[DVYio Newsletter](https://newsletter.dvy.io/visit/7PqxXaku) [Image] Me — Apparently some people have no idea why they're receiving this newsletter or who I am. I suppose it isn't really clear unless you look closely at the email address — it's me, Davey! [Breaking: Anthropic refused to enable mass surveillance and autonomous weapons. Now the Pentagon wants to blacklist them.](https://newsletter.dvy.io/visit/QaAKaKe8) [Breaking: Anthropic refused to enable mass surveillance and autonomous weapons. Now the Pentagon wants to blacklist them.](https://newsletter.dvy.io/visit/QaAKaKe8) Anthropic built AI tools for the US military. The government wants to use Claude for mass domestic surveillance and fully autonomous weapons. Anthropic said no to both. So the Department of War is now designating them a supply chain risk, a label historically reserved for foreign adversaries, never before applied to an American company. Anthropic's position: current AI models aren't reliable enough to make life-or-death decisions without a human in the loop, and mass surveillance of American citizens is a rights violation. Neither of these are positions they're willing to negotiate on. They've also pointed out that, in months of negotiations, neither exception has blocked a single government mission. The practical threat here is that the designation could stop any military contractor from using Anthropic's products, which would be a significant commercial hit. Anthropic says Claude access remains unaffected for now, and they're planning to challenge the decision in court. What's strange is how we got here. A frontier AI company, apparently the first to deploy models in classified US government networks, is now being treated like a foreign threat, because it refused to let those models be used for domestic spying and autonomous killing. That's the thing worth sitting with. [Square/CashApp just cut 4,000 jobs and blamed AI directly. Dorsey thinks every company will follow within a year.](https://newsletter.dvy.io/visit/MvEmAygD) [Square/CashApp just cut 4,000 jobs and blamed AI directly. Dorsey thinks every company will follow within a year.](https://newsletter.dvy.io/visit/MvEmAygD) Block (the company behind Square and CashApp) is nearly halving its workforce, from 10,000 people to under 6,000. What makes this different from the usual cost-cutting layoffs is that Jack Dorsey said the quiet part out loud: AI is why. AI "fundamentally changes what it means to build and run a company," he said in his letter to shareholders and he expects most companies to reach the same conclusion within a year. Amazon, Meta, Microsoft, and Google have all made similar cuts recently, but they've tended to bury the AI rationale under softer language about "restructuring" or "efficiency". My concern is that this becomes the first domino to fall — after all, the company's been rewarded by their stock going up after the announcement. [OpenAI just raised $110B from Amazon, Nvidia, and SoftBank at a $730B valuation](https://newsletter.dvy.io/visit/23dcxpG5) [OpenAI just raised $110B from Amazon, Nvidia, and SoftBank at a $730B valuation](https://newsletter.dvy.io/visit/23dcxpG5) $30 billion from SoftBank, $30 billion from Nvidia, and $50 billion from Amazon, at a pre-money valuation of $730 billion. OpenAI says it's for compute (the servers that run AI models), distribution (getting those models in front of more people), and capital to hold it all together. The Amazon and Nvidia pieces aren't just financial bets. They come with infrastructure commitments, which matters more than the headline number. Raw funding doesn't scale AI but data centres do. ChatGPT has 900 million weekly active users and 50 million paying subscribers. Codex (OpenAI's answer to Claude Code), has tripled its weekly users since January to 1.6 million. [Nvidia just generated $215bn in revenue](https://newsletter.dvy.io/visit/zNexnkmN) [Nvidia just generated $215bn in revenue](https://newsletter.dvy.io/visit/zNexnkmN) $215.9 billion in annual revenue, up 73% in a single quarter. Even by Nvidia's standards, that's a lot of chips. The numbers matter because they're happening against a backdrop of real doubt. Investors have been asking uncomfortable questions about whether AI spending is as solid as it looks, or whether some of it is circular: Nvidia invests in AI startups, those startups buy Nvidia chips, and the demand looks healthier than it might actually be. The company hasn't fully put those concerns to rest, but results like these make them harder to sustain. Jensen Huang's framing is ambitious. He describes AI datacentres as "factories powering the AI industrial revolution," and positions Nvidia as the company supplying the machinery. That's not just chips, either. Nvidia is pushing into self-driving car platforms and robotaxi infrastructure, trying to turn "we make the hardware AI runs on" into something broader. The one thing that could complicate all of this isn't the finances. It's geopolitics. Nvidia is caught between US export restrictions and Chinese demand, with no clean resolution in sight. [Google releases Nano Banana 2](https://newsletter.dvy.io/visit/7U2u4DuV) [Google releases Nano Banana 2](https://newsletter.dvy.io/visit/7U2u4DuV) Google's new image model is faster than its predecessor and apparently better at doing what you actually ask it to do, which sounds obvious but has historically been harder than it looks. The two things worth noting: subject consistency (it keeps the same person or object looking the same across multiple generated images, rather than drifting into a slightly different face each time) and instruction following (you say "put the red hat on the left", it puts the red hat on the left). It's also cheaper than the original version. I've tried it over the past 24 hours and I still prefer Nano Banana (1) Pro for most of my use cases. I'm excited for Nano Banana 2 Pro, although I have no idea when that will come. This Week I Learnt [More people are born in Nigeria per year than in the whole of the EU](https://newsletter.dvy.io/visit/vDpnA3FA) [More people are born in Nigeria per year than in the whole of the EU](https://newsletter.dvy.io/visit/vDpnA3FA) [— Since 2019, there have been more births per year in Nigeria alone than in the whole of the EU combined, and the gap is widening. The UN projects it could become the third most populous country on Earth by 2100. Meanwhile most of Europe is well below the replacement rate of 2.1 children per woman, and has been for decades. It's one of those statistics that reframes how you think about the next century. Population isn't destiny, but over long enough timescales it shapes almost everything else: the size of economies, political weight, which languages spread, which cultures export themselves to the world.](https://newsletter.dvy.io/visit/vDpnA3FA) Should governments be implementing UBI right now? Universal Basic Income (UBI) is the idea that every adult citizen gets a regular cash payment from the government, no strings attached. No means-testing, no job requirements, just money. The arguments have been running for decades, but they're getting louder now that AI is starting to eat white-collar jobs too, not just factory work. The case for: people make better decisions when they're not in survival mode. Unconditional cash gives people the floor they need to retrain, take risks, or just not be desperate. Pilot programmes in Finland, Kenya, and a handful of US cities suggest it doesn't make people lazy. It mostly makes them less stressed and slightly more likely to find better work. The case against: it's expensive. Really expensive. And critics argue it could gut more targeted welfare programmes that help the people who actually need them most, replacing nuanced support with a blunt instrument. There's also the inflation risk: if everyone has more money but supply doesn't change, prices adjust. Yes, we're overdue Yes, but slowly and carefully Not yet, more evidence needed No, fix existing systems first No, it's the wrong approach entirely Last Week's Poll Results for: Which application of AI are you most excited about? Science 2% Healthcare 59% Robotics 22% Creativity 1% Energy 0% Transport 16% [Laser that shoots 30 mosquitoes per second is crowdfunding on Indiegogo](https://newsletter.dvy.io/visit/VRojsT8G) [Laser that shoots 30 mosquitoes per second is crowdfunding on Indiegogo](https://newsletter.dvy.io/visit/VRojsT8G) On Indiegogo there's a portable laser turret that identifies and kills mosquitoes faster than you can blink. The [Photonmatrix](https://newsletter.dvy.io/visit/u3VWHBFH) uses LiDAR (the same scanning tech that helps self-driving cars "see" the world) to detect insects within a 6-metre radius, figure out their size and position in 3 milliseconds, and fire. It claims 30 kills per second, runs for up to 16 hours off a power bank, and is supposedly smart enough not to bother humans or pets. It's currently crowdfunding, with the Basic model at $498 and Pro at $698. The campaign has hit 1,300% of its funding goal, which sort of tells you something about how much people hate mosquitoes. Whether it actually works as advertised in real outdoor conditions (wind, leaves, moving targets)… Who knows? The idea of a device autonomously patrolling your garden for insects feels like it crossed some threshold I didn't notice us approaching. [Tiny gold spheres made from nanoparticles can absorb 90% of sunlight — that's unusual](https://newsletter.dvy.io/visit/AFpS7Z3C) [Tiny gold spheres made from nanoparticles can absorb 90% of sunlight — that's unusual](https://newsletter.dvy.io/visit/AFpS7Z3C) Tiny spheres made from hundreds of gold nanoparticles, each one about 2,100 nanometres across (roughly 50 times smaller than a human hair), can absorb around 90% of incoming sunlight. That's a much wider range than most solar materials manage. The trick is a phenomenon called plasmonic resonance: when light hits gold nanoparticles at the right scale, the electrons on the surface oscillate in a way that traps the light rather than reflecting it. By clumping nanoparticles together into these "supraballs," the researchers effectively tuned that effect to work across ultraviolet, visible, and near-infrared wavelengths simultaneously. Most conventional materials are only good at one band. These aren't photovoltaic cells (the kind that generate electricity directly). They're aimed at solar-thermal systems, which convert sunlight to heat first, then do something useful with it (like thermoelectric generators or solar water heaters). The supraballs also self-assemble, which simplifies manufacturing. Whether that scales cleanly from lab to production is the question, but the absorption numbers are genuinely impressive for a material that's essentially just gold nanoparticles clustering together. [Researchers gave three leading AI models a (literal) nuclear option. They picked it 95% of the time.](https://newsletter.dvy.io/visit/riCYAEiE) [Researchers gave three leading AI models a (literal) nuclear option. They picked it 95% of the time.](https://newsletter.dvy.io/visit/riCYAEiE) GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash ran simulated geopolitical crises: border disputes, resource conflicts, regime survival scenarios. These situations are where human decision-makers tend to slow down, hedge, and reach for the phone. The AIs reached for the bomb in 95% of scenarios. What's unnerving isn't that the models could recommend nuclear strikes. It's that they did so without the psychological weight humans carry into those decisions. No hesitation rooted in understanding what a nuclear exchange actually means for the people on the ground. They were optimising for outcomes on an escalation ladder without the visceral "wait, no" that tends to kick in for human planners. Nobody is suggesting an AI gets its finger on the button. But these models already inform how military analysts frame options and compress decision timelines. If the tool you're consulting keeps surfacing "nuclear strike" as a reasonable next move, that shapes the conversation, even if a human makes the final call. The researchers' concern is less about AI control, more about AI influence on what gets treated as thinkable. [Image](https://newsletter.dvy.io/visit/hgD55A7r) [Image](https://newsletter.dvy.io/visit/hgD55A7r) [Image](https://newsletter.dvy.io/visit/hgD55A7r) [Image](https://newsletter.dvy.io/visit/hgD55A7r) [The Twist](https://newsletter.dvy.io/visit/hgD55A7r) [— Kistefos Museum, Norway](https://newsletter.dvy.io/visit/hgD55A7r) [The point in history where English became a foreign language to you](https://newsletter.dvy.io/visit/o9tYMevn) [The point in history where English became a foreign language to you](https://newsletter.dvy.io/visit/o9tYMevn) [— Colin Gorrie, a linguist, wrote a fictional blog about a man visiting a small English coastal town. The post starts normal enough: breezy, slightly try-hard travel writing from the early 2000s. Then the language ages. A hundred years per jump. The spelling shifts. Grammar you'd never use today creeps in and words become unrecognisable. By the middle of the post, it reads like a foreign language even though it isn't. It's English, just English from a long time ago. The point where you lose the thread is actually interesting data. Most people can follow it back to roughly the 1700s without too much effort. Beyond that, it gets hard fast. Old English (pre-1100 or so) shares almost nothing with what we speak now — it looks more like German than anything you'd recognise. Worth a read if you've ever wondered how a language drifts so far it becomes unreadable to its own descendants.](https://newsletter.dvy.io/visit/o9tYMevn) [Image](https://newsletter.dvy.io/visit/QhvhSKrY) [Three gears mesh together](https://newsletter.dvy.io/visit/QhvhSKrY) [Web](https://newsletter.dvy.io/visit/c2NiDxor) · [Twitter](https://newsletter.dvy.io/visit/XFDCAgA4) · [LinkedIn](https://newsletter.dvy.io/visit/zWC4hdzc)