure unveils fireproof robot battery with 94% more energy
Figure's new F03 battery packs 2.3 kWh into their humanoid robots' torsos, delivering 94% more energy density than their original design while cutting costs by 78% compared to the previous generation. The integrated battery doubles as a structural component and includes flame arrestor systems that successfully contained thermal runaway events during testing.
🔋🤖 Walker S2: Autonomous Battery Swapping
Meanwhile China's UBTech has unveiled the Walker S2, claiming it's the world's first humanoid robot capable of fully autonomous battery swapping in just three minutes, enabling true 24/7 industrial operations without human intervention. The robot features dual-battery redundancy and integrates with UBTech's "BrainNet" framework for coordinated multi-robot factory deployment.
🤖🍽️ Robot Metabolism: When Bots Start Eating Each Other
Columbia University researchers have developed modular "Truss Link" robots that can grow, self-repair, and morph by literally absorbing parts from other robots, complete with their own "laws of robot metabolism." The magnetic bar-shaped modules can cooperatively reshape themselves, help each other change form, and even jettison dead batteries while scavenging working components from their mechanical brethren.
🐕KLEIYN: Japanese robodog that climbs chimneys
University of Tokyo researchers have developed KLEIYN, a quadruped robot with a flexible waist joint that can squeeze through tight spaces and climb vertical walls at 170mm per second—50 times faster than comparable climbing bots. Rather than using grippers, the 13kg robodog relies on leg joint pressure against opposing walls to defy gravity, trained through a novel reinforcement learning platform called Contact-Guided Curriculum Learning.
🐘 Programmable foam skeleton
EPFL researchers have cracked the code on soft-rigid robotics with a programmable foam lattice that mimics animal musculoskeletal systems, enabling robots to seamlessly transition from flexible to rigid across their bodies. Their elephant robot demo showcases a twisting trunk and jointed limbs, all built from customisable foam cells that can generate over 75 million geometric configurations. The breakthrough lies in continuous spatial blending of material properties—essentially giving robots the ability to be squishy where they need to bend and sturdy where they need to bear weight, just like biological systems.
MT1Vac: Robot vacuum with 200% more suction power
Pudu Robotics has unleashed the MT1Vac, a commercial cleaning robot with 200% more suction power than standard units, targeting airports and metro stations with AI-driven surface recognition and 20-litre capacity.
🪱 Worm Blob Robotics
This one’s a completely novel robot form factor: Harvard researchers have created foot-long worm-like robots that physically entangle to move as a collective unit, mimicking the behaviour of California blackworms that form protective, writhing blobs:
The synthetic polymer bots use pressurised air chambers to curl and interlock, enabling coordinated movement across land and water surfaces that individual units couldn't achieve alone.
Narrative: as swarm robotics evolves beyond traditional wireless coordination, Harvard's entangled approach suggests that tomorrow's robot collectives might literally entangle each other to solve problems too complex for individual machines.
🦿 e-OPRA: MIT's bone-integrated bionic knee
MIT researchers have developed e-OPRA, a bionic knee system that integrates directly with bone and muscle rather than using traditional socket fittings, enabling remarkably natural movement for above-knee amputees. The system combines titanium bone implants with neural interfaces that read muscle signals, allowing users to control their prosthetic with thought and feel genuine ownership of the artificial limb.
Check out this video of the SOTA for bionic prosthetic limbs. Wow:
Energy
⚛️💸 Sizewell C's £38 billion reality check
The UK's Sizewell C nuclear project has seen costs balloon from £20 billion to nearly £38 billion. Commentary focuses on Europe's nuclear programmes stumble through decades-long construction timelines and fragmented approaches, while renewables have quietly scaled with standardised designs and predictable delivery schedules.
Narrative: Nuclear power is historically far more complex and costly than most other forms of electricity generation. This clearly hasn’t changed in this case … so it’s reasonable to expect similar cost blowouts for the new wave of nuclear build-out currently ramping up in the US as well. (Which will add to the $trillions already being invested to power AI.)
⚛️♻️Stellaria raises $27M for waste-destroying nuclear reactor
That said… French startup Stellaria just raised US$26.88 million to build the world's first commercial nuclear reactor designed to destroy more long-lived waste than it produces—a feat no reactor has ever achieved. Their fourth-generation molten salt design promises 20+ years of operation without refueling, passive safety systems, and enough electricity for a city of 400,000 people with energy density 70 million times greater than lithium-ion batteries. They're targeting first fission by 2029 and commercial deployment by 2035
🦅Solar farms confuse migrating birds with water illusion
Apparently solar panels are creating massive optical illusions that confuse migrating birds, who mistake the reflective surfaces for water bodies and get thrown off their ancient flight paths. Murdoch University researchers found that while solar farms help fight climate change, their rapid expansion—covering nearly 38,000 square kilometres globally—is creating unexpected biodiversity challenges through "polarized light pollution" and wildlife-trapping fences.
Solutions like nano-coatings to reduce glare and wildlife-friendly design practices (including fence openings for desert fauna) are already showing promise in places like Nevada.🌊China builds underwater data centre to cut cooling costs
China is constructing an underwater data center 6 miles off Shanghai, set to begin operations in September, as a novel solution to the massive cooling costs plaguing AI infrastructure amid global water scarcity. The project, moving from proposal to operation in just 2.5 years, contrasts sharply with Microsoft's abandoned Project Natick underwater data centre experiment off the coast of Orkney, Scotland from a decade ago.
🔋Cleaner lithium extraction digital twin could cut emissions 50%?
University of Connecticut researchers are developing an AI-powered "end-to-end digital twin" to extract lithium from geothermal brines, potentially cutting extraction emissions by 50% while creating a domestic US supply chain. Their mathematical models optimise locations for extraction, battery manufacturing, and EV production facilities, targeting mineral-rich sources.
Narrative: Much of the real-world value from AI is going to be found in advanced data analytics and digital twin simulations like this example, not LLMs. If you have accurate, proprietary data, then you have competitive differentiation to improve operations.
Transport tech
🚲⚡🔌 Tiler Compact
Dutch startup Tiler is relaunching their wireless ebike charging system as the more affordable "Compact" version (€250/~US$290), which tops up batteries through a chunky kickstand that connects to weatherproof 150W charging pads on the ground.
🚗🌏 Uber + Baidu + Lucid
Uber has inked a multi-year deal with Chinese search and AI giant Baidu to deploy thousands of Apollo Go robotaxis across Asia and Middle East markets before year's end. This follows Uber's pattern of partnering with autonomous vehicle leaders like Waymo and WeRide to maintain platform relevance as the industry shifts toward fully driverless rides.
The move represents Baidu's biggest international expansion yet, with Apollo Go having already delivered over 11 million rides across 15 Chinese cities.🚗🤖💰Uber is also investing US$300mn in struggling EV maker Lucid Motors while committing to purchase at least 20,000 of their Gravity SUVs equipped with Nuro's self-driving tech for its first robotaxi fleet. The deal provides crucial lifeline funding for Lucid, which racked up nearly US$6bn in losses over 2023-2024, sending shares soaring 36%.
🚀 Invictus: Phoenix rising from SABRE's ashes
Artist's impression of a space plane that could result from Europe's Invictus program. (Image credit: Frazer-Nash Consultancy) After Reaction Engines collapsed in 2024, a British-led consortium has resurrected the legendary SABRE hypersonic engine technology in the new Invictus programe, targeting a Mach 5+ spaceplane by 2031 with €7 million in ESA backing. The key breakthrough remains the pre-cooler system that chills incoming air from over 1,000°C to ambient temperature in less than 1/20th of a second using liquid helium and hydrogen fuel as coolant.
🛰️Colombia seizes first unmanned narco-sub equipped with Starlink
Colombian authorities have seized their first unmanned narco-submarine equipped with Starlink connectivity, capable of carrying 1.5 tonnes of cocaine across 1,287 km without human crew. The Gulf Clan-operated vessel represents a significant evolution in drug trafficking technology, using SpaceX's satellite internet for remote control and surveillance cameras for navigation.
(If the Ukrainian navy can do it, so can the narcos…)
Biotech
🧬👶🔬 Three-parent babies doing fine
Mitochondrial donation is an in vitro fertilisation technique involving the DNA of three people.Image credit: Phanie/Science Photo Library Britain's pioneering mitochondrial donation technique has produced eight healthy children, marking a breakthrough in preventing mothers from passing fatal genetic diseases to their offspring. The IVF process involves extracting a couple's genetic material from their fertilised egg and injecting it into a donor egg with healthy mitochondria, effectively creating babies with DNA from three people. While three children showed detectable levels of diseased mitochondria, all eight are developing normally—a promising sign for families facing devastating mitochondrial disorders.
🧬🥽Nanome
Startup Nanome is pushing spatial computing into drug discovery with MARA (their AI copilot) and immersive XR environments that let scientists collaboratively manipulate 3D molecular structures in virtual space. The platform targets computational chemists and budget-conscious managers, promising to shave 3-6 months off research timelines by making molecular interactions visible in ways 2D screens can't match. Harvard and Novartis are already publishing papers on the approach, suggesting this isn't just another fancy tech demo.
Narrative: It’s taken a long long time but it seems that finally productivity-charging 3D VR UIs for information manipulation may be upon us.
(I anticipated this way back in 2022 covering the VRNetzer research)
“It will almost certainly be several years until the virtual productivity UX matures to something which expands individual productivity, but when it does (combined with AI) there may well be a corresponding leap in what humans can achieve.”
Credit: Pirch, S., Müller, F., Iofinova, E. et al. The VRNetzer platform enables interactive network analysis in Virtual Reality. Nat Commun 12, 2432 (2021). 🔬Synthetic biology could enable Mars medicine production
Researchers are exploring how synthetic biology could keep future Mars and lunar colonists healthy by managing inflammation using insights from HIV research. The key lies in controlling inflammasomes—protein complexes that trigger immune responses but can cause chronic inflammation in space's harsh environment of microgravity and radiation. This could enable astronauts to produce custom medicines on-demand using bioreactors rather than hauling entire pharmacies to Mars.
Materials science
❄️ AI-Designed Thermal Meta-Emitters
Researchers at UT Austin used machine learning to create over 1,500 new thermal meta-emitter materials that can selectively control heat emission, achieving 5-20°C better cooling than commercial paints. The breakthrough could save the equivalent of 15,800 kWh annually per apartment building in hot climates like Bangkok—that's more than 10 times what a typical air conditioning unit consumes.
🌙⚗️ Lunar Alchemy: Turning moon dirt into life support
Researchers at Chinese University of Hong Kong, Shenzhen, have cracked the code on extracting water from lunar soil and converting astronaut-exhaled CO2 into oxygen and fuel using a single photothermal process powered by sunlight. The breakthrough could slash the eye-watering US$83,000-per-gallon cost of shipping water to the Moon, though significant technical hurdles remain around the lunar environment's extreme conditions and inconsistent soil composition.
⏳ Zeitgeist
🧬Cancer death rates in the US keep falling
More than five decades after Nixon's declaration of war on cancer, the world is actually winning through a "steady war of attrition". In the US, age-adjusted death rates have been falling since the 1990s thanks to thousands of incremental advances in screening and treatment. Childhood leukaemia survival rates jumped from 60% in 1975 to 90% today in the US, while prevention efforts like reduced smoking have quietly prevented millions of lung cancer deaths.
Compound technological progress and public health interventions can deliver generational wins — through initiatives like India's mass HPV vaccination campaign currently underway.
Ecological collapse
🧲 Soft Tipping: ecosystems fall apart like magnetic patterns
New research reveals that ecosystem collapse isn't always the dramatic cliff-edge we imagine—some systems "sag" rather than "snap," gradually reorganising under stress like complex magnetic materials in a lab. The study suggests we may already be crossing climate tipping points without realising it, as large-scale systems like forests and ocean currents quietly unravel in ways that are easy to miss.
🌲💧Clear-cutting forests increases flood frequency 18-fold
A University of British Columbia study tracking 70 years of data reveals that clear-cutting forests can make catastrophic floods 18 times more frequent, with effects persisting for over four decades.
Climate
⚡China's clean tech exports cut global emissions 1.5%
China's clean energy exports—solar panels, wind turbines, and EV batteries—cut global carbon emissions by 1% in 2024, rising to 1.5% when including Chinese-owned factories worldwide (equivalent to Australia's entire annual footprint).
Clean tech now represents over 10% of China's GDP for the first time, cementing its role as the world's manufacturing hub for climate solutions.⚖️🌊🏝️ICJ to deliver landmark climate ruling
The International Court of Justice is set to deliver a landmark advisory opinion defining countries' legal obligations to prevent climate change and the consequences for major polluters—essentially codifying what we all know but have been legally tiptoeing around for decades. Over 100 countries made submissions in what Vanuatu called a "David versus Goliath" battle, with small island states demanding accountability while major emitters like the US and India argued existing frameworks are perfectly adequate. The ruling won't be legally binding, but could become the legal scaffolding for a new generation of climate litigation—assuming anyone actually pays attention to international law anymore.
Extreme weather roundup
Typhoon Wipha batters Hong Kong with T10 warning
Typhoon Wipha hit Hong Kong on Sunday, triggering the city's highest T10 warning for seven hours and cancelling over 500 flights as 140km/h winds and torrential rain swept through the financial hub.
South Korea floods kill 17 amid record rainfall
The country's typically robust monsoon preparations proved insufficient against 800mm of rain in just five days.
🔥🇨🇦Canada wildfires burn area size of Croatia
Canada's wildfires have torched 5.5 million hectares this year—roughly the size of Croatia—making it one of the country's most destructive fire seasons behind only the apocalyptic 2023 record.
Pakistan's monsoon season has killed 180 people since late June, with 54 deaths in just 24 hours as torrential rains triggered flash floods and building collapses across Punjab province.
🎰🌊🏝️ Climate migration lottery: Tuvalu's exodus begins
Over 5,000 Tuvaluans—nearly half the nation's 11,000 residents—have applied for the world's first climate migration visa lottery, allowing 280 people annually to relocate to Australia as their Pacific island home faces inevitable submersion. The Australia-Tuvalu Falepili Union treaty represents humanity's first planned migration of an entire nation, with recipients gaining full Australian residency rights while retaining the option to return home.
🚀🌍⚠️Rocket launches threaten ozone layer recovery
ETH Zurich researchers warn that surging rocket launches—from 97 globally in 2019 to a projected 2,040 by 2030—could delay ozone layer recovery by years or decades through chlorine and soot emissions that persist 100 times longer than ground-based pollutants. The study projects a 0.3% global ozone decline with seasonal Antarctic reductions up to 4%, while satellite re-entry effects remain poorly understood but likely significant.
💧Nearly billion people face water scarcity by 2100
Northeastern researchers using advanced climate models (CMIP6) project that 850 million people will face water scarcity by 2100—more than triple previous estimates of 260 million. The improved models, with higher resolution and better physics, show 40% of the world's 30 major rivers experiencing decreased runoff. Lower carbon emissions could reduce affected population from 900 to 500 million in a "greener world" scenario.
Other headlines of interest from around the world this week. There’s so much going on…
Singapore battles 'serious' cyberattack on critical infrastructure.
China starts building world's largest hydropower dam China has commenced construction of the world's largest hydropower dam on Tibet's Yarlung Tsangpo river, a US$167bn project that will generate three times more energy than the Three Gorges Dam when completed. The massive infrastructure undertaking has triggered serious concerns from India and Bangladesh, whose downstream populations depend on the river system that flows into the Brahmaputra and other major waterways.
🌊⛏️ Deep Sea Mining regulation deadlocked The International Seabed Authority remains deadlocked on deep sea mining rules after two weeks of negotiations, while Trump's April decision to fast-track US permits using a 1980 law has Canadian firm TMC pivoting from the multilateral process to American regulatory arbitrage.
🏗️🪞 The Line: Reality check incoming Saudi Arabia's audacious 170km mirrored desert megacity "The Line" is facing a strategic review as spiralling costs and low oil prices force authorities to reassess the project's feasibility. Whoever would have thought it…
💰🌍 The Offshore Playbook: How Home Country Governance Shapes Elite Money-Hiding Strategies Dartmouth researchers used machine learning to decode the offshore financial strategies of nearly 3,000 billionaires and elites, revealing three distinct patterns based on their home country's governance quality. The study found that elites from authoritarian regimes use a "confetti strategy" (scattering assets), those from countries with weak civil rights employ "concealment strategies" (blacklisted jurisdictions), while others use hybrid approaches. Apparently 70-90% of assets from countries like Peru and Thailand end up in blacklisted jurisdictions compared to just 30% from China and Russia.
🔌🇺🇸 US bans Chinese technology in submarine cables The US Federal Communications Commission is set to ban Chinese technology from submarine cables connecting to America, targeting companies like Huawei in response to ongoing cyberattacks including the massive "Salt Typhoon" breach.
Houthi arms dealers exploit Meta and X platforms Arms dealers linked to Yemen's Houthi rebels are openly selling US military weapons through X Premium subscriptions and WhatsApp Business accounts, with some weapons bearing "Property of U.S. Govt" markings. Get that Blue Tick.
📱🇨🇳🕵️ Massistant: China's phone forensics dragnet Chinese authorities are deploying a new forensic software tool called "Massistant" that apparently hoovers up everything from Signal messages to location histories from a seized phone, requiring only physical access to unlocked devices.
Developed by surveillance tech giant Xiamen Meiya Pico (which holds 40% of China's digital forensics market), the tool is reportedly in widespread use across Chinese police agencies since legal powers expanded in 2024 to allow warrantless device searches.
It feels like it’s best to wipe one’s phone every time before you enter the US and China.France's parliament urges Europe to partner with China France's National Assembly has released a 153-page report calling for Europe to abandon its US-led confrontational stance toward China and forge an independent, cooperative relationship with Beijing instead. The document includes 50 recommendations spanning trade to space cooperation and explicitly criticises what it terms American "vassalisation" of European policy. The beginning of Europe's strategic autonomy moment, or just French diplomatic theatre?
💭Meme stream
What a monster issue this week…. so much going on!
🌈☄️Cosmic rainbow comet
Astronomers have captured stunning technicolour timelapse images of 3I/ATLAS, only the third confirmed interstellar object ever detected, as it hurtles through our solar system at 210,000 km/h. The comet, potentially up to 24km across and possibly 3 billion years older than our solar system, will reach its closest approach to the sun in late October before disappearing forever.
🪑Google Veo 3 does Ikea
Nowadays, sometimes a single video prompt will do better than a $100,000 VFX production budget:
(Video: @minchoi. Full prompt here.)
🚀Space jellyfish
Incredible footage of the “space jellyfish” phenomena from China’s launch of Long March 7 this week.
(Video via @PettitFrontier)
🏠Architecture gems
A couple of pieces of architecture which grabbed my attention via New Atlas:
🤔Explain bitcoin to me
An oldie but a goldie…

🌌The observable universe in one image
Pablo Budassi’s mesmerising Circular Map of the Universe is now in its 13th year, with updates added for 2025:
“The solar system is located in the center. Towards the edges, the scale is progressively reduced to show in detail the most distant and biggest structures of the observable universe sphere.“
Take the full tour:
🙏🙏🙏 Thanks for making it to the end! As always big appreciation to everyone who takes the time to get in touch with links and feedback.
Ha' det godt
Ben
Classic Bill Hicks on this topic, way back in 1993:
”There is a new party being born: The People Who Hate People Party. ‘People who hate people: come together!’
‘No!’
We’re kinda having trouble getting off the boards, but you know.
‘Are you gonna be there?’
‘Yeah.’
‘Then I ain’t fucking coming.’
‘You’re our strongest member.’
‘Fuck you!’
‘That’s what I’m talking about, you asshole.’
‘Fuck off!’
Damn, we almost had our meeting going. It’s so hard to get my people together.”