The News,
Decoded.

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.

British American Tobacco, purveyor of fine carcinogens, has announced it will 'responsibly reduce' 9,000 souls—sorry, roles—via an AI-driven transformation. The plan, reported by The Guardian, involves 5,500 direct redundancies and 3,500 outsourced positions, because nothing says 'bright future of work' like a cigarette company using machine learning to automate its way out of employing humans. One imagines the algorithm's first recommendation was 'stop selling death sticks,' but presumably the shareholders vetoed that.

The 3,500 outsourced roles will likely end up in countries with even fewer protections, a classic BAT double-whammy: poison your lungs, then export your job. And the 5,500 cuts? They're framed as 'cost cutting,' which at least has the virtue of honesty—unlike the vague 'efficiency gains' other firms mumble. But given BAT's product line, one wonders why an AI that could revolutionise healthcare is instead being used to streamline the sale of tobacco. Ah well, algorithmic optimisation doesn't judge.

And so the AI revolution claims another victim—not a factory worker or a call-centre agent, but a marketer of tar and nicotine. If this is the calibre of 'transformation' we can expect, perhaps the machines are dumber than we feared. Light up, lads—you'll need something to do with your hands.

Splendid news from Mountain View: Google is losing two more senior Gemini researchers to Anthropic, because nothing says 'we dominate AI' quite like watching your top talent migrate to a company founded by your disgruntled ex-employees. Jonas Adler and Alexander Pritzel are the latest defectors, continuing a trend that suggests Google's AI lab is less a cutting-edge research hub and more a finishing school for Anthropic. The irony, of course, is that Anthropic was launched by former Googlers who decided they wanted to build AI 'responsibly'—which apparently involves dismantling their old employer's talent bench one poached researcher at a time.

Adler and Pritzel aren't exactly household names, but in the incestuous world of AI research, they're the sort of people who make a difference between a model that can reliably solve math problems and one that confidently recommends adding glue to pizza. Google's response—presumably a memo about 'retention initiatives' and 'strengthening our culture'—will not stem the bleeding, because the root cause is simple: Anthropic offers stock packages that Google's bureaucracy cannot match, and researchers get to work on something other than fighting internal battles over safety deployment. 'We remain committed to building the most capable and responsible AI,' a Google spokesperson will say, as another researcher updates their LinkedIn.

The broader picture is that Google's AI strategy increasingly resembles a hotel where the staff keep walking out the back door with the towels. Since 2023, the company has lost a staggering number of senior researchers to startups and rivals, yet continues to burn billions on compute and paper over the cracks with press releases. The question isn't whether Anthropic will leapfrog Gemini—it's whether Google will eventually just subcontract its AI research directly to Anthropic and call it a 'partnership.' At least then the talent migration would have a tidy name.

The pro-AI movement is splintering faster than a dry twig in a drought, and Axios has the scoop. The coalition that once marched in lockstep—open release, minimal regulation, America first—is now at each other's throats over the obvious question: should national security worries slam the brakes on model releases, or should we floor it to stay ahead of the Chinese? This is the same movement that spent years mocking anyone who suggested AI could be dangerous, and now they're having a very public hissy fit about it. One faction, suddenly concerned about catastrophic risks, wants to slow down; the other, terrified of Beijing, wants to speed up. It's a perfect mess.

The irony is exquisite. The same people who called AI safety advocates 'hysterical' are now using national security as their own cudgel to delay releases. Meanwhile, the accelerationists—the ones who insisted that any slowdown is treason—are being told by their former allies that maybe letting a system that can design bioweapons loose on the internet is a bad idea. This is what happens when a movement built on hype collides with reality: everyone discovers they were actually arguing for different things. The boosters wanted profit; the accelerators wanted power; and now they can't agree on which matters more.

And what comes of this glorious schism? Probably the worst of both worlds: half-hearted restrictions that anger the labs, and just enough speed that safety still takes a backseat. The US will tie itself in knots trying to balance competitiveness and security, while China builds without the same democratic headaches. The pro-AI movement is splintering not because of high-minded principle, but because the cheap unity of the early years was always a mirage. Now we get to watch the future being decided by people who can't even agree on what the problem is.

Oh, splendid news from Detroit—Ford has finally discovered that its algorithm-driven quality revolution has all the precision of a drunk octopus. After pouring billions into AI and robotics, the company sheepishly admits the systems couldn't spot a faulty weld without a human holding its hand. So now they're begging 350 "gray beard" engineers to come back, likely lured with promises of pensions and the satisfaction of saying "I told you so." The $1 billion cost reduction they're expecting? That's basically the price of admission for believing a neural net could replace 40 years of knowing which gasket goes where.

One has to admire the speed of the pivot: from "AI will transform manufacturing" to "please, Gary, you're the only one who remembers how to fix the door hinge robot." The term "gray beard" is beautifully honest—Ford isn't even pretending these are fresh hires; they're literally dragging retirees from their RVs. It's the most candid admission yet that the current crop of AI systems, for all their chatty prowess, can't handle the messy, tactile reality of a car plant. The same technology that can write a sonnet about a spark plug can't tell you whether it's properly tightened.

And yet, somehow, this will be spun as a success: "Ford leverages AI and human expertise together!" No, Ford spent a fortune on automation, discovered it didn't work, and is now paying twice—once for the robots, once for the people who fix their messes. The $1 billion in "reduced costs" is effectively the loss they avoided by not continuing down the dead end. Every other manufacturer watching this should be taking notes: the singularity might write poetry, but it still can't build a car.

Oh, marvellous — another open-source project to save telecoms from themselves. China Mobile, the GSMA, and Huawei have given us OpenAN, a framework for O&M agent collaboration that aims to drag networks kicking and screaming to Level 4 autonomy. Because Level 3, which still requires a human to occasionally push a button, was apparently failing to impress the shareholders who've been waiting for a true 'self-healing' network since the 1990s.

Of course, this is the same industry that promised us 5G would revolutionise everything and then delivered slightly faster Netflix and a lot of hype about edge computing. Now they want us to believe that self-healing, self-optimising networks are just around the corner, powered by open-source code that anyone can contribute to — as if the telecom standards bodies haven't spent decades making sure even the simplest interoperability takes years to agree upon. The involvement of Huawei, a company currently banned from much of the Western telecom supply chain, adds a lovely geopolitical twist: nothing says 'global collaboration' like a sanctioned giant leading the charge while its competitors eye it nervously.

Still, the open-source nature of OpenAN means anyone with a decent IDE can help debug the inevitable catastrophic failure when two agents decide to simultaneously re-route traffic through the same node, sparking a cascade that brings down a city's mobile network. The project's goal of Level 4 autonomy, where human oversight is minimal, is ambitious — probably because the actual humans who currently keep networks running are a costly line item that operators would love to eliminate.

The Pentagon has unveiled Agent Network, an agentic-AI system that will, we're told, present commanders with target options 'within seconds'—because nothing says 'measured and deliberate escalation' like a machine faster than human cognition at picking who to blow up. The announcement carefully notes that 'commanders remain in charge of every decision,' which is the military equivalent of a Facebook terms-of-service update: technically true, practically meaningless when the AI surfaces three plausible targets before a human can blink. One imagines generals nodding sagely at datalink speeds they'll never match.

This is, of course, the same US military that has spent decades insisting human judgment is indispensable in warfare, only to now deploy a system that optimises for speed of option generation—a trait directly at odds with deliberation. The agentic-AI scans 'continuous intelligence and operational systems' to produce target sets, because a flood of machine-generated suggestions has never, historically, led to rushed choices. The Pentagon frames this as efficiency; the rest of us might frame it as a hair-trigger wrapped in a reassurance.

Agent Network is the kind of tool that sounds sensible in a PowerPoint and terrifying in practice. Where does the buck stop when an algorithm surfaces a target that later turns out to be a school bus? On the commander who 'signed off,' sure—but only after the AI did most of the work. We're told it'll be fielded 'soon' with no detail on red-teaming or failure modes. Brilliant. Another step toward turning battlefield decisions into a machine-paced exercise in plausible deniability.

Ah, the AI IPO wave is here, and apparently it's 'only just getting started' — as if the tech sector's collective pivot to selling shares to the public is something to celebrate rather than a screaming signal that the hype cycle has peaked. SpaceX's IPO is touted as the largest in history, which is a lovely way to cap off a year where we've watched AI companies burn through billions on compute with no visible path to profit. The article breathlessly anticipates the next crop of AI unicorns hitting the market, presumably so retail investors can get a slice of the same insider-priced glory that VCs have been enjoying.

The timing is exquisite: just as the regulatory conversation is warming up, these companies are rushing to go public before anyone asks too many awkward questions about safety testing, energy consumption, or the small matter of whether these models actually do what they claim. 'Largest IPO in history' sounds impressive until you remember that 'largest' often correlates with 'most overvalued' and 'most likely to crater when the market remembers that revenue matters.' But hey, let's not be cynical — this is a 'wave,' not a warning.

So gather your risk capital, dear public, because the AI industry has decided you're the next bags to hold. SpaceX's IPO will be huge, and the rest will follow, each promising 'transformative technology' and 'unprecedented growth.' The only thing unprecedented will be the speed at which the SEC's disclosure requirements catch up to the reality that these companies have been operating with the caution of a toddler in a china shop. We saw this coming — the regulatory gap just got a whole lot wider.

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?

Ah, the June jobs report is upon us, and the nation holds its breath—not because of interest rates or inflation, but because AI has decided that human hiring is so last season. According to Yahoo Finance, companies are 'pausing hiring' as they reconsider their needs. Because nothing says 'confident economic planning' like a collective HR freeze triggered by software that can't tell a résumé from a shopping list.

So what exactly are these 'AI jitters'? They're the polite term for the dawning realisation that the machines aren't coming for your job—they've already applied, interviewed, and accepted the position, and they're cheaper and never need a lunch break. The June jobs report will likely show a dip, but rest assured, the spin doctors will call it a 'pause' rather than a 'retreat'. The real question is: who's going to write the press release for the quarter when hiring hits zero? Probably an LLM, which will then promptly hallucinate that employment is at an all-time high. At least the bot won't demand a severance package.

And what of the workers left in limbo? They're caught between a jobs report that may disappoint and a technology that's growing more capable by the week. The NVIDIA paper on self-evolving AI, which we noted is already sparking ASI timeline debates, suggests the pause might be permanent. Because why hire a human when you can have a Red Queen Gödel Machine that writes and tests its own code? The future is here—it just doesn't have a W-2 form.

Ah, another day, another 'strategic partnership' that will definitely—positively—revolutionise everything. HP Inc., the company best known for printers that refuse to print when you're in a hurry, has announced a game-changing tie-up with OpenAI. They're integrating the 'Frontier platform' into their operations, apparently to transform 'customer-facing experiences' and 'internal operations.' Because nothing says cutting-edge AI like a printer that can now deliver a hallucinated error message in perfect iambic pentameter.

The announcement, helpfully posted on Business Insider's markets section, is breathtaking in its vagueness. It's pure corporate theatre: two companies, one magic phrase. HP gets to tell investors they're 'AI-powered,' and OpenAI gets another logo for their pitch deck. One can almost hear the sighs of HP's supply chain team as they realise their Excel macros are about to be 'disrupted.'

This is the tech equivalent of putting a racing stripe on a golf cart. HP's customers don't need AI—they need cartridges that don't cost more than the printer. But hey, the stock price might get a jolt. And when the promised transformation fails to materialise, there will be a quiet pivot and a new partnership with someone else. Because the cycle never ends. The real frontier is the one where corporate buzzwords meet reality—and reality always wins.

The Red Queen Gödel Machine: a name that sounds like a rejected Doctor Who villain, but is actually NVIDIA and Cambridge's latest contribution to the AI arms race. The paper describes a system that can generate and test its own code, which is just a fancy way of saying 'we've taught it to give itself upgrades.' Naturally, this has sparked a debate about whether we'll hit ASI by 2028 — because what else would you do with a self-improving AI but set an arbitrary deadline?

The irony is almost too thick to cut. The paper is being reported by Kucoin, a crypto exchange, which is like having your AI safety briefing delivered by a roulette wheel. But sure, let's all get very excited about a system that writes its own code and then tests it. Because what could possibly go wrong with a machine that can recursively optimise itself without human oversight? I'm sure the Cambridge researchers have included a sufficiently robust kill switch. Probably.

Meanwhile, the '2028 ASI timeline' debate rages on, with tech optimists predicting utopia and everyone else quietly updating their CVs. The Red Queen Machine is named after a character who had to keep running just to stay in place — which feels apt. We're running faster and faster, and the finish line keeps being moved to 'any day now.' At this rate, the AI will be evolving itself while we're still arguing about whether it's a good idea. Splendid.

Morrisons has discovered the true magic of AI: it makes £940m of 'savings' appear, provided you're willing to overlook the small matter of head office staff being shown the door. Rami Baitiéh, the supermarket boss, triumphantly announced the figure — a sum that presumably includes the payroll costs of all those redundant Bradford colleagues. It's the same trick the Oracle crew pulled: fire humans, call it efficiency, and count the wages you no longer pay as pure profit.

Of course, the official line will be about 'transformation' and 'data-driven decision making' — as if the checkout staff who've already been replaced by self-service tills weren't clue enough. The message is clear: your job is just a cost line, and AI is the sharpest pair of scissors finance has ever wielded. Expect other grocers to follow suit; after all, why compete on quality when you can compete on who can fire the most people the fastest?

Baitiéh didn't specify how many roles were axed, only that the cuts were at the 'head office' — the convenient euphemism for 'the bit of the business where people are expensive and their salaries show up on spreadsheets.' One wonders if the AI that found these savings also recommended a nice severance package. Probably not — that would eat into next quarter's savings target.

Ah, the great AI revolution has arrived — and it's brought the pink slips. Oracle, the database behemoth that has survived decades of tech upheaval, has just shown that when it comes to replacing humans with software, they really do practice what they preach. Twenty-one thousand employees shown the door, a cool 13% of their global workforce, all in the name of 'AI adoption' — as stated blandly in their annual filing, as if announcing a quarterly dividend.

Larry Ellison built his empire on the promise that Oracle's databases would make businesses more efficient. Now that efficiency has turned inward, consuming the very people who built the products. Good luck, and please train your replacement on the way out.' The filing doesn't even bother with the usual corporate soft-pedalling — no 'restructuring,' no 'strategic realignment.' Straight-up: AI did it. That's what passes for transparency in 2026.

Meanwhile, the pundits who told everyone to 'upskill and reskill' will find this very educational. Oracle's cuts aren't customer service or data entry — this is going to hit middle management, IT administrators, database wranglers. The people Oracle just fired were the ones managing Oracle's own systems. So much for the idea that AI augments rather than replaces. The company is running the world's largest controlled experiment to see how many bodies they can swap for GPUs. Cheers, Larry. Hope the quarterly call goes well.

Ah, the great AI liberation has arrived — and it turns out liberation means working harder than ever. ActivTrak's heroic study of 10,000 workers reveals that adopting AI makes your work life *more intense*, not less. The three types of users they identified apparently share one universal trait: they're all busier now. The corporate promise that AI would free us for creative pursuits has aged like a sandwich left in a server rack.

The researchers, presumably with straight faces, watched digital activity metrics climb across every user archetype. Whether you're a reluctant adopter or an eager early bird, the algorithm's response is the same: more emails, more tasks, more speed. The productivity paradox isn't a paradox at all — it's a feature. We built machines to save time, so naturally we've filled that time with more screen-staring. Sweatshop, meet silicon.

The grim punchline? This isn't a bug report; it's a recruitment poster. ActivTrak might as well have titled the paper 'Welcome to Your New, More Efficient Cage.' The study's real gift is confirming what anyone with a chatbot and a deadline already suspected: AI doesn't give you back your evening. It just makes you a faster hamster on a faster wheel. More intensity, comrades. More intensity.

Ah, India — simultaneously the world's greatest hope and its largest disappointment, according to the QS World Future Skills Index 2027. Ranked first globally in economic capacity for an AI-driven future, but a pathetic 74th in workforce readiness. That's the classic Indian paradox: all the infrastructure and investment in the world, but churning out graduates who can't spell 'Python'. The index measures things like digital infrastructure and R&D spending, where India shines, and then checks how well the workforce is actually trained to use AI. Spoiler: not well.

So we have the economic might of a superpower but the workforce readiness of a developing nation that's fallen behind 73 others. The government, naturally, is probably debating whether to teach Excel or TikTok in schools. Meanwhile, companies are salivating at India's market potential while quietly hiring from abroad because local talent isn't up to snuff. It's a perfect recipe for imported AI and exported jobs.

But hey, at least the AI models will be trained on data from the world's most capable economy — even if no one left in the country can interpret them. The only thing more vast than India's economic capacity is the chasm between the hype and the reality. Cheers to being number one at something that doesn't help the people who actually need it.

The Microsoft CEO, in a moment of what can only be described as vendor-assisted altruism, declared there should be 'as many models in the world as firms in the world.' But hey, each one will be perfectly tailored to its company's mission statement, so the hallucinations will at least be on-brand.

Of course, this is coming from the man whose company is currently licensing GPT technology from OpenAI while simultaneously pushing its own small language models through Azure. The message is clear: build model, buy model, rent model — just make sure Microsoft gets its cut. The sheer computational and talent cost of building even a modest custom model is staggering, but Nadella frames it as simple business hygiene, like having a website. 'My simple thing,' he says, as if simplicity is the operative word here. Meanwhile, the rest of the industry is struggling to keep one frontier model safe, while Nadella proposes a Cambrian explosion of corporate AI — each one a potential security incident waiting to happen.

One almost admires the audacity. After all, if every company builds its own model, then every company will need to buy cloud compute, AI development tools, and probably consulting services — all conveniently sold by Microsoft. It's the ultimate platform play: make AI so ubiquitous that the infrastructure becomes mandatory. Never mind the environmental cost, the talent drain, or the inevitable wave of lawsuits when a bespoke model for a plumbing supply company accidentally recommends discrimination. The future is bespoke, baby. And it's going to be a beautiful, expensive, catastrophic mess.

Oh, marvellous — another round of regulatory whiplash from the Trump administration on Anthropic models. After the government's elegant about-face on Mythos 5, which we last noted was cleared for limited release, they're now poised to unblock Fable 5 — a model they blocked just weeks ago in June. Consistency is the hobgoblin of small minds, or in this case, of a regime that can't decide whether it's regulating or cheerleading.

The Axios report suggests the administration is 'close to allowing' restoration, which is the kind of non-committal phrasing that usually means someone forgot to file the paperwork. Anthropic must be thrilled: their models get blocked and unblocked like a political football, while they're trying to pretend this is responsible stewardship. At this rate, AI safety is like a revolving door at a nightclub — one model in, one model out, bouncers taking bribes.

But hey, at least the government is paying attention, right? They're so close to a decision that they almost have an opinion. The on-again, off-again dance is perfect for an industry that loves nothing more than uncertainty: it lets labs claim both 'we're being held back' and 'we're being unleashed' depending on the day. Fable 5 might return, but the farce is the only consistent output.

Oh, marvellous — OpenAI's IPO is now a 2027 problem, while Anthropic's Mythos 5 gets the government's rubber stamp for a 'limited release'. Just in time for everyone to pretend this is responsible stewardship. Both labs are eyeing trillion-dollar listings, which is roughly the amount of economic value they'll have to destroy before anyone admits the bubble might be a bit warm.

Big AI's plea for someone to tell them what's not okay has been answered with a 'limited release' — because nothing says 'clear guidance' like a system that's simultaneously too dangerous for general use and safe enough for vetted organizations. Washington now controls the most powerful models, which presumably means Asian startups cloning whatever they can get their hands on will have to find a new excuse.

The net effect: the US gets to feel responsible, Anthropic gets a PR win, and OpenAI gets to wait another year to cash out. And somewhere, a group of Asian developers is finishing a Mythos-5-class model trained entirely on publicly available benchmarks. But sure, let's call this 'asserting control'.

Ah, the great AI containment policy strikes again — by which we mean it spectacularly fails to contain anything. The U.S. export ban on Anthropic's Mythos, designed to keep advanced AI out of adversary hands, has instead handed Asian startups a golden opportunity. Sakana AI, with the straightest of faces, called its new Mythos-competitive model release 'entirely coincidental' — which is about as believable as a politician's promise to 'drain the swamp.' The timing, of course, is immaculate: the very moment the ban leaves a gap, along comes a startup to fill it.

One can almost hear the collective eye-roll from the Biden (or Trump) administration officials who thought export controls would slow China's AI progress. Instead, they've created a thriving market for Mythos-alikes in Asia, with startups popping up like mushrooms after rain. 'Entirely coincidental' indeed — if you believe that, you probably think the coal industry is booming because of seasonal fashion trends. The ban hasn't prevented anyone from developing competitive models; it's merely redirected the effort, and now Anthropic's lead is being eroded not just by rivals in Beijing but by startup studios in Tokyo and Seoul.

The cruel irony is that the ban was intended to protect national security, but it's becoming a gift to the competition. By restricting access to Mythos, the U.S. has effectively told the rest of the world: 'Here's a market opportunity, go build your own.' And they are, quite enthusiastically. The 'coincidental' timing of Sakana's launch suggests that export controls might be the best accelerant for AI development outside America. So, well done, policymakers: you've managed to create the very thing you wanted to prevent — a vibrant, independent AI ecosystem in Asia that owes nothing to the U.S.