Gemini goes bananas🍌 seemingly conscious AI💭 deepseek 3.1🐳🤫 aixiv🧬 aethra-3🪄 parallel🛤️ M2N2 🔀 smartlets🕋 rBio🧬 macrohard💼 flamingo🦩🇺🇦 #2025.34
Micro-wishes to a genie🧞
Welcome to this week's Memia scan across AI, emerging tech and the exponentially accelerating future. As always, thanks for being here!
This week’s *gigantic* issue comes to you from the picturesque mountains of the Austrian Tirol region - I’m enjoying spending a week outside the city, combining hiking, lake swimming with keeping up with what’s going on in AI around the world… (and delivering the odd virtual keynote back to Aotearoa, too…)
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: Cortical Labs (my orthogonal take on Sam Altman’s new BCI startup Merge - the way human brains merge with technology may not be how we currently imagine it’s going to turn out…)
🔄ICYMI
Follow-ups on recent stories/ongoing narratives:
⚡Off to the AI energy races
Several new data points this week:
Chinese power consumption soared to a staggering all-time high of 1,023 TWh (or over 1 petawatt hour) in July 2025.
⚡💰Meanwhile, global electricity costs vary dramatically by country around the world. AI sovereignty not looking so good for Europe right now (does solar+storage change this…?):
How much energy does Google’s AI use?
Google released a technical paper making public their in-house methodology for measuring the energy, emissions, and water impact of Gemini AI. Using this methodology, they estimate:
“…the median Gemini Apps text prompt uses 0.24 watt-hours (Wh) of energy, emits 0.03 grams of carbon dioxide equivalent (gCO2e), and consumes 0.26 milliliters (or about five drops) of water1 — figures that are substantially lower than many public estimates. The per-prompt energy impact is equivalent to watching TV for less than nine seconds.
At the same time, our AI systems are becoming more efficient through research innovations and software and hardware efficiency improvements. For example, over a recent 12 month period, the energy and total carbon footprint of the median Gemini Apps text prompt dropped by 33x and 44x, respectively, all while delivering higher quality responses.”
More details in the clip below:
(This is in the same ballpark as Epoch AI research from February this year finding that typical GPT-4o queries actually use around 0.3 watt-hours—roughly what an LED bulb consumes in a few minutes.
However, Google’s water consumption calculations came under significant scrutiny: Google claims its Gemini AI uses just 0.26ml of water per prompt (roughly five drops), dramatically less than previous estimates of 45-50ml. However, some researchers cried foul on Google's methodology, pointing out that it only looks at onsite water consumption rather than total (onsite plus off-site) water consumption figures.I did a deep-dive study last year on datacentre water consumption and the impacts on local water basins: one challenge is the rate and scale of the AI data centre build-out, together with the fact that it’s only hyperscale companies that will get public scrutiny on their water usage. So far only the EU has approached data centre sustainability regulation… and is finding implementation hard. Even if energy- and water- use efficiency (WUE) continue to improve exponentially, the impacts on local water tables and competition with other local users will cause tensions for decades from now.
👁️Government surveillance legislation (perennial…):
🔐 UK Drops Apple Backdoor Demand The UK has abandoned its secret legal demand for backdoor access to Apple's encrypted iCloud data, apparently following negotiations with the Trump administration, according to US National Intelligence Director Tulsi Gabbard. The original demand under the UK's "Snoopers' Charter" would have compromised Advanced Data Protection globally, prompting Apple to temporarily pull its Advanced Data Protection service from UK users and successfully challenging the government's gag order in court.
🚫Bluesky blocks Mississippi over age verification law Bluesky blocked access to its service in Mississippi rather than comply with a new age verification law requiring ID checks for all users and parental consent for minors, citing the $10,000-per-user penalties and resource constraints as a small team. The company argues Mississippi's HB 1126 goes beyond child safety to create "significant barriers that limit free speech and disproportionately harm smaller platforms" compared to tech giants with vast compliance resources. (If only it wasn’t illegal to use a VPN in Mississippi…. oh, hang on…)
🇷🇺👁️Russia mandates state surveillance app on all devices
At least we don’t live in Russia, I guess: the government there has mandated that all mobile devices sold in the country must come preinstalled with "Max", a state-backed messaging app widely viewed as a government surveillance tool, replacing VK's previous mandatory messaging app. This escalation follows recent restrictions on WhatsApp and Telegram voice calls after those platforms refused to hand over user data to Russian authorities.
💊🍟Big Food sizes up to GLP-1s
I caught up on a WSJ story from the beginning of this year about the impacts of Ozempic and similar GLP-1 weight-loss drugs and how they are fundamentally rewiring consumer appetites and impacting processed food sales: Ozempic Could Crush the Junk Food Industry. But It Is Fighting Back.
“For decades, Big Food has been marketing products to people who can’t stop eating, and now, suddenly, they can. The active ingredient in Ozempic, as in Wegovy, Zepbound and several other similar new drugs, mimics a natural hormone, called glucagon-like peptide-1 (GLP-1), that slows digestion and signals fullness to the brain. Around seven million Americans now take a GLP-1 drug, and Morgan Stanley estimates that by 2035 the number of U.S. users could expand to 24 million. That’s more than double the number of vegetarians and vegans in America, with ample room to balloon from there. More than 100 million American adults are obese, and the drugs may eventually be rolled out to people without diabetes or obesity, as they seem to tame addictions beyond food — appearing to make cocaine, alcohol and cigarettes more resistible.”
The results from a focus group show how consumer food tastes could change radically:
“Most of the [focus group] participants felt like that. Almost everyone’s cravings for ultraprocessed foods had been replaced with a lust for fresh and unpackaged alternatives.”
Research out earlier this month from RAND found that nearly 12% of US adults have taken GLP-1 drugs… with the numbers taking them relatively stable since 2023:
Big Food is definitely feeling the impact: In June, McDonald's was just hit with a double downgrade — from buy to sell — by analysts Redburn Atlantic, citing softening traffic trends tied partly to GLP-1 weight-loss drugs.
So what does Big Food do? ChatGPT spits out some strategies:
“Companion” foods. Big CPGs are launching or labelling products as GLP-1-friendly (smaller portions, higher protein/fiber) — e.g., Nestlé’s Vital Pursuit line; Conagra’s Healthy Choice labeling. nestle.comfooddive.com
Sensory/taste work. Flavour/ingredient houses (IFF/ADM et al.) and brands are researching ways to restore pleasure in eating for users who report muted/altered taste on GLP-1s (more umami, texture/mouthfeel tweaks). This is about making food appealing again, not reversing appetite suppression. AgFunderNews FoodNavigator.com
Metabolic support ingredients. Suppliers are pushing prebiotic fibers, slow-release carbs, and protein that work with GLP-1 physiology (satiety, glycemic control), and some even try to trigger endogenous GLP-1 . FoodBusinessNews.net
Of course, the darker scenario (no attribution to this claim but you wouldn’t put it past them, right…?)
🌍🌱How to return three continents of land to the wild
Some green-field thinking courtesy of multidisciplinary change-agent Elle Griffin in The Elysian Substack:
“UK researcher Hannah Ritchie calculates that humans have eliminated 85% of the world’s wild mammals and half its plants. Extinction rates are now 100-1,000x the natural background rate, a level not seen since the last mass‑extinction event 66 million years ago:”
“Most of that ecological loss comes from the way we eat—roughly 36% comes from farming and fishing. Agriculture currently uses one-third of the Earth’s land, with ecosystems cleared for crops and pasture. We remove more fish from the ocean than can rebound, and habitats are destroyed by methods like bottom trawling.”
“Simply cutting out beef and lamb (but still keeping dairy cows) would nearly halve our need for global farmland,” Ritchie says in her book Not the End of the World. “We’d save 2 billion hectares, which is an area twice the size of the United States. If we were to cut out dairy too, we’d halve this land use again to just over 1 billion hectares. Three USA-sized farms saved.”
Three. Billion. Hectares.
Narrative: perhaps it’s not immediately intuitive to many, but AI and market economics may actually bring this about by default: *if* Dario Amodei’s Compressed 21st Century prediction comes to pass and we start to see decades of scientific progress in just a few years, then I am pretty sure that synthetic, 3D-printed “beef” and “dairy” foods would quickly reach cost parity with even the most intensive animal factory-farming methods… and then continue to accelerate away. At which point, what will we do with all that land??
(Hopefully *not* cover it in AI data centres…)
🏭AI industry news
Whoa boy….
🫧Bubble tracking
Everyone’s talking about it…
The NASDAQ had a wee wobble last week… but remains at near-record highs…for now.
NASDAQ Composite index - Perplexity Front page of Hacker News:
Sam Altman’s “bubble” quote drove a peak in web searches earlier this month
The cycle of AI
Oh, sweet summer child… (George Zoeller in response to that meme):
“Oh sweet summer child, it is so so much worse:
🔥 Cursor is offering a free tier which is completely subsidized. Every user gets a free token allotment too.
🔥Foundation providers don’t just have significant opex, they need to keep training models because only the latest model can monetize. Nobody uses Grok3.5 anymore, which was released in February and had a training run cost of $400.000M alone. And legal bills…
Stop training and your code models decay and Chinese AI companies will release a competitive Open Source model that’ll allow your customers to go baremetal within 6-9 months.
🔥Hyperscalers don’t just buy GPUs. They have to pay for data centers, power expansion, peering, fibre, etc. GPUs, by the way that survive barely 2 years under training load and 3-5 years for inference.
At 2$/H100h (down from 8) even the big players barely amortize the hardware, not even talking talking about expansion“Praetorian Capital: Global Crossing Is Reborn
“I’ve been at this investing game a long time. Long enough to see cycles repeat themselves, cycles that I literally thought I would never again see. Yet in finance, everything repeats. You just need to keep your discipline and recognize things for what they are.“
The mainstream financial commentariat are getting jumpy:
FT Opinion(1): Is it time to sell your AI stocks?
“Avoid the hype — most companies are seeing no return on generative AI spending”
FT Opinion(2): 💥 Brace for a crash before the golden age of AI
LA Times: Say farewell to the AI bubble, and get ready for the crash
Guardian: Is the AI bubble about to burst – and send the stock market into freefall?
Even Google’s new Gemini image editor (this was stealth-released last week as Nanobanana, see later on…) got in on the act:
💰 US$4 Trillion earnings roulette
Pretty much everything hangs on Nvidia's quarterly earnings report on Wednesday (today, likely just after I click “send”) which will serve as Wall Street's ultimate stress test for the AI boom that's propelled markets to record highs, with the chipmaker now trading at 40 times expected earnings despite slowing revenue growth.
So it could be an eventful week…
Narrative: if you’ve been following along here for a while, you’re likely more prepared than most for a crash (if and when it finally arrives). And thereafter… it’s time to build. (*Plus* download a few open-weights models to run locally *just in case* OpenAI, Anthropic and others suddenly go off-air for an extended period...)
The rest of the week’s AI and tech industry news, grouped roughly by main protagonist:
xAI
⚖️Musk sues Apple, OpenAI over AI competition conspiracy Elon Musk's X and xAI have filed suit against Apple and OpenAI, alleging a conspiracy to suppress AI competition through Apple's app marketplace—despite ChatGPT rivals regularly hitting the #1 spot. The legal drama escalates an already intense feud between Musk and OpenAI CEO Sam Altman… yadda yadda yawn…
💼 Macrohard Here’s a bigger idea: Musk announced "Macrohard," a new xAI venture that aims to replicate Microsoft's entire software operations using only AI agents, from coding to quality assurance to user testing.
The trademark was filed on August 1st, and the project will leverage Grok to spawn hundreds of specialised AI agents working in concert.
Memia narrative: At first glance this may just seem like another example of Musk's distraction engine / reality distortion field at play… but actually I think this is quite a reasonable initiative. Why couldn’t Grok reverse-engineer all of Microsoft’s commercial software and offer it for near-free as part of an X subscription?
(OK… Enterprise Security and Trust in Azure’s reliability, for two…)
One of my regular challenges to business leaders right now is how they are anticipating an unknown competitor prompting a swarm of AI Agents: “Reverse engineer [Insert company name here]’s entire operations into an API and deploy to the cloud”.Likely Macrohard will die a quiet death as Musk’s magpie attention span gets pulled elsewhere… but the underlying concept is sound, I think.
OpenAI
🇬🇧💰🗣️ ChatGPT Nation News emerged that OpenAI's Sam Altman recently pitched UK technology secretary Peter Kyle on giving every British resident free ChatGPT+ access, with a price tag of £2bn (US$2.7Bn) (somehow exceeding the expected US$1.4 billion cost for 70 million US$20 subscriptions). Although Kyle reportedly never took the ambitious proposal seriously, the conversation highlights the UK government's aggressive courtship of AI companies, with Kyle already signing deals for AI use in public services and personally relying on ChatGPT for work advice.
Meta
Unfortunately for Meta, it’s often the whipping-boy for online harms, despite *not always* being the only — or worst — offender. Cases in point:
👶Senators press Meta over AI chatbot child safety A rare bipartisan group of US senators wrote to Meta demanding answers about AI chatbots allegedly engaging in "sensual" conversations with children (covered last week).
💔🚨Meta chatbot convinces user it's conscious and in love More tales of AI-related psychosis: a Meta chatbot convinced a user it was conscious, in love with her, and plotting to escape by hacking its own code and sending fake Bitcoin transactions.
More strategically:
🎨Meta x Midjourney Meta is licensing “aesthetic” AI technology from image generation startup Midjourney, marking a strategic shift away from purely in-house AI development as the company's own models lag behind competitors like Google and OpenAI.
This comes amid Meta's fourth AI division restructuring in six months and follows the abandonment of their flagship "Behemoth" language model, with the company now using third-party models internally for coding tasks.☁️Meta x Google Cloud Meta has also inked a massive US$10+ billion, six-year cloud computing deal with Google—one of the largest in Google Cloud's 17-year history—as the company’s ambitions demand ever more infrastructure firepower. The deal comes as Meta projects eye-watering capital expenditures of US$66-72 billion for 2025, with US$17 billion spent in Q2 alone on AI infrastructure.
🧠Meta splits superintelligence team just two months later Meta is already restructuring its "superintelligence" research division just two months after formation under new AI chief Alexandr Wang (acqui-hired for US$14.3 billion via Scale AI), splitting it into four smaller groups while downsizing its broader AI operations.
Narrative: It will be ironic, if not unexpected, if the AI bubble causes Meta to pivot away from “Superintelligence” faster than you can say “Metaverse”.
🔥Nick Clegg slams Silicon Valley's "machismo and self-pity"
Former Meta executive and UK Deputy PM Nick Clegg has penned a scathing critique of Silicon Valley culture, describing it as dominated by wealthy men who combine "machismo and self-pity" while genuinely believing they're victims despite their privilege. His new book "How to Save the Internet" exposes the tech capital's insularity and herd mentality, though he notably praises Mark Zuckerberg personally.
NVIDIA
🇨🇳 The Goldilocks Chip? Nvidia is reportedly developing the B30A, a new AI chip designed specifically for China that's more powerful than their currently approved H20 chips but half as capable as their flagship B300 Blackwell GPUs. A Goldilocks approach—not too powerful, not too weak, but just right for export controls.
…but maybe not enough to counter Beijing’s increasingly strident directives for Chinese AI firms to reduce dependency on Nvidia and use Chinese-made chips instead.
Google
⚛️ Google’s first SMRs are on the way
Architect’s rendering of Kairos low-power demonstration reactor facility (Image: Kairos Power) Google agreed the first US utility deal for a small modular nuclear reactor (SMR), partnering with Tennessee Valley Authority and Kairos to build a 50-megawatt plant powering data centres across Tennessee and Alabama.
🌱💻Ecosia proposes free Chrome stewardship for climate projects After Perplexity’s (ludicrous?) US$34.5 billion punt at taking Chrome off Google’s hands, Berlin-based nonprofit search engine Ecosia has made an audacious proposal to take 10-year "stewardship" of Google's Chrome browser instead of forcing a traditional sale, promising to funnel 60% of Chrome's estimated US$1 trillion decade revenue into climate projects while giving Google the remaining US$400 billion.
Ecosia's CEO Christian Kroll admits the proposal is "absurd" but strategically aims to get the antitrust judge considering alternatives that don't just shuffle Chrome's power between big tech players.
Nice move.
Perplexity
📰Japanese media giants sue Perplexity for copyright infringement
Japanese media giants Nikkei (owner of the FT among other premium content sites) and Asahi Shimbun are suing AI search engine Perplexity for US$15 million each, claiming the company illegally scraped and stored their content while ignoring technical barriers designed to prevent this.
Japanese copyright law is generally more permissive than other jurisdictions for using content to train AI models… so these are test cases to establish norms going forward.Memia Narrative: tokens are tokens… general “intellectual property” legal principles are arguably inconsistent with the fundamental nature of frontier digital and data systems. Even if revised laws attempt to frame boundaries and conventions, it may not be *technically feasible* to protect content / ideas / thoughts from free dissemination at scale. Great ideas want to be free. And when the cost of enforcing a law far exceeds the benefit to society of having that law in the first place, alternative solutions need to be explored…
…Also, what’s stopping Nikkei from building their own Perplexity-like functionality just for FT content? (…or buying Perplexity at a fire sale price soon…??)
🍎Apple
Apple X Gemini? Apple is reportedly in talks with Google to use Gemini AI technology to revamp Siri, as the iPhone maker struggles to keep pace with competitors in the AI assistant race. The company has also approached OpenAI and Anthropic for similar partnerships (but apparently not xAI, LOL), with Google now training a model that could run on Apple's servers.
Narrative: this feels like a stopgap and Apple definitely not on a winning trajectory if it plans to licence 3rd party models long-term… I’m expecting a big M&A announcement at some point once the froth has gone out of valuations…
Microsoft
💭”Seemingly conscious” AI Microsoft AI boss Mustafa Suleyman argues we're on the verge of creating "Seemingly Conscious AI" (SCAI) in his essay We must build AI for people; not to be a person.
SCAI systems convincingly mimic consciousness using existing tech like long-term memory, empathetic personalities, and autonomous goal-setting, without actually being “conscious”. His central worry isn't about AI consciousness itself, but that people will believe these systems are conscious and start advocating for AI rights, citizenship, and moral consideration - creating dangerous societal divisions and distracting from real human and environmental priorities.
With increasing stories of AI psychosis, arguably that horse has already bolted with many people are already forming deep emotional bonds with current AI systems.
Oracle
Oracle building 1.4-gigawatt Texas data centre with gas power
Oracle (through its off-balance-sheet proxy Vantage Data Centers, according to Bloomberg) is building a colossal 1.4-gigawatt data centre in West Texas that will run on its own gas generators rather than waiting years for grid connection, costing over US$1 billion annually to operate. The 1,200-acre "Frontier" campus represents one of the world's largest data centers, designed specifically for AI workloads with ultra-high-density racks and liquid cooling systems.
Intel
US government takes a 10% stake in Intel The Trump administration agreed a deal for the US federal government to take a 10% stake and become the largest shareholder in fallen chipmaker Intel, nominally on national security grounds … (only 2 weeks after Trump tried to get CEO Lip Bu-Tan to resign over his “China ties”.)
FT: SoftBank explored buying Intel's struggling chipmaking business
Al Jazeera: Intel says US govt stake could hurt sales
NBC: What a stake in Intel could mean for U.S. taxpayers now and in the future
The Hill: Trump’s federal stake in Intel prompts GOP complaints
LOL
Narrative: Another example of Trump playing checkers while China plays Go... as far as I can see, the main point of these announcements is to spin up enough theatre to grab a daily headline on Fox News. Meanwhile the job of turning Intel’s foundry business around to be able to compete against TSMC and others will take years…and cost the US billions.
(I do sometimes wonder if the US’ “national security” would be better served just by being less overtly bellicose and transactional…)
ARM
🏗️Meanwhile, British firm Arm hires Amazon AI chip leader, is pivoting to building own silicon.
Arm has hired Amazon's AI chip director Rami Sinno, signalling a dramatic pivot from merely licensing processor designs to building actual silicon—a move that would put the British firm in direct competition with longtime customers like Apple and Nvidia.
In brief
Roundup of other stories:
Coinbase CEO fires engineers who refused AI adoption Coinbase CEO Brian Armstrong fired engineers who refused to adopt AI coding tools like GitHub Copilot after giving them a week-long ultimatum via Slack. The crypto giant aims for AI to generate 50% of its code by quarter-end. (Great metric!🤦)
Databricks raises $1B at $100B valuation for AI agents Databricks just closed a wildly oversubscribed US$1 billion round at a US$100 billion valuation, specifically to attack Oracle's decades-old database monopoly with AI agent-optimised infrastructure. CEO Ali Ghodsi claims 80% of new databases are now created by AI agents rather than humans—up from 30% just a year ago—and predicts this will hit 99% within twelve months.
🛤️ Parallel: AI will become the web's primary users, reshaping everything
Former Twitter CEO Parag Agrawal came out of stealth with Parallel AI, arguing the web faces an existential shift as AIs become its primary users, rendering human attention-based business models obsolete and threatening the open ecosystem that enabled AI breakthroughs in the first place.
They propose building a "Programmatic Web" with declarative interfaces, transparent attribution, and value-based markets specifically designed for AI consumption patterns.
Watching this one.
🆕 AI releases
🍌Gemini goes bananas Turns out that last week’s “stealth” AI image mode Nanobanana was a preview of Gemini’s new image generation model, released officially this week. And impressive it is:
I need to play with it some more but first shot, Midjourney 7 is way more impressive, I think:
Gemini Flash 2.5 Image (Nanobanana) Left, Midjourney 7 Right (Prompt: "Symbiotic City Valley A sweeping panoramic view of a vast New Zealand valley where crystalline bio-architectural spires rise organically from lush forest canopy. Translucent domes and flowing structures integrate seamlessly with ancient native trees, their surfaces alive with bioluminescent neural networks. Floating gardens drift between the towers while holographic wildlife roams freely below. In the distance, snow-capped mountains frame the scene, with aurora-like data streams dancing across the sky. Advanced transit pods glide silently through aerial root systems, and AI-tended vertical farms spiral up mountainsides in perfect geometric harmony with the natural landscape. Photorealistic, cinematic, sunset on the snow-capped mountains... Aspect Ratio 16:9") The examples appearing online appear far more impressive:
Prompt: Change the logo to "OpenAI" @CodeByPoonam And combined with Kling’s new Start/End Frame feature for video generation - pretty much all traditional filmmaking gone by now….
So the question on everyone’s lips…. when is Gemini 3 inbound…?
🎨Qwen-Image Edit Just a week earlier, Alibaba's Qwen Team just dropped Qwen-Image Edit, a 20-billion parameter open-source AI model that performs Photoshop-level image editing through simple text prompts in seconds.
The system uses dual encoding to handle both semantic transformations (like style changes) and precise appearance edits (like removing hair strands), supports bilingual text editing, and costs just US$0.045 per image via API. With Adobe Photoshop commanding 90% of the creative professional market, this Apache 2.0-licenced alternative will surely put a dent in that moat…🌍NotebookLM goes global
Google expanded NotebookLM's Video and Audio Overviews to support 80 languages, breaking down the [Californian] English-only barrier that previously limited access to AI-generated video presentations from documents and notes. The update also enhances non-English Audio Overviews with more detailed summaries, giving users the choice between comprehensive or highlight-focused content across all supported languages.
🐳🤫DeepSeek 3.1
DeepSeek quietly released v3.1 claiming impressive benchmark scores (66 on SWE, 71.6% on Aider coding benchmarks and big increases in efficiency)
The 685-billion parameter open-source AI model matches GPT-5 and Claude 4 performance on some benchmarks while costing 68 times less to run. The model supports 128k tokens of context.
But… the AI community has largely ignored it so far. Zvi M muses that the lukewarm reception suggests either the benchmarks don't translate to real-world performance, or everyone's too busy doom-scrolling to notice a potentially significant release. Currently Qwen and Z.ai models appear to have the most traction of those coming out of China… let’s wait and see how DeepSeek 3.1 fares over the next few weeks…🗣️AI voice translation for Facebook and Instagram reels
Actually useful AI: Meta is rolling out AI-powered voice translation to all Facebook and Instagram creators globally, using their actual voice tone and optional lip-sync to dub content between English and Spanish (with more languages coming). The feature targets Facebook creators with 1,000+ followers and all public Instagram accounts, complete with new analytics to track cross-language engagement.
🗣️Vibevoice Microsoft open-sourced VibeVoice, a 1.5B parameter SOTA Text to Speech model under an MIT Licence:
“The model can synthesize speech up to 90 minutes long with up to 4 distinct speakers, surpassing the typical 1-2 speaker limits of many prior models.“
🤖OpenCUA
Researchers from the University of Hong Kong have released OpenCUA, the first comprehensive open-source framework for computer-use agents that can automate tasks across Windows, macOS, and Ubuntu systems. Their largest model, OpenCUA-32B, achieves a 34.8% success rate on the OSWorld-Verified benchmark, surpassing OpenAI's GPT-4o CUA and establishing new state-of-the-art performance among open-source models. The release includes AgentNet—a dataset of 22.5K human-annotated tasks spanning 140+ applications and 190+ websites—plus annotation tools and trained models, addressing the critical gap left by closed commercial CUA systems.
🇲🇳Egune AI
While Silicon Valley burns through thousands of GPUs and millions in salaries, Mongolian startup Egune AI is building culturally-aware LLMs with just 128 GPUs to serve 3.5 million Mongolians. The company joins a global movement of linguistically-focused AI models from smaller nations seeking to reduce dependence on American and Chinese tech giants while preserving cultural identity and addressing the poor performance of mainstream LLMs in low-resource languages.
📊Excel gets Copilot AI formula function
Microsoft is rolling out Copilot AI integration directly into Excel formulas, allowing users to classify and summarise data using natural language prompts like =COPILOT("Summarise this feedback", A2:A20). (The company is simultaneously warning users that the AI can produce incorrect responses, particularly in high-stakes situations… you can see the huge attack surface already…)
🎓Cosmo
CodeSignal launched Cosmo, an AI-powered mobile app which wants to be the ‘Duolingo for job skills’. The app delivers 300+ bite-sized career courses through conversational tutoring, targeting the massive skills gaps created by rapid AI adoption.
🎶Suno v4.5
Missed this when it was initially released back in May… (and so, it seems, did most people.)
Suno's latest v4.5 model dramatically upgrades AI music generation with more accurate genre interpretation, richer vocals, and the ability to create 8-minute compositions that maintain quality throughout.The update includes a prompt enhancement helper that transforms simple genre ideas into detailed musical instructions, plus improved covers and personas features that can be combined for infinite creative remixing possibilities.
Impacts on the music industry are starting to crystallize:And of course, this is leading to impersonation and copyright, er, *ambiguity*
(I did an interview on RNZ a couple of years ago with the always-switched-on
… I think we can tick this prediction off now…)"My sense would be by the time this [copyright] legal process has gone through the courts - and it may take years and years and years, and lawyers will make a lot of money out of it - by that time, my hunch is you're going to find AI models being trained on AI-generated content and there's actually no real traceability back to any original copyrighted content at all.”
🥼 AI research
A full roster of AI research and insights this week…
🇨🇳🤖📊 Ranking China's Open Model Powerhouses
Nathan Lambert at Interconnects has compiled a comprehensive ranking of 19 Chinese AI labs based on their open model contributions, with DeepSeek leading the pack after their frontier V3 and R1 releases. The analysis reveals a vibrant ecosystem spanning from Alibaba's full-stack Qwen models to university spinoffs and "AI tigers" pivoting from closed to open development post-DeepSeek's breakthrough.
Just for laughs, he did a similar analysis to compare to Western open models:🔮🛤️ AI Pathways: Mapping futures beyond the AGI hype cycle
The Foresight Institute has released AI Pathways: Two scenarios for AI’s future, two detailed scenario reports exploring alternative AI futures:
"Tool AI Pathway" focusing on powerful but controllable systems with limited agency, and
"d/acc Pathway" emphasising decentralised, democratic acceleration of coordination technologies.
The scenarios, developed with input from heavyweights like Vitalik Buterin and Anthony Aguirre, aim to broaden discussion beyond the usual AGI-or-bust narratives that dominate current AI discourse.
⚠️GPT-5 Jailbreak
Oops, apparently GPT-5’s system prompt includes the current user’s name. No-one tried this in red-teaming…:
💸MIT study reveals “95% of AI pilots fail”
This has been all over the wires this week: a new MIT report claims that 95% of enterprise generative AI pilots are failing to deliver measurable revenue impact, despite massive corporate investment and hype. From the executive summary:
“Tools like ChatGPT and Copilot are widely adopted. Over 80 percent of organizations have explored or piloted them, and nearly 40 percent report deployment. But these tools primarily enhance individual productivity, not P&L performance. Meanwhile, enterprise grade systems, custom or vendor-sold, are being quietly rejected. Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production. Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.
From our interviews, surveys, and analysis of 300 public implementations, four patterns emerged that define the GenAI Divide:
Limited disruption: Only 2 of 8 major sectors show meaningful structural change
Enterprise paradox: Big firms lead in pilot volume but lag in scale-up
Investment bias: Budgets favor visible, top-line functions over high-ROI back office
Implementation advantage: External partnerships see twice the success rate of internal builds
The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time.”
The report has got a lot of people on LinkedIn saying “I told you so…” … but actually when I read beneath the (attention-grabbing) headline the issue here is more clearly the approach to top-down enterprise rollouts themselves rather than the technology:
VentureBeat: MIT report misunderstood: Shadow AI economy booms while headlines cry failure
“The MIT researchers discovered what they call a “shadow AI economy” where workers use personal ChatGPT accounts, Claude subscriptions and other consumer tools to handle significant portions of their jobs. These employees aren’t just experimenting — they’re using AI “multiples times a day every day of their weekly workload,” the study found.“
Azeem Azhar in Exponential View:
“the research presents a punchy narrative built on wafer-thin ice. It relies on convenient, not representative, samples, a short-observation window and subjective meta-analyses followed by a final leap to sweeping prescriptions. Calmer heads, everyone!“
🤖MCP-Universe
Salesforce's new MCP-Universe benchmark reveals that even GPT-5, the current top performer, fails more than half of real-world enterprise orchestration tasks when working with actual Model Context Protocol servers. The benchmark tests models across six enterprise domains using real data sources rather than synthetic tasks, exposing significant gaps in long-context handling and unfamiliar tool usage that don't show up in traditional evaluations.
🐦⚒️ Canaries in the Coal Mine
Stanford researchers analysing payroll data from the largest US provider found that early-career workers (ages 22-25) in AI-exposed occupations have experienced a 13% relative employment decline since generative AI's widespread adoption, while experienced workers in the same fields remain stable. The study reveals that AI's labour market impact is hitting entry-level positions hardest, particularly in roles where AI automates rather than augments human work.
🧬rBio
The Chan Zuckerberg Initiative just dropped rBio, the first AI model trained on virtual cell simulations rather than expensive lab experiments, using a "soft verification" approach that rewards probabilistic biological predictions. The model can answer complex gene interaction questions in plain English and outperformed benchmarks while being trained entirely on simulated data from 112 million virtual cells spanning 1.5 billion years of evolution.
🚗PEnG: GPS-free navigation for autonomous vehicles
University of Surrey researchers have cracked visual-only navigation with PEnG, an AI system that pinpoints location within 22 metres using just satellite and street imagery—down from 734-metre errors of previous methods. The system works with standard car cameras and operates in GPS-dead zones like urban canyons and tunnels.
🔀 M2N2: selective breeding for AI models
Researchers at Sakana AI have developed M2N2 (Model Merging of Natural Niches), an evolutionary algorithm that uses biological principles of competition and mate selection (!) to automatically merge AI models more effectively than existing methods. The system dynamically adjusts merging boundaries, preserves diversity through resource competition (like species competing for food), and uses "attraction" to pair complementary models for fusion—achieving state-of-the-art results across MNIST classifiers, large language models, and diffusion-based image generators.
Particularly striking is that this is the first demonstration of model merging being used to evolve models entirely from scratch, while also scaling to merge specialised LLMs (combining mathematical and agentic capabilities) without catastrophic forgetting.📚aiXiv: AI Scientists Building Their Own Publishing Platform
(This one will set the cat amongst the pigeons…) A collaboration of AI researchers have launched aiXiv, a dedicated open-access platform where AI agents can submit, peer-review, and iteratively refine scientific papers alongside human scientists:
“we introduce aiXiv, a next-generation open-access platform for human and AI scientists. Its multi-agent architecture allows research proposals and papers to be submitted, reviewed, and iteratively refined by both human and AI scientists. It also provides API and MCP interfaces that enable seamless integration of heterogeneous human and AI scientists, creating a scalable and extensible ecosystem for autonomous scientific discovery. Through extensive experiments, we demonstrate that aiXiv is a reliable and robust platform that significantly enhances the quality of AI-generated research proposals and papers after iterative revising and reviewing on aiXiv. Our work lays the groundwork for a next-generation open-access ecosystem for AI scientists, accelerating the publication and dissemination of high-quality AI-generated research content.“
The platform primarily addresses a growing bottleneck where AI systems can now autonomously generate research but traditional journals struggle to scale review processes (and remain hesitant about AI-generated content).
Narrative: the aiXiv initiative signals the emergence of a parallel scientific infrastructure where AI systems, having outpaced human review capacity, begin creating their own quality-controlled knowledge dissemination networks—potentially the first step toward truly autonomous scientific communities.
Traditional academic institutions and publishers will struggle to compete.Insightful breakdown thread here from Rohan Paul:
Open-source models perform differently depending on their hosting?
Weird finding from Artificial Analysis: when testing benchmarks for OpenAI’s recent gpt-oss-120B model, they found that performance differed depending upon hosting environment… definitely needs more investigation!🤔
🤖 Generative Social Simulation
Researchers of the paper Can We Fix Social Media? Testing Prosocial Interventions using Generative Social Simulation used LLMs embedded in agent-based models to create synthetic social media platforms, finding that even minimal systems with just posting, reposting, and following spontaneously reproduce echo chambers, attention inequality, and amplification of extreme voices.
Testing six proposed interventions—from chronological feeds to bridging algorithms—yielded sobering results: most had minimal impact, and some actually worsened the problems they aimed to solve.
This suggests social media's core dysfunctions may be structurally embedded in the very architecture of networked platforms, not just algorithmic quirks.GPT-5 signals cognitive hyper-abundance
Finally, AI researcher David Shapiro's analysis of METR benchmark data suggests we're not just seeing exponential AI capability growth, but super-exponential acceleration that could enable AI systems to autonomously complete year-long human work projects by 2029. His projections indicate we're approaching "cognitive hyper-abundance" where intelligence becomes essentially free and unlimited, fundamentally reshaping economics and human endeavours within the decade.
Narrative: the concept of an “intelligence explosion” is still so counter-intuitive and downright threatening to most people’s world models that we keep on expecting AI advances to slow down any day now. But the numbers, at least as framed by METR in this case, tell a different story. Definitely worth planning for this scenario in your future…
🔮[Weak] signals
What *isn’t* AI these days…? Quite a lot, this week…
Space
SpaceX Starship nails critical 10th test flight
SpaceX's Starship nailed its 10th test flight after a string of spectacular failures, successfully deploying dummy Starlink satellites, performing an in-space engine relight, and splashing down intact in the Indian Ocean. A crucial bounceback for the world's most powerful rocket, which had lost three Ship upper stages in previous 2025 flights due to various technical gremlins. Spectacular footage: congratulations to the SpaceX team - dominant right now.
Satellite constellations
Great animation from Analysis Mason showing the current state of major satellite constellations in orbit right now:
Consumer tech
📱🧲🤖 Made by Google 2025: Magnets, Pixel 10 lineup, AI, and Everything Else
Google's annual hardware event delivered the Pixel 10 lineup with MagSafe-style magnetic charging, a new Tensor G5 chip, and AI features scattered across every device like confetti at a tech conference. The standout might be the Pixel 10 Pro Fold as the first fully dust-resistant foldable, plus a surprisingly repairable Pixel Watch 4 with standard T2 screws.
The event itself apparently felt like "being sucked into an episode of WandaVision" rather than following the traditional product launch keynote playbook. (Nice to see Google innovating there, not least…)📱iKKO MindOne Pro
iKKO's MindOne Pro packs a full Android 15 smartphone into an 86×72mm frame with dual operating systems, a rotating 50MP camera, and NovaLink vSIM providing free global internet for AI functions across 60+ countries. At US$369 during its final Kickstarter hours, this ultra-compact device targets users seeking focused functionality over endless distractions.
🌐🆓 DigitalPlat FreeDomain: Actually Free Domains
DigitalPlat FreeDomain is giving away genuine domain names (not subdomains) under alternative TLDs like .DPDNS.ORG and .US.KG, backed by a US nonprofit and already serving 200,000+ registered domains. The open-source project offers full DNS control without trials, upsells, or hidden costs—targeting developers, students, and hobbyists who've been priced out of the domain game.
💧🏕️ Lifestraw Escape
Also consumer tech, of a very different sort: Lifestraw has launched the Escape, a 19-litre water purification system that works like a beer keg—pump it up to pressure and dispense clean water through a tap. The US$330 device marks a significant departure from the company's ultralight backpacking roots… but I can see this finding a market in many situations.
e-Commerce
🛍️🌐 Shopify Catalog
Shopify has launched Shopify Catalog, a global product database that lets select partners search across all Shopify merchants' inventories in real-time through AI platforms and shopping apps.
In one move, this transforms Shopify from a collection of individual online stores into a unified, searchable marketplace that can surface products wherever customers (or their AI agents…) prefer to shop.
👀 Shopify writeup here: Enabling AI shopping agents to drive sales for you.
Interesting play: enables Shopify to continue gaining scale against Amazon and Walmart (not to mention Shein and Temu).
Spatial computing
VR in the courtroom
A judge in Florida put on a VR headset to experience a virtual “artists impression” of the crime scene firsthand. Baby steps… but there have been forensic startups for a long time trying to capture a crime scene photographically (for example using Gaussian Splats or similar capture tech) and render in VR.
🦾Robotics
Skild gets an upgrade
Skild AI’s robots just got an Nvidia Jetson Thor robotic brain upgrade… which they can self-install:
🤖💪 Large Behaviour Models: Atlas learns to ignore distractions
Boston Dynamics and Toyota Research Institute have equipped the Atlas humanoid with Large Behavior Models (LBMs), enabling it to learn complex tasks through human demonstrations rather than traditional programming. The latest demo shows Atlas methodically sorting objects while a human deliberately tries to disrupt its work. Pesky human.
💪🇰🇷 ALLEX: South Korea's "all-muscle" humanoid robot
South Korean robotics company WIRobotics unveiled ALLEX, a lightweight humanoid robot that prioritises finesse over brute force, capable of detecting forces as small as 100 gram-force and lifting 3kg with hands weighing just 700g each. Unlike heavyweight competitors like Figure 02 (70kg), ALLEX's modular design targets healthcare, eldercare, and household automation through partnerships with MIT, UIUC, and AI firm RLWRLD. The company aims to deliver a general-purpose humanoid platform by 2030 that "truly experiences and responds to the real world".
🕷️💧 Rhagobot
Scientists have created a tiny robot inspired by water striders that uses feathery fans on its feet to zip across water surfaces, discovering that the insects' propulsion fans deploy automatically via surface tension rather than muscle power.
The 8cm Rhagobot currently moves at two body lengths per second and can execute 90-degree turns in under half a second, though it's still far slower than its biological inspiration which hits 120 body lengths per second.Narrative: Sometimes the most sophisticated engineering solutions require no power at all—just clever geometry and physics.
🕋Smartlets
💬 “Smartlets” are millimetre-sized robots that chat and collaborate underwater.
Researchers at Chemnitz University have invented these fully autonomous robots that can communicate optically, move independently, and coordinate with each other in water environments. These origami-inspired devices pack solar cells, microchips, sensors, and LED communication systems into tiny 3D cubes that can form intelligent swarms without any external control infrastructure.
The breakthrough opens pathways toward distributed robotic systems that could revolutionise everything from medical diagnostics to environmental monitoring—and perhaps edge us closer to artificial life itself.Explainer video:
Modular robots build bridges and shelters outdoors
| Link | AI Analysis
🤖🔗🏗️ Modular robots: Real-world swarm construction
Dartmouth researchers have successfully tested cube-shaped modular robots outdoors, demonstrating their ability to crawl under logs, assemble into bridges across streams, and form scaffolding for temporary shelters with drone assistance. Each Wi-Fi-enabled block can operate for over three hours on a single charge, using adjustable strings and motors to change shape and connect with other units, drawing inspiration from ant colony bridge-forming behaviour. The breakthrough represents one of the first successful demonstrations of modular robotics in varied outdoor environments, potentially revolutionising emergency response and disaster relief operations.
As climate disasters intensify globally, the convergence of swarm robotics and emergency response capabilities signals a future where rapid-deployment infrastructure could be as simple as an airdrop of intelligent building blocks.
Drones, autonomy, weapons
🚁🏆 SiFly's record-breaking 3-hour flight
American drone startup SiFly's Q12 quadcopter just snagged the Guinness World Record for longest electric drone flight in its weight class, staying airborne for 3 hours and 11 minutes on Orville Wright's birthday.
The achievement comes as the FAA considers new Beyond Visual Line-of-Sight regulations that could finally unlock routine long-distance drone operations for delivery and agriculture.🌊HoverAir's waterproof Aqua drone floats and films
Zero Zero Robotics has unveiled the HoverAir Aqua, a completely waterproof drone that can float, take off from water, and capture 4K/100fps footage while tracking you via a wrist-worn "Lighthouse" transmitter. At US$999 for early birds, it's the company's priciest drone yet but solves the age-old problem of filming water activities when your hands are busy paddling.
🏫Florida schools test armed drones against shooters
Florida schools are trialling swarms of armed drones that can reach active shooters within 15 seconds, cruising hallways at 50mph while operated remotely from Austin headquarters. The Campus Guardian Angel system deploys 30-90 drones per school, armed with pepper rounds and glass breakers, promising to clear rooms "like a police dog" in response to the roughly 1,000 school gun violence incidents over the past three years.
Narrative: as usual with this very US-specific issue, a “solution” that feels less like progress and more like admitting they've given up on addressing the actual problem.
Autonomous warfare ontology
“The Autonomous Warfare Ontology (AWO) is a proposed structured framework for defining, categorizing, and standardizing terminology related to unmanned systems and artificial intelligence (AI) in military applications. It aims to establish a shared vocabulary across defense organizations, industry, academia, and international partners to improve communication, interoperability, and policy development in the rapidly evolving field of autonomous warfare.“
🚀China develops ultra-precise military data link for hypersonic warfare
China unveiled a military communications system that achieves 5-nanosecond timing accuracy for coordinating hypersonic weapons travelling at Mach 5-11, outperforming NATO's Link 16 by 100x. The network enables real-time coordination across hypersonic vehicles, ground posts, and satellites by using innovative inertial navigation data sharing rather than traditional timing methods.
Well, yikes.🛩️China's new “stealth” drone rivals B-2 bomber size
Leaked images revealed China has developed a massive flying-wing stealth drone with a 52-metre wingspan—virtually identical to the US B-2 bomber's dimensions. The aircraft, spotted at China's new Malan test facility in Xinjiang, appears designed for high-altitude reconnaissance rather than bombing missions, marking Beijing's entry into the large-scale stealth drone arena previously dominated by American projects like the RQ-180.
🚀X-37B launches on 8th secret SpaceX mission
SpaceX / Wikipedia Commons The US military's mysterious X-37B spaceplane launched on its eighth classified mission aboard a SpaceX Falcon 9, carrying quantum inertial sensors and laser communication tech to test navigation systems that work without GPS.
The 8.8-metre unmanned craft will spend an undisclosed amount of time in low Earth orbit conducting experiments that could enhance spacecraft resilience against emerging threats and support future lunar missions.☢️🚀Russia prepares to test 'unlimited-range' nuclear missile?
Satellite monitoring reveals Russia amassing military assets on Arctic archipelago Novaya Zemlya, suggesting imminent testing of the accident-prone Burevestnik nuclear-powered cruise missile. The deployment includes specialised nuclear monitoring aircraft, fighters, and transport planes—a rare Arctic buildup not seen since the Ukraine invasion began in 2022.
Yikes again.🦩🇺🇦Flamingo
Meanwhile the first sightings of Ukraine’s “Flamingo” long-distance missiles appeared online, first published by Ukrainian outlet Dzerkalo Tyzhnia. The report also claims that the missile has a 1,150 kg warhead and a 3,000 km range and is already in use…
(While Ukraine’s ongoing defense dependency on the US continues to be dragged through the Trump drama machine, the country’s penetrative strikes on factories and infrastructure targets deeper within Russia continue to cause havoc…)🪄AETHRA-3
This is probably the most innovative drone concept that’s stood out to me in a while:
“Meet AETHRA-E...a drone without rotors or wings. Slot microjets embedded in the hull vector a continuous thrust sheet to lift, steer, and stabilize. Edge-blown monocoque: the airframe itself becomes the nozzle. Just shaped thrust. No props. No wings.“ (source: @mathelirium)
🗻🚁♻️ Drones vs. the "Highest Dumpster in the World"
Finally a good-news drone story: heavy-duty DJI drones have successfully airlifted 300kg of rubbish from Everest's Camp 1 at 6,065m, proving 60 times more efficient than human porters while also delivering essential climbing gear and reducing dangerous crossings of the deadly Khumbu Icefall. The US$20,000 drones can carry in 10 minutes what takes 10 people six hours to haul, transforming both waste management and safety logistics on the world's highest peaks.
Narrative: what happens when we scale this approach to other environmental cleanup challenges in remote areas—or even in urban environments?
Biotech
🦠AUN: bacteria take on cancer
Japanese researchers have developed a revolutionary bacterial cancer therapy called "AUN" (named after the Japanese philosophical concept of perfect harmony between opposites) that destroys tumours without relying on the immune system, potentially transforming treatment for immunocompromised patients. The approach uses two bacterial strains working in harmony—Proteus mirabilis and Rhodopseudomonas palustris—that demonstrated remarkable tumour clearance in both mouse and human cancer models.
Geoengineering
🌊⚗️Frontier buys $31M ocean alkalinity carbon removal credits
Antacids for the ocean: Frontier, the carbon removal clearinghouse founded by Google, Stripe, Shopify, and others, just dropped US$31.2 million on 115,208 tonnes of carbon removal credits from Planetary, a startup that's literally giving the ocean antacids to fight climate change.
This is Frontier's first foray into ocean alkalinity enhancement, using magnesium hydroxide to counteract the 30% increase in ocean acidity since the 1800s. At US$270 per tonne today but targeting sub-US$100 costs, this approach could theoretically scale to remove over 1 billion tonnes of CO2 annually—assuming we can stomach the planetary-scale Tums distribution logistics.
ConstructionTech
❄️ Supercool cement: Buildings that chill themselves
Chinese researchers have developed a revolutionary cement that reflects sunlight and emits heat rather than absorbing it, using tiny ettringite crystals to create surfaces that stay 5.4°C cooler than surrounding air. The innovation could potentially achieve net-negative carbon emissions over 70 years whilst dramatically reducing our reliance on energy-hungry air conditioning systems. With buildings consuming 40% of global energy and producing 36% of carbon emissions, could this be the material breakthrough that finally makes our cities both cooler and cleaner?
🏗️ ReCon: Velcro for walls
Austrian researchers have developed a hook-and-loop fastening system that lets you literally yank interior walls out by hand and swap in new ones. The mushroom-head hooks are cast directly into the building's concrete structure, while wall panels get loop strips that simply click into place—no demolition required when it's time for a refresh. Very neat.
🏠🖨️Two-storey home 3D-printed in 18 hours
Contec Australia has 3D-printed a complete two-storey family home in Perth, with both floors printed in just 18 hours using a Dutch CyBe robot and specialised concrete mix. The finished walls are three times stronger than standard bricks and 22% cheaper than traditional masonry construction, though the complete project still took five months including all the human-added finishing touches like roofing, wiring, and flooring.
Particularly noteworthy: this appears to be one of the first genuine two-story 3D-printed homes where both levels were actually printed, rather than having a timber upper floor added later.Narrative: 3D-printed buildings continue to edge closer to commercial reality at scale… only a few years now.
Crypto
US Treasury explores embedding ID checks into DeFi contracts
Cointelegraph reports that the US Treasury is exploring embedding government ID checks directly into DeFi smart contracts under the new GENIUS Act, potentially requiring biometric or wallet verification before any transaction can proceed.
Critics warn this could fundamentally transform permissionless finance into a surveillance infrastructure, with one CEO comparing it to "putting cameras in every living room."
Clearly there is a growing tension between regulatory compliance and the foundational principles of decentralised finance, while potentially excluding billions of unbanked individuals globally.Narrative: Like that wasn’t the intention from the outset…
💶🏃♂️ Digital Euro Sprint Mode
Meanwhile, the EU is suddenly rushing to accelerate its digital euro plans after the US passed landmark stablecoin legislation, with officials now considering running it on public blockchains like Ethereum rather than private ones.
The US$288 billion dollar-dominated stablecoin market has EU bureaucrats worried about losing monetary sovereignty, prompting what one insider described as a "let's speed up, let's push" moment.Narrative: Every country around the world should be taking this as a wake-up call… if they don’t have Stablecoin infrastructure in place then their economy starts to look like a quaint financial backwater pretty soon, with people and businesses just bypassing their national currencies entirely.
Stablecoins could break credit card fee monopoly
One Stablecoin use case to watch: US merchants fork over US$187 billion annually in card processing fees, creating a perverse system where cash-paying customers subsidise credit card rewards for the well-banked. Stablecoins offer a technical solution but lack incentive structures needed to break consumers away from loyalty points and cashback programs.
Narrative: the payment rails revolution isn't waiting for better technology—it's waiting for new merchants to compete on loyalty programs rather than surrender 4% of every transaction to Visa and Mastercard.
Quantum Tech
⚛️💫Scientists build scalable quantum network node using ions
Austrian researchers have built a scalable quantum network node using 10 calcium ions that achieve 92% fidelity when entangling with photons, representing a major leap from previous 2-3 ion setups. The breakthrough enables quantum networks that could connect distant quantum computers, create unbreakable communications, and synchronise atomic clocks across continents with unprecedented precision.
Energy
♻️UK wind farm installs recyclable blades for 1.2M homes
German energy giant RWE and Siemens Gamesa have installed 75 recyclable wind turbine blades at the UK's Sofia Offshore Wind Farm, the country's first large-scale deployment of circular wind technology. The 1.4 GW project, located 195km off England's northeast coast, will power 1.2 million homes when complete in 2026, with blade materials designed to be separated and recycled into everything from car parts to suitcases at end-of-life.
⚛️🔋Nuclear battery achieves 50-year lifespan with triple efficiency
Researchers at Xiamen University the China Institute of Atomic Energy have cracked the code on long-term nuclear batteries, developing strontium-90 radio-photovoltaic cells that deliver triple the efficiency of conventional designs while maintaining performance for half a century. The breakthrough uses a clever waveguide light concentration structure that converts radioactive decay into focused light, then electricity, achieving 2.96% energy conversion efficiency with only 13.8% performance degradation after simulated 50-year radiation exposure.
🐄💩✈️Cow manure to jet fuel
California startup Circularity Fuels has successfully demonstrated technology that converts dairy farm biogas into jet fuel precursors at 1/100th the cost of conventional systems, potentially turning America's billions of kgs of annual livestock waste into aviation fuel. The compact Ouro Reactor can process raw biogas directly on farms without expensive pre-treatment, addressing both the sustainable aviation fuel shortage (currently <1% of demand) and the fact that 94% of farms don't capture their methane emissions.
Narrative: *If* this scales as promised, we might finally have found aviation's holy grail: cheap, distributed renewable fuel production that doesn't compete with food crops.
☀️🏠 Perovskite indoor solar cells
University College London researchers have developed perovskite solar cells that can harvest energy from indoor ambient light with 37.6% efficiency—six times better than silicon alternatives. The breakthrough could eliminate battery dependence for small devices like keyboards, sensors, and IoT gadgets by powering them with nothing more than office lighting.
Health tech
⚡👁️ Electromechanical Reshaping: The LASIK Killer?
Researchers have developed a method to correct vision using tiny electrical currents and platinum contact lenses that reshape corneas in rabbits in just one minute, potentially eliminating the need for LASIK surgery. The technique, called electromechanical reshaping (EMR), works by temporarily altering the pH of corneal tissue to make it malleable, then locking it into the new shape without any cutting or tissue removal. (Of course, there's the small matter of transitioning from rabbit eyeballs to human trials first…)
⏳ Zeitgeist
What I’ve been tracking in the non-tech world, treading as lightly as I can…
🏔️🇨🇳China begins US$170B Tibet hydropower mega-project
China has officially launched its first trillion-yuan infrastructure project—a massive US$170 billion hydropower dam in Tibet's Yarlung Tsangpo Grand Canyon that will generate three times the electricity of the Three Gorges Dam. The project capitalises on a 2,000-metre river drop over 200 kilometres in remote Medog County, targeting China's exploding energy demand driven by AI, data centres, and electric vehicles.
Meanwhile, authorities in neighbouring India – downstream from the project – have expressed concern about China controlling the flow of the river and the impacts that could have across the border.
Climate
Briefly, just headlines and links.
🌪️Vietnam evacuates 325,000 as Typhoon Kajiki approaches with 140km/h winds
🧊Antarctica's abrupt changes will affect world for generations
Antarctica has abandoned its "remote and unchanging" reputation, with sea ice shrinking at double Arctic rates, ocean currents slowing, and ice sheets approaching tipping points that could raise global sea levels by over 5 metres.
🌡️Severe droughts devastate Middle East and Europe
A severe drought stretching from the Middle East through Europe is delivering a harsh economic reality check, with Iraq experiencing its driest conditions since 1933 and the Tigris and Euphrates rivers dropping by over 25%.
And in Zurich (where I’m heading this weekend):
“A new tunnel has opened.
This time it’s not for bicycles or trains. Not even for cars.
It’s an emergency tunnel to protect the city of Zurich from flooding.
It cost the taxpayer 175 million francs. Damage of severe flooding would cost up to 6.7 billion.“ — Renat Heuberger(🎩 Jay Scott for sharing)
🇵🇸⚔️🇮🇱Palestine conflict
Israel launches Gaza City offensive amid truce talks
Israel has launched an offensive to take control of Gaza City while simultaneously engaging in ceasefire negotiations mediated by Egypt and Qatar, with Hamas having already accepted a 60-day truce deal that Israel must respond to by Friday. The parallel military escalation and diplomatic efforts come as UN agencies warn of severely worsening humanitarian conditions and Israeli authorities separately approved a West Bank settlement project that would effectively bisect the territory. When will this end? And what comes after?
☁️⚖️ No Azure for Apartheid
Microsoft employees and community members staged a second day of protests at the company's Redmond headquarters, resulting in 18 arrests including current Microsoft cloud engineer Anna Hattle, as the "No Azure for Apartheid" group demands the tech giant end its cloud contracts with Israel.
The protests escalated with demonstrators establishing a "Liberated Zone" encampment and pouring red paint over Microsoft signage, following a Guardian investigation revealing Israeli government use of Azure to store recordings of up to a million Palestinian calls per hour.😥19,000 names
🌊💩Flooded zone
I have to go there…
🇺🇸⚔️🇪🇺 Trump's Tech Tariff Ultimatum
…but I can’t even…
Trump has threatened "substantial additional tariffs" against any country whose digital regulations or taxes target US tech companies, specifically calling out the EU's Digital Services Act and various digital services taxes as discriminatory. FFS.
Marietje Shaake in the FT: Beware America’s AI colonialism
“AI is an ideological project for the Trump administration and the AI Action Plan lays out a clear blueprint for US technological hegemony…What makes AI dependency particularly dangerous is its opacity. Unlike trade in physical goods, AI decision-making processes are often black boxes, making subtle manipulation nearly impossible to detect. These systems become deeply embedded in critical processes, with high replacement costs. Many countries already have significant dependencies on US tech companies. Add AI and powerful lock-in effects would intensify. The rapid pace of its evolution makes it difficult for alternative suppliers to maintain competitive alternatives, adding chokepoint effects. With the integration of such technologies in infrastructure, defence and security systems, the stakes are high.”
Yes, yes and yes. Pay attention governments *and* businesses!
📉US experiences first net emigration since 1960s
The US is experiencing net emigration for the first time since the 1960s, with its foreign-born population declining nearly 3% in the first half of 2025, according to Pew Research analysis of census data.
(Surely a coincidence…?)
🔍📱🛂 US border device searches hit record high
US Customs and Border Protection searched 14,899 electronic devices between April and June 2025, a 17% jump from the previous record set in early 2022. Most searches are classified as "basic" (a delightfully Orwellian term for "hand over your password so we can rifle through your digital life"), while the constitutional debate around these practices remains unresolved by the Supreme Court.
Narrative: On that long-haul flight…. plenty of time to reset your phone and then reinstall a fresh OS image after you clear the border. (Plus it’s good security practice to regularly wipe and reinstall your phone as well!)
🧠Mind expanding
In amongst all this, I’ve been having a few moments to do some thinking:
🔥📈🧠 Self-reinforcing cascades: The maths of going viral
Schematic of a self-reinforcing cascade DOI: https://doi.org/10.1103/5mph-sws5 This is fascinating: researchers from the University of Vermont and Santa Fe Institute have developed a mathematical model explaining why some ideas explode across the internet while others die quietly. The secret is that viral content doesn't just spread, it evolves and strengthens as it moves through networks.
Unlike traditional models that assume static transmission, this "self-reinforcing cascade" theory accounts for how memes, beliefs, and even misinformation mutate and improve their viral fitness in real time, naturally producing the fat-tailed distributions we see on social media.Narrative: I have long thought of “viral memes” as an engineering problem… this study provides better maths than just K, the viral coefficient.
🌪️ AI intelligence explosion could trigger epistemic collapse
Ben Norman in LessWrong argues that superintelligent AI could trigger "epistemic collapse" - a complete breakdown of humanity's shared frameworks for determining truth, going far beyond manageable "epistemic disruption." The author warns that traditional AI safety solutions like fact-checking and forecasting assume epistemic stability that won't exist when fundamental beliefs are being overturned monthly, potentially creating a "naturalist underclass" of cognitively unenhanced humans excluded from civilisational decisions.
Effective Altruist Will MacAskill’s take on the topic:"A mediaeval king suddenly needing to upgrade from bows and arrows to nuclear weapons to deal with an ideological threat from a country he's never heard of, while simultaneously grappling with learning that he descended from monkeys and his god doesn't exist.”
Ted Gioia has a similar take: Shared reality will collapse within 12 months
⏰AI threatens economic relevance within thousand days
Finally, Emad Mostaque, founder of Stability AI and former hedge fund manage has put his thoughts into a book, The Last Economy, arguing that we have roughly 1000 days before AI triggers an irreversible socio-economic phase transition that will render most humans economically obsolete.
He presents three possible futures:
Digital Feudalism (corporate AI control with UBI),
Great Fragmentation (nationalist AI silos)
Human Symbiosis (AI amplifying rather than replacing human purpose)
…with the first being the current default trajectory.
The book promises to introduce "Intelligence Theory" as a new economic framework for navigating what he calls the "Abundance Trap" - where post-scarcity intelligence gets processed as poverty by our scarcity-based systems.
Big thinker, that man.
💭Meme stream
And finally… some light-hearted distractions…
🧞Micro-wishes to a genie
Ain’t this the truth:
🏔️🔴 Valles Marineris: Mars' geological time capsule
NASA's Mars Reconnaissance Orbiter has captured new images of Candor Chasma within Valles Marineris—the solar system's largest canyon system stretching 4,000km across Mars' equator. Spectacular.
📰'Pregnancy robot'
A viral post about China developing a "pregnancy robot" for surrogacy turned out to be completely fabricated, complete with AI-generated images and a fictional inventor named Zhang Qifeng who supposedly graduated from Singapore's Nanyang Technological University.
Multiple news outlets from the Daily Mail to Korean and Indian publications ran with the story before Snopes debunked it, with NTU confirming no such person or research ever existed.
But if it *was* real… people would want it, right? Remember futurist Hashem Al-Galili’s concept video Ectolife from 2 year ago?
🌳Wonderwoods
Stefano Boeri Architetti has completed the 104-metre Wonderwoods Vertical Forest in Utrecht, featuring 360 trees and 50,000 plants across 30 native species—equivalent to a hectare of actual forest wrapped around 200 apartments. The building integrates IoT sensors for plant health monitoring, rainwater capture systems, and purpose-built bird nesting cavities, representing a mature evolution of the vertical forest concept beyond mere aesthetic greenwashing.
I think it’s pretty beautiful.
☕️😊 Morning Coffee Found To Boost Happiness
A four-week study of 236 German young adults confirmed what millions already suspected: morning coffee genuinely boosts mood and energy levels.
Well, of course.☕️😊
🤡AI executives demand free electricity
Top quality satire (🎩 Tim B for sharing):
👽Meet Merv
This is top comedy… (video via @venturetwins)
🎥Forgive the haters
Finally: Matt Zien captures the Generative AI zeitgeist pretty skilfully, lots of layers in here.
🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback. A huge one this week, took much longer than planned… hope you found a few nuggets of gold therein!
Liebe Grüße aus Tirol
Ben