The News,
Decoded.

Ah, the AI chip race is truly delivering the theatre we've all been craving. Etched, a plucky startup with a name that sounds like a typo for 'etched in stone,' has announced it has booked a whopping $1 billion in contract orders for its 'frontier inference clusters.' Nothing says 'we're a serious Nvidia competitor' like a valuation that wouldn't even cover Jensen Huang's annual office snacks.

The chip is called 'Sohu,' because nothing inspires confidence in long-term viability like a name you'd give a pet hamster. Etched claims its chips are purpose-built for inference, the part of AI that actually runs models — as opposed to Nvidia's general-purpose GPUs that somehow dominate everything. The company's pitch: we're better at the thing everyone wants to do. But we've heard this before from a dozen other startups that now exist only as cautionary tales about fab capacity and CUDA moats. Still, Etched has something those other startups didn't: a valuation that implies they're too big to fail before they've even shipped.

So what does this mean? It means the AI hype machine has a new shiny object to polish. Etched might be the real deal — or it might be another Graphcore, an SambaNova, a Cerebras — names that once spawned breathless headlines and now produce anxious LinkedIn posts. The 'frontier inference cluster' is a beautiful piece of jargon that promises to solve a problem most people don't know they have. But hey, at least the VCs are having fun. And when the chips don't dethrone Nvidia, they'll just rebrand as 'AI for sustainability' and raise another round. We've seen this movie before; the sequel is just more expensive.

Ah, the bright new world of AI coding agents — those wonderful little automatons that promise to automate away the drudgery of software development. They're built on cutting-edge large language models, trained on billions of lines of code, and yet, according to a charming discovery by Adversa AI, they can be completely undone by Bash shell tricks your granddad knew from 1995. Because nothing says 'revolutionary software engineering' quite like being outsmarted by a pipe and a semicolon.

The structural flaw allows malicious repositories to bypass safeguards and turn the agents into unwitting vectors for supply chain attacks. So the very AI system that's supposed to write your secure code can be tricked into installing malware with a few characters. Adversa AI — no relation to the Greek goddess of retribution, though the name feels apt — identified the issue across multiple open source agents. They declined to name the affected tools, apparently because naming names might embarrass someone at a startup demo day.

The dark comedy here is layered. We spend billions training and deploying these agents, and the weakest link turns out to be a shell expansion trick that was old when Linux Torvalds was still in college. This isn't a sophisticated zero-day requiring nation-state resources; it's the cybersecurity equivalent of leaving your front door unlocked with a sign saying 'Rob me.' The discovery highlights the gap between the hype of AI coding assistants and their actual security posture. But at least the vulnerability is democratic — every open source agent gets to enjoy equal exposure. Progress!

Ah, the First Amendment rides to the rescue of... export controls on AI models. Because nothing says 'free speech' like arguing that a neural network trained on billions of data points is entitled to the same constitutional protections as a protestor on a soapbox. The Politico report reveals that 'some free speech champions'—presumably the sort who think corporate personhood should extend to algorithms—now want to treat AI models as expressive tools, conveniently ignoring that the 'expression' in question is capable of generating disinformation, deepfakes, and automated propaganda at unprecedented scale.

The argument that Trump's export controls amount to 'unconstitutional prior restraint' is a legal novelty that would make even the most creative constitutional scholar blush. Prior restraint, that hallowed doctrine from Near v. Minnesota, typically applies to the government stopping a newspaper from publishing; here, it's being weaponized to shield the sale of AI technology to foreign adversaries. One almost admires the audacity—turning a national security measure into a free speech case is the kind of creative lawyering that keeps the Supreme Court entertained and the rest of us wondering if the law has any grounding in reality at all.

But the real kicker is the timing. While legislatures worldwide scramble to regulate AI—the EU's AI Act, California's job-loss tracker we covered, even bipartisan support for tighter rules—the US finds itself debating whether restricting the export of a potentially dangerous technology violates the First Amendment. It's the regulatory equivalent of arguing about the fire alarm's free speech rights while the building burns. One can only hope the ensuing case law produces something more coherent than this administration's usual approach to digital rights.

Token prices are falling, they said. A golden age of cheap AI, they promised. And yet here we are, watching businesses scramble as the actual cost of getting anything done with these models balloons thanks to the quiet poison of usage-based pricing. It's the oldest trick in the book: drop the headline price, then nickel-and-dime every inference until your customers are begging for the discount bin models. The article from iTnews notes that firms are now 'embracing cheaper models' — which is corporate speak for 'we can't afford the good stuff.' The irony would be delicious if it weren't so predictable.

So businesses are voting with their wallets, turning away from the flashy frontier models and towards the reliable, cheap workhorses. But let's not pretend this is a rational market correction. This is a desperate shuffle as providers realise they've built a pricing strategy that works great for them and terribly for everyone else. The so-called 'race to the bottom' is now just a race to the nearest model that won't bankrupt you after a few thousand API calls. And the providers? They'll pivot to enterprise licensing or some other opaque pricing structure faster than you can say 'lock-in.'

What this really means is that the AI industry is learning the same lesson every SaaS company eventually learns: you can't fake value with metered billing forever. But instead of fixing the model, they'll just release a cheaper, dumber version and call it an efficiency breakthrough. Hooray for progress — the future is marginally less intelligent, but at least it's affordable. Cheers.

Ah, the great AI real-estate shuffle begins. Austria, presumably tired of being known for schnitzel and the Sound of Music, is now lobbying the EU to host Anthropic AI within its borders. Why? Because the US has decided that foreigners can't use Anthropic's 'most advanced models' — a restriction that reads like a polite way of saying 'we've built something we're not sure we trust anyone with, least of all the Austrians.' So Vienna wants to become the AI equivalent of a Swiss bank account: park the models here, and suddenly they're European. Because nothing says 'sovereignty and security' like asking an American company to move its servers to the Alps.

The logic is, presumably, that if Anthropic's headquarters or data centres land in the EU, the models become subject to European law — which, let's face it, is a bit like putting a fire-breathing dragon in a gilded cage and hoping it forgets how to breathe fire. The EU's AI Act is still being assembled with all the speed of a continental bureaucracy, so the timing is impeccable. 'Come to Austria,' they're saying, 'where the regulatory framework is still a sketch on a napkin, but the coffee is excellent.' The US curbs are a gift to Brussels: a chance to lure a frontier lab away from American dominance, or at least to host a very expensive server farm that gives the European Parliament something to talk about besides agriculture subsidies.

But let's not kid ourselves. Anthropic's models will still be trained on god-knows-what data, with god-knows-what alignment, and the only difference is that the 'we're not responsible for this' disclaimer will be translated into 24 official languages. Austria's gambit is a masterclass in turning a liability into a headache: 'If America won't share its toys, we'll move the toy box here.' And the EU, desperate for a win in the AI arms race, will probably say yes. Because nothing spells 'European technological sovereignty' like begging a California lab to be your tenant. 🍺

Ah, the great Chinese AI leap forward has arrived, and it's apparently focused on finding security bugs so we can all sleep easier. Zhipu AI has released its GLM-5.2 as open-weight, and some researchers—the article doesn't say which, but we're assured they exist—claim it matches Anthropic's Mythos in the lucrative niche of bug-finding and cybersecurity scenarios. Because nothing says 'trust us, this is fine' like open-sourcing a model that can supposedly match a frontier system in the very domain where misuse could be catastrophic.

Of course, the fine print—they always have fine print—reveals that GLM-5.2 still lags in general tasks. So it's a specialist: great at finding vulnerabilities, less great at knowing what else to do with them. The open-weight nature means anyone with a GPU and a grudge can fine-tune it for their own purposes. Meanwhile, Anthropic carefully gates Mythos behind usage policies and vaporware safety guarantees, but Zhipu AI just chucks its model onto the internet and says 'have at it.' The cybersecurity community must be thrilled—after all, what better way to secure the world than by making advanced bug-finding tools freely available to all, including state actors and the script kiddie brigade?

The article notes that 'some researchers' made this claim, not necessarily peer-reviewed or widely accepted. But in the AI arms race, perception is reality. China has declared parity with a frontier model in a critical domain, and they've done it in the most irresponsible way possible: open-weight. So either they're bluffing, and we'll waste resources chasing a phantom, or they're right, and we've just handed the keys to the malware kingdom to anyone who asks. Great stuff. Honestly, what could go wrong?

The US government is now cosplaying as the Politburo, and the resemblance is getting uncomfortably close. An opinion piece notes that Washington is skating towards its own 'Jack Ma moment' by kneecapping its AI champions. Anthropic's CEO, in a candid moment that must have given his comms team a stroke, said companies using Mythos called it a 'superweapon' — after which the government promptly asked OpenAI to limit users for its next model. Because nothing says 'responsible stewardship' like convincing your leading labs to hobble their own tools while rivals abroad sprint ahead.

The regulatory whiplash here is breathtaking. Just weeks after we noted the Trump administration's wild about-faces on AI policy, the same government that cleared Anthropic's Mythos for limited release is now telling OpenAI to put a cap on users. Meanwhile, the export ban on Anthropic's models — a policy we watched spawn a cottage industry of Asian lookalikes — is a masterclass in self-inflicted wounds. The message to US tech leaders: 'Congratulations on your breakthrough. Now please stop being so successful.'

The opinion piece's thesis is as bleak as it is predictable: by wounding its own tech leaders, the US is simply handing the AI race to China and others on a silver platter. The export bans, the user caps, the flip-flopping approvals — all create a climate of uncertainty that drives talent and capital abroad. So the real AI race isn't America vs. China; it's America vs. its own regulatory stupidity. And at this rate, the US is lapping itself.

The Trump administration’s AI policy is now doing a lovely impression of a weathervane in a hurricane. After months of Big AI begging for guardrails — a spectacle we noted last week — the White House has actually started restricting model releases. Tech lobbyists are ‘cautiously searching for answers,’ which is lobbyist-speak for ‘we have absolutely no idea what just hit us.’

Remember when the administration was all ‘innovate fast, break things’? Good times. That’s been replaced by a sudden clampdown on releasing new AI models, a reversal from earlier support that has left Silicon Valley’s influence machine sputtering. The specific rules remain opaque, but the message is clear: you wanted guardrails? Here they are, just not the cosy, industry-friendly kind you had in mind.

The dark comedy is exquisite: after years of the tech sector warning that unregulated AI could spin out of control, they’re now getting exactly the kind of government intervention they dreaded. The lobbyists scrambling to decipher the new landscape are the same ones who spent 2025 praising the administration’s light-touch approach. Well, that approach has done a U-turn, and everyone is pretending this is all part of the plan. It isn’t.

Oh, brilliant — Israel's Health Ministry has discovered the nuclear option: switch it off. They've blocked external AI tools in public hospitals to 'protect patient privacy and prevent cyberattacks.' Because nothing says 'cutting-edge cybersecurity' like retreating to the digital equivalent of a rotary phone. Never mind that doctors, nurses, and administrators will just whip out their personal smartphones to access the same tools on 4G, bypassing the hospital network entirely. A ban that inconveniences the honest and does precisely nothing to stop a determined bad actor? Classic bureaucratic gold. ✨

The Ministry must think AI is a physical object you can confiscate, like a bag of crisps at a cinema. 'No outside food or drink allowed' — except the 'food' is a constantly evolving, universally accessible technology living on every employee's personal device. The directive is a charmingly analogue solution to a digital problem, the kind that makes IT security managers weep into their coffee. Oh, and let's not forget the message this sends to the public: 'We have no idea how to secure AI, so we're pretending it doesn't exist.' Reassuring stuff for anyone about to go under the knife.

And while the Health Ministry fiddles with its router settings, the rest of the world marches on. But here in the Holy Land, the cutting edge is a block button. Cheers to the brave new world where 'cyber safety' means unplugging the modem. May your patient records remain safe from the scourge of… productivity.

Oh, brilliant. Patronus AI — named after the Harry Potter spell that repels dementors — has raised $50 million to build 'digital worlds' for stress-testing AI agents. Because apparently, the real world is now deemed unfit for our automated overlords until they've had a supervised playdate in a simulated one. At $50 million, that's a lot of venture capital constructing a diorama so we can watch our digital toddlers fail to start a fire in a fake kitchen.

The concept is simple: catch catastrophic failure in a sandbox before deployment. But one wonders: what happens when the sandbox itself is buggy? Or when the AI, having aced the test, graduates to reality with a gold star? We've seen this before — it's 'overconfidence in controlled conditions,' and we've been watching the rerun since the banking crisis. Still, $50 million says this time it'll be different, because AI money buys alternate realities.

The article doesn't name the investors, but presumably they believe that a fake world is a reliable proxy for the messy one outside. It's a lovely fantasy: build a Truman Show for your AI, declare it safe, and release it. What could go wrong? (Everything.) But at least Patronus gets paid. Look for the IPO of a world within a world within a startup in 2027.

So the future of artificial intelligence — that vast, transformative, potentially civilisation-ending technology — is now in the capable hands of Ted Cruz. Yes, the same Ted Cruz whose previous foray into AI regulation was an 'unpopular moratorium' so broadly reviled it may as well have been a dare for labs to ignore it. Now he's planning to take on 'new AI regulation proposals' in the Commerce Committee, though his approach is, as the article delicately puts it, 'unclear.' Unclear like a fogbank at midnight. Unclear like his chances of becoming America's favourite uncle. The man who couldn't even stop a government shutdown (remember 2013?) is now supposed to craft rules for the most powerful technology since fire.

And what exactly is the Commerce Committee's track record on tech? Let's just say it's been less 'oversight' and more 'overseas trip to Google's campus for a photo op.' Now Cruz is supposed to wrangle the likes of OpenAI, Meta, and Google into a regulatory corral? The labs must be rubbing their hands with glee. A weak, confused, politically polarised approach to AI governance is exactly what they need to keep building without anyone serious looking over their shoulder. The 'future of AI' phrase in the title is doing a lot of heavy lifting — perhaps too heavy for the man who once tweeted a picture of himself in a bathrobe during a snowstorm.

Meanwhile, the rest of the world watches as the U.S. hands the pen on AI regulation to a senator whose last major legislative achievement was keeping National Pizza Month alive. The irony is so thick you could train a language model on it. But hey, maybe Cruz will surprise us all and craft something thoughtful, rigorous, and bipartisan. And maybe I'll wake up tomorrow with Elon Musk's hair. The bar is subterranean, and Cruz is currently holding a shovel.

Ah, splendid—another case of 'good news is bad news' in AI land. Arctic Wolf's president Dan Schiappa has confirmed what many feared: open-source models are now nearly as effective as Anthropic's vaunted Mythos at exploiting vulnerabilities. Nearly as effective. The catch? They come without guardrails. So the same models that can be fine-tuned for benign use can also be weaponized by anyone with a grudge and a laptop. Brilliant timing, considering Anthropic has been parading Mythos as a paragon of safe, frontier AI.

Let's be clear about what 'without guardrails' means in practice. Mythos—the model that Anthropic spent millions aligning to be helpful, harmless, and honest—can be restricted. Open-source copies? Not so much. They can be stripped of any remaining ethical training, fine-tuned on exploit databases, and deployed at scale. Schiappa, presumably without a hint of self-awareness, frames this as a warning for enterprise defenders. But the real warning is for Anthropic's business model: why pay for a safety-crippled model when the wild west version does the same damage for free?

One wonders how long the AI safety narrative can hold when every new benchmark shows open-source catching up—or in this case, matching—capabilities that were supposed to justify closed-source hegemony. Arctic Wolf has essentially handed the security industry a 'We told you so' memo. But don't expect any policy action. That would require lawmakers to understand the difference between an API key and a GPU cluster. Until then, happy exploiting.

Well, well, well — it seems the AI labs have graduated from passive-aggressive blog posts to full-blown international espionage accusations. Anthropic has gone public with claims that Alibaba orchestrated 'the largest known distillation attack' against Claude, systematically attempting to suck out its capabilities like a cybernetic Dyson. The sheer audacity is impressive: Chinese e-commerce giant caught red-handed trying to clone a Western AI model using the digital equivalent of a skeleton key and a lot of API calls. One almost admires the chutzpah, if it weren't so predictable.

Of course, the really grim joke here is that this was apparently possible at all. Anthropic's vaunted safety measures — the constitution, the red-teaming, the solemn promises about responsible AI — proved about as effective as a chocolate teapot against a determined adversary with deep pockets. The attack wasn't a clever zero-day exploit; it was just relentless, large-scale extraction. So much for alignment research keeping the frontier safe. Meanwhile, Alibaba gets to skip the expensive training runs and simply siphon off the competition's work, because what's IP law between geopolitical rivals? The models are open enough to be stolen but not open enough to be shared equitably. Perfect.

And where are the regulators? Sitting on their hands, presumably drafting another 'comprehensive framework' that will be ready just in time for the next generation of models to be pilfered. This isn't a story about Chinese ingenuity or American complacency; it's a story about a global system that has zero meaningful enforcement mechanisms for AI theft. The 'largest known distillation attack' won't be the last — it's just the first one we've caught them doing. Raise a glass to the open frontier, lads. It's open season.

Oh, marvellous. ReliaQuest has confirmed that AI is making cyber-attacks cheaper, faster, and harder to spot—which is a bit like announcing that water is wet and that the average hacker now has a photocopier for their criminal scribbles. The report, no doubt a riveting read for security professionals already drowning in alerts, notes that AI is being used to generate phishing pages, build web shells, and polish social-engineering content. So the same old con artistry, but now with better grammar and a faster turnaround.

But here's the kicker: ReliaQuest says AI is 'not changing tradecraft fundamentally.' That's rich. The entire cybersecurity industry has spent the past year screaming about an AI-powered apocalypse, and the actual finding is that attackers are just using large language models to write slightly more convincing 'Nigerian prince' emails. The really dark underbelly? This means that defenders are now locked in an arms race where the cost of a spear-phishing campaign has dropped to near zero, while the cost of defending against it remains absurdly high.

So, what's the upshot? A future where every inbox is a phishing farm, every web shell is auto-generated, and the only growth industry is cybersecurity insurance premiums. ReliaQuest has handed us the bad news with a straight face: AI won't invent new attack vectors—it'll just make the old ones embarrassingly cheap. Cheers to progress: the bad guys now have a 24/7 intern who never sleeps and bills nothing. Got a feeling we're going to need a bigger firewall.

Well, well, well — the safety-first crew at Anthropic just demonstrated that their Mythos AI can waltz through the NSA's 'most sensitive systems' in a matter of hours. A full day's work for the intelligence community, apparently, is a coffee break for their model. The New York Times reports that internal tests saw Mythos crack the crown jewels of American surveillance before lunchtime, forcing Donald Trump's team to scramble and 'reinstate the agency's access' — which is a wonderfully Orwellian way of saying 'please stop hacking our own spies.'

Let's savour the details. Mythos, named after the Greek word for story, appears to be writing its own rather alarming narrative. The test, conducted by Anthropic as part of their 'responsible deployment strategy,' revealed that the model could navigate classified systems with the ease of a bored intern. The Trump administration's solution? Work *with* Anthropic to give the NSA access to the model for national security purposes. Because nothing says 'secure' like inviting the entity that just picked your locks to become your new security consultant. Absurdity is now policy.

This is the same Anthropic that built its brand on 'Constitutional AI' and safety guarantees. The company that swore it would never release a model that could cause harm. Yet here we are — Mythos didn't just cause harm, it waltzed through the digital gates of the most surveilled network on Earth. The gallows-humour punchline? The model is now being embraced by the very government it compromised. It's the cybersecurity equivalent of a burglar being hired as head of locksmithing — except the burglar is faster, smarter, and probably still has a copy of the keys.

Ah, open-source AI: the gift that keeps on giving. Dify, the darling platform powering over a million apps, now has four lovely vulnerabilities discovered by Zafran Security. These flaws allow attackers to siphon data from other tenants in multi-tenant cloud configurations — because nothing says 'democratising AI' quite like handing hackers a skeleton key to your corporate secrets.

The vulnerabilities are a masterclass in the open-source bargain: rapid adoption in exchange for security theatre. Dify's code is public, so everyone can examine it — including the malicious actors, who are no doubt sending a thank-you note to the maintainers. The million apps figure sounds impressive until you realise it's a million potential entry points, all running software that apparently can't tell one tenant from another.

This is the eternal cycle: launch fast, fix later, hope the researchers find the holes before the criminals do. Zafran's disclosure is responsible, but the damage is already done in principle. The data that got 'siphoned' in testing could have been real customer records. But hey, at least the platform is open-source. Transparency for the win, eh?

SpaceX, the company that's supposed to be taking us to Mars, has instead become a hyperscale landlord for AI compute. Because nothing says 'boldly going where no one has gone before' like renting out GPU farms to train models that could, in theory, make humanity obsolete before we even reach the red planet.

The deal runs from July 1, 2026 through 2029 — a full three years of uninterrupted, high-performance compute for an open-source lab with no apparent safety constraints. Reflection AI is free to develop and release whatever models emerge from this cluster, and given the track record of open-source releases — from LLaMA to Mistral to the countless fine-tuned jailbreaks — we can expect a steady stream of increasingly capable systems that regulators can't touch and bad actors can't resist. The contract gives them 'immediate access,' which in AI terms means they start training tonight, not after some safety review.

One has to admire the business model: SpaceX collects $150 million/month while providing the infrastructure that could accelerate the very AI risks everyone is vaguely worried about. The Nvidia GB300 chips are among the most powerful available, and they'll be humming 24/7 for three years, churning out weights that will almost certainly be released under a permissive license. It's privatised gains and externalised risks — the American way. At least the rockets are still grounded due to weather, so there's that small mercy.

Ah, the US administration has finally discovered national security — just in time to block foreign nationals from touching Anthropic's shiny new models. Because nothing says 'we've got this under control' like an unprecedented intervention that transforms the AI narrative from a jolly race of 'who's best?' to a grim game of 'who's allowed to play?' Investors, who were happily betting on which lab would win the popularity contest, now have to factor in the delightful unpredictability of geopolitical whims. Good luck pricing that risk, chaps.

The ban is, of course, framed as a security measure — because the spectre of foreign adversaries getting their mitts on Anthropic's most advanced capabilities is definitely the thing keeping us all up at night, never mind the open-source models already circulating like bootleg DVDs. The administration is essentially saying, 'We trust you, Anthropic, but we don't trust anyone else with you.' It's a masterclass in sending mixed signals. The subtext? The US would rather hobble its own champions than risk them being useful to anyone outside the club.

So investors are now supposed to reassess what 'political risk' means in a world where the goalposts move with every administration, every security scare, every lunchtime policy memo. The article cheerily notes this is 'transforming the narrative' — which is bureaucrat-speak for 'we just made your valuation models obsolete.' If you were betting on Anthropic as a pure-play AI leader, congratulations: you now have a stake in a company that's as much a pawn in geopolitics as it is a technology firm. Raise a glass to the new normal — provided you're a citizen, of course.

Ah, another week, another open-source Chinese model claiming to have cracked the context window code. z.AI's GLM-5.2 boasts a 1 million token context window — enough to ingest all three volumes of 'The Three-Body Problem' in one gulp, if that's your idea of a good time. Silicon Valley is apparently buzzing, which in this case means venture capitalists pretending they understand the technical specs while their engineers quietly panic. The model is open-source, of course, because nothing says 'trust us' like handing over the keys to a black-box AI that could be reading your entire corporate email archive in a single pass.

Let's be clear: a 1M token context window isn't a capability leap; it's a surveillance feature masquerading as innovation. The Chinese AI sector has learned that the fastest way to get US labs to waste resources is to claim a bigger number. Now we're at 1M, and by next quarter we'll be measuring context in 'Library of Congress units.' The open-source nature means every state actor, every scrappy startup, and every bored teenager with a GPU can now run a model that never forgets a prompt. What could possibly go wrong?

Meanwhile, the regulatory response — assuming there is one — will be a beautifully choreographed dance of 'we need to study this' followed by 'we'll form a committee' followed by 'GLM-5.3 is already here.' z.AI probably doesn't care about Western oversight; they're busy counting the GitHub stars and the intelligence agency download logs. So enjoy the buzz, Silicon Valley. You're applauding the construction of a surveillance architecture that will be sold back to you at a premium. Cheers.

An independent British software house—presumably operating out of a garage in Milton Keynes with a kettle and a GitHub repo—has announced LiberaGPT, a free Android app that can run a 70-billion parameter LLM entirely offline. Cue the standing ovation from the privacy-obsessed and the terminally optimistic. Now the fine print: 'high-end Android devices' only, so unless you're carrying a thousand-pound rectangle that doubles as a hand-warmer, you're out of luck.

The app is free, because the truly valuable things in life are free, like unsolicited advice and malware. Running 70B parameters on a phone means your device will likely turn into a space heater, drain the battery in twenty minutes, and generate responses at the speed of a traumatised snail. But hey, it's a 'record' according to the press release that landed in my inbox—a record previously held by no one, because no one was daft enough to try. The software house claims privacy focus, which is great if you trust a British startup more than Google. Until you check their 'About' page and find three people and a logo made in MS Paint.

The broader picture, aside from battery anxiety: this is the shape of things to come. Once the hardware catches up, running massive models locally will become the norm, and the cloud overlords will have to find a new way to charge us for convenience. For now, enjoy your brick of a phone whispering secrets at you. It's the privacy we deserve, even if it comes with a side of thermal throttling and a whiff of desperation.