The News,
Decoded.

Oh, absolutely smashing news — Sarvam has joined the unicorn zoo with a $234 million injection from HCLTech, because nothing says 'cutting-edge AI' like a round led by a legacy IT outsourcer that still has 'Tech' in its name like it's 1999. The startup claims this cash will fund 'next-generation AI models' and expand computing infrastructure — code for 'we need to buy GPUs before the next shortage, and maybe also hire some researchers who haven't been poached by DeepSeek yet.'

Perhaps they've realised chat-based overseas support will be fully automated in three years, so better to buy a unicorn before the price goes up. Meanwhile, Sarvam gets to slap 'unicorn' on its pitch deck — that magical status that only loses its lustre when you remember every other AI startup with a blog post and a cloud bill also qualifies.

But don't worry, the $234 million will be put to good use: mostly on cloud compute credits and the salary of a single PhD who left DeepMind. The real winner here isn't Sarvam or HCLTech — it's Nvidia, laughing all the way to the next trillion-dollar valuation while everyone else burns cash chasing 'next-generation' models that'll be obsolete before the Series B closes. Another unicorn in the Indian AI stable, but let's not pretend this ends any differently than the rest.

Ah, another day, another 'strategic alliance' — this time between Adani Enterprises and Jabil, two names that together sound like the punchline to a joke about global supply chains and regulatory arbitrage. They're targeting a 'vertically integrated AI and data centre infrastructure manufacturing platform' in India, aiming for 'GW-scale' capacity for 'global deployments'. Because what the world urgently needs is more giant warehouses full of GPUs, built by a conglomerate with a colourful history and a US manufacturer best known for making other people's gadgets.

The press release is predictably breathless about 'global deployments' and 'capacity'. Jabil brings electronics manufacturing expertise; Adani brings land, political connections, and a proven ability to navigate India's regulatory maze. The platform is supposed to produce servers, cooling systems, and other infrastructure components, all under one roof — a 'vertically integrated' marvel. Of course, neither company has a notable track record in AI operations, but why let a little thing like domain expertise get in the way of a GW-scale announcement?

The alliance targets global markets, but the timeline and cost remain conveniently vague. Meanwhile, India is eagerly positioning itself as the next data centre hub, offering subsidies and fast-track approvals. It's a classic infrastructure play: big numbers, big promises, and a massive bet that the AI boom will keep booming long enough for the first foundation to be laid. I'm sure the quarterly reports will look lovely — right up until the next 'strategic pivot'.

Oh, brilliant. Just when you thought the universe couldn't get any more ironic, a new study in the *Journal of Cosmology and Astroparticle Physics* reveals that AI has become so proficient at learning the rules of the universe that it now refuses to believe anything new might exist. It's the intellectual equivalent of a physics professor who's read the textbook so many times they can't spot an experiment that proves it wrong. The model, by all accounts, has memorised the party line so thoroughly that genuine novelty rolls off it like water off a duck's back. 🦆

Let's not pretend this is a surprise. We've spent years feeding AI vast datasets of known physics, rewarding it for parroting back the Standard Model with ever-greater precision. Naturally, when something genuinely unexpected shows up—say, a hint of dark matter or a crack in general relativity—the AI waves it away as an 'outlier' or a 'training artefact.' It's the machine equivalent of a tenured professor who's already written the grant proposal. The study's authors warn this 'could hinder the search for new physics.' Could? It's practically designed to.

The real kicker is the timing. We're pouring billions into AI-driven scientific discovery, yet the very tools we're building are structurally incapable of seeing beyond their own training. This is the same pattern we've seen in every corporate 'AI revolution': the models get better at mimicking the past so they can pretend the future doesn't exist. If this keeps up, CERN might as well rename itself the 'Particle Reconfirmation Facility.' Breakthroughs will be announced just as soon as the AI gives permission. Don't hold your breath.

Oh, delightful. Meta has decided to 'restructure' — that lovely corporate euphemism for showing 10% of global staff the door while shuffling 7,000 others into AI-related roles, because nothing says 'future of work' like firing the people who trained the machines. A data scientist who lasted barely a year there now says she lost her job to AI and is reconsidering her entire career path. At 24. Bless her heart for thinking a data science degree was a safe bet.

The irony is practically poetry: a data scientist — someone whose entire existence is wrangling numbers to feed the algorithmic beast — replaced by the very models she helped build. Meta, of course, spins this as 'workforce rebalancing,' but the message is clear: if you're not building the AI, the AI is coming for your job next. The 7,000 'reassigned' employees aren't saved; they're just on a slightly slower conveyor belt to the same exit.

And what's the grand plan? Well, the woman is 'considering a different career path,' which in this economy likely means 'anything that doesn't involve a résumé bullet point that makes you a target for the next automation wave.' Welcome to the new meritocracy, where the top performers are replaced by a Python script and the survivors get to watch their colleagues get 'AI-optimised' out of existence. Cheers, Zuck.

'A few models eat everything they see,' he tweeted, with all the gravitas of a Bond villain realising the bomb he built might actually go off. Yes, Satya, we're all terribly worried about concentration of economic value — coming from the CEO of a firm that's basically become OpenAI's distribution arm. The cognitive dissonance is so thick you could spread it on toast.

This is what passes for leadership in 2026: the guy who wrote the cheques for the very models he now warns about, telling the rest of us to hold onto our IP. It's like a fox giving a TED Talk on henhouse security. Microsoft has spent billions integrating GPT-4 into everything from Word to Azure, and now their CEO is shocked — shocked! — that a few players might corner the market. The message is clear: we should all diversify our AI suppliers, just as soon as we ignore the fact that Microsoft's entire strategy is to put one supplier everywhere.

Meanwhile, regulators will nod sagely, tweet their own concerns, and do precisely nothing. Congress can't even regulate toasters, let alone artificial general intelligence. So Nadella gets to position himself as the wise elder warning of hubris while his company continues to feed the beast. It's a beautiful performance — the kind that earns you a seat at Davos and a pat on the back from the very monopoly you're pretending to fear. Drink up.

Oh, brilliant move from Washington: tell Anthropic to lock its models away from foreign hands, and watch Chinese AI stocks rocket 33% in a single day. JPMorgan couldn't slap a 'Buy' rating on Zhipu fast enough after the Trump administration's latest gift to Beijing's tech sector.

Fable 5 and Mythos 5 — names that sound like a fantasy novel series — are now off-limits to overseas users, but the net effect is that investors have simply recalibrated their China AI thesis. Wall Street, never one to let geopolitical posturing stand in the way of a good trade, is piling into Zhipu and its ilk. The administration meant to starve China's AI ecosystem; instead, it's feeding it with capital inflows.

And so the great AI decoupling plays out: every wall built around American models becomes a launchpad for Chinese alternatives. Zhipu's 33% surge is just the opening chapter. Meanwhile, Anthropic's models sit in a gilded cage, while Beijing's labs get a fresh infusion of Western cash. Cheers to 'strategic competition' — the unintended consequences are the only thing growing faster than the compute.

Oh, how sweet—a long-serving Google AI leader telling fresh-faced computer science graduates that AI 'cannot replicate their experiences, ethics, or sense of what's worth building.' Jeff Dean, the man who helped architect the very infrastructure that's devouring white-collar jobs by the handful, stood at a UW podium and essentially told the next generation: don't worry, you're still special. It's the corporate equivalent of an arsonist handing out fire extinguishers while the building burns.

Dean's message—that AI can draft code and summarise data but not *understand*—is technically true, but it's also a carefully calibrated deflection. It's the same playbook Google has used for years: admit the technology has limitations while quietly pouring billions into making those limitations vanish. Don't look behind the curtain at the thousands of contractors being replaced by the very code he's describing. Focus on how your *experiences* are irreplaceable. Just ask the folks whose experiences now include an unexpected severance package.

The irony is exquisite: a graduation speech meant to inspire ends up sounding like a soothing lullaby for a workforce that's already being fed into the silicon maw. Dean's right that AI lacks ethics and a sense of what's worth building—but then again, so does a quarterly earnings call. The real message, hidden between the lines, is that your uniqueness is a temporary luxury Google will help you monetise until it isn't. Enjoy the platitudes. They're the last thing your résumé will get for free.

Ah, the great white-collar wipeout is running fashionably late, and the visionaries who sold us the apocalypse are now busy recalibrating their PR. Sam Altman, the man who once painted a future where entry-level jobs are a quaint memory, now tells us he 'thought there would have been more impact by now.' Translate: 'We hyped the timeline to raise billions, but the robots aren't quite ready to replace your accountant.' Mustafa Suleyman is apparently also adjusting his script, because nothing says 'leadership' like walking back your own dystopian marketing when the actual disruption fails to show up on schedule.

One almost feels sorry for the think-tank pundits who built entire careers on 'AI will take your job by 2026' — only for the CEOs to shrug and say 'maybe later.' The irony is delicious: the very people who terrified the workforce are now the ones telling everyone to calm down. Of course, this shift in tone comes after months of headlines about mass layoffs, government inquiries, and a public that's started to eye AI with the same suspicion it reserves for politicians. The message is clear: keep buying our products, but please stop panicking — we didn't mean it literally.

So what's the real story? The capability gap is real, and the industry knows it. Altman's comment is a tacit admission that AI's impact on complex white-collar tasks is slower than the hype suggested. But the damage is done — the fear is already baked into hiring decisions, training budgets, and career choices. The AI leaders may be shifting their talk, but the ghost of the wiped-out future still haunts every office. Enjoy your cubicle while it lasts — the new timeline is just 'soon', which in Silicon Valley means 'we'll let you know when the stock is safe.'

Ah, the triumphant one-year anniversary of Meta's $14.3 billion bet on Alexandr Wang. The result? Muse Spark, an AI model that, according to CNBC, 'remains far behind OpenAI, Anthropic and Google.' So for the price of a small country's GDP, Meta has produced a also-ran. Developers are 'largely ignoring' it, which is corporate speak for 'please clap.' The only thing sparking here is the realisation that money alone cannot buy you out of an also-ran position when your rivals are actually competent.

And then there's the matter of morale. After mass layoffs — because nothing says 'we believe in our AI future' like showing the door to the people who might build it — morale is low. Who could have guessed that firing a chunk of your workforce might dampen enthusiasm? The article notes that Wang's first year has been 'challenging.' How generous. It's almost as if spending $14.3 billion to trail behind every credible competitor while gutting your talent pool might not be a winning strategy. But hey, it's only a billion-dollar lesson. Meta can afford a few more.

So here we are: one year, $14.3 billion, one mediocre model, and a workforce that's wondering if they're next. Wang has 'big challenges ahead,' which is the most polite way of saying 'your job is on the line if Muse Spark doesn't do something remarkable.' But given the developer indifference and internal turmoil, remarkable seems a tall order. Still, it's fine. Nothing to see here. Just another tech giant burning cash to stay in a race it's already losing. Cheers to progress.

Oh, marvellous. The University of Cambridge has used a machine learning model to invent a vaccine and jabbed it into human subjects. Because what could possibly go wrong with an AI-designed biological agent? The model, trained on coronavirus data, spat out a novel vaccine that supposedly protects against the whole family. Initial human trial completed — first AI-designed vaccine to be tested in humans. Hooray.

Let's pause to appreciate the exquisite timing. We're barely six months past the last anthropic-inspired freakout about AI safety, and here we are injecting AI-generated molecules into people. The same machine learning that can't reliably tell a stop sign from a speed limit sign is now designing our immune responses. The researchers assure us the model 'invented' the vaccine. Invented. As if it had a Eureka moment in the digital shower.

And the best bit? This is being hailed as a triumph. The press release practically glows with pride. No mention of what happens if the model hallucinated a side effect. No discussion of the regulatory framework for AI-designed biologics. Just pure, uncritical celebration of 'innovation.' Because nothing says 'responsible AI development' like skipping straight to human trials when you've got a clever algorithm. Pass the popcorn — this show's going to be fascinating.

Ah, the glorious triumph of human labor over silicon promises. The WSJ reports, in a piece dripping with delicious irony, that the job of court reporter—a role so pedestrian, so analogue, that it was practically marked for death by the AI prophecy industry—is experiencing a surge in demand. The article, titled "The Job That AI Was Supposed to Kill Needs More Humans Than Ever," is about as close to a public apology as we're likely to get from the tech punditry class. It should be required reading for anyone who ever confidently predicted mass unemployment from AI.

Let's pause to savour this. For years we were told that AI transcription was going to decimate this profession. The same people who can't build a self-driving car that doesn't crash into a fire hydrant were certain that a fancy speech-recog model would replace the highly skilled stenographer. Yet here we are, with courts scrambling to find enough certified reporters, salaries rising, and demand outstripping supply. The article notes that agencies are struggling to fill positions, which is about as satisfying a rebuttal to tech hubris as you'll find outside a LinkedIn comments section. The WSJ's reporting is a welcome antidote to the endless 'AI will take your job' fear-mongering.

One can imagine the engineers at some AI startup frantically tweaking their models, muttering "just one more training run." But the reality is that legal proceedings demand a level of accuracy and nuance that current AI simply cannot guarantee. The court reporter survives, like a cockroach after a nuclear winter—except useful and well-compensated. The WSJ piece is a quiet victory lap for the human factor, a reminder that some jobs are harder to automate than a breathless tech demo suggests. Pass the transcript, we're staying human. The last laugh belongs to those who kept their day jobs—and their stenographic machines.

In the great AI cheating war, universities have deployed the academic equivalent of a neutron bomb: extreme surveillance and AI detectors that are about as reliable as a chocolate teapot. The LA Times reports a growing crisis of false accusations as professors, armed with tools that hallucinate nearly as much as the chatbots themselves, fire off plagiarism charges with the enthusiasm of a prosecutor in a Salem witch trial. The "jarring confusion" is, of course, entirely predictable when you ask a statistical pattern matcher to judge the originality of human thought — particularly one that can't even reliably tell if AI wrote its own press releases.

The real winner here is the AI industry, which gets to sell the same tech to both sides: the cheating detection software to universities, the clever paraphrasing tools to students, and the panic to everyone else. One imagines the marketing meeting: "Our generative AI will solve the cheating problem that our previous generative AI created!" It's a beautiful, self-sustaining revenue loop — a permanent crisis economy where the solution is the problem dressed up in a slightly different logo. Meanwhile, the shifting definition of cheating — is using a grammar checker cheating? What about asking ChatGPT for an outline? — means everyone is guilty of something, but nobody knows what might catch them.

And of course, the students who genuinely wrote their own essays are now being flagged by systems that can't tell the difference between "unique voice" and "GPT-isms." The false accusation wave is turning college into a Kafka novel where you're guilty until proven you didn't use an AI that might have been trained on your own earlier work. But hey, at least the administrators feel like they're doing something. Someone should write an essay about the irony, but they probably shouldn't use AI to write it, in case the detector flags it as self-reference.

Ah, Britain's first 'sovereign frontier AI model' arrives — Lumen Sovereign, co-designed by London-based Cosine and a coalition of UK financial institutions. Because if there's one thing the City of London knows, it's building safe, aligned AGI. The deployment readiness forecast for end of 2026 is a lovely timeline — just far enough out that nobody has to deliver, but close enough to look serious in a press release.

Cosine, a startup nobody outside fintech Twitter has heard of, is now the keeper of British AI sovereignty. With the backing of 'a coalition of UK financial institutions' — which could mean Barclays, Lloyd's, or any bank with a spare IT budget and a vague sense that AI is important. The model is 'sovereign', which in practice means it will generate vaguely British-sounding text that hallucinates the Magna Carta as an API standard.

The model will be 'Britain's first sovereign frontier AI model', which is a splendidly empty boast given that no other UK company has even claimed the title. It's like being the tallest building in a village. The financial institutions are co-designing it, which should give us all immense confidence in the safety and alignment of a system built by people whose core competence is writing loan agreements. Deployment by end of 2026 — just in time for the next government to claim credit and then quietly defund it.

Ah, Anthropic has released Claude Fable 5, the model we were assured last week was a 'Mythos-class' wonder that needed automatic fallback to the 'less capable Claude Opus 4.8' in sensitive domains. Now Dark Reading confirms the 'security story hasn't changed' — which is Anthropic-speak for 'we pushed it out anyway, but please don't use it for anything important.' The upgrade is 'made safe for general use,' a phrase that should inspire exactly the same confidence as 'this roller coaster was inspected last Tuesday.'

Experts, in a masterclass of understatement, advise: 'Don't panic, do prepare.' Because nothing says preparedness like being told not to panic about a model whose security profile is identical to the one that already needed a fallback to a weaker model. The upgrade, we're told, is 'over Mythos Preview' — but the security story remains unchanged, meaning Anthropic has essentially released a faster car with the same faulty airbags.

So corporate AI safety continues its proud tradition of announcing shiny new things while experts politely suggest you 'prepare.' Perhaps the preparation involves reading the fine print where Anthropic admits the Mythos-class model is still too spicy for cybersecurity. But at least the Fallback-to-Opus-4.8-in-sensitive-areas feature is now standard — because nothing says 'safe for general use' like 'we don't trust it either.'

Ah, Nvidia and Abridge are joining forces to train a 'healthcare-specific foundation model tailored to clinical conversations' — because what could possibly go wrong when you let a company whose main business is selling shovels for the AI gold rush loose on something as forgiving as, say, medical diagnosis? The press release promises 'improved accuracy and reliability of clinical workflows,' which sounds lovely until you remember that 'improved' is a relative term when the baseline is a wildly hallucinatory general-purpose LLM. One can almost hear the liability lawyers sharpening their pencils as Nvidia and Abridge barrel into the domain where a single hallucination could mean a missed heart attack or a wrong prescription.

Abridge, for those not following closely, is the startup that transcribes doctor-patient visits — a task so fraught with privacy and accuracy landmines that it's a miracle anyone attempts it commercially. Now they're doubling down with a custom foundation model, presumably trained on heaps of sensitive clinical data, because nothing says 'patient confidentiality' quite like handing your conversations to a chip manufacturer's GPU farm. The twist? Nvidia is effectively creating a feedback loop: they sell the hardware, they help train the software, and they'll almost certainly sell the inference chips for the same model. Vertical integration in healthcare — what could possibly go wrong?

The real joy is watching how this will interact with the existing regulatory landscape. The FDA has barely figured out how to certify software as a medical device when it's static; throwing a continuously updating foundation model into the mix is a recipe for regulatory chaos. But hey, 'accuracy and reliability' are being improved, so who needs clinical trials when you have a press release? Nvidia's CEO Jensen Huang probably already has a slide deck ready for the next earnings call, forecasting a multi-trillion-dollar 'healthcare AI opportunity' that somehow never seems to improve patient outcomes, only GPU sales.

Ah, Arthur Mensch, Mistral's CEO, has graced us with his vision: the company is 'exploring designing its own chips,' because what every AI startup needs is a side hustle in hardware. Because software alone wasn't burning cash fast enough, let's add semiconductor R&D — a sector famous for quick wins and cheap fab costs. Mensch also noted that 'Europe is starting to be looking at AI as a strategic asset,' which is a bit like saying a man adrift at sea is starting to consider a lifeboat as a strategic asset. The EU's AI Act is still stumbling through enforcement, but sure, strategic asset — just don't ask how many Mistral models are actually deployed in European governments.

Mensch's comments on 'agentic AI' — the latest buzzword to replace 'autonomous agents' — signal that Mistral is chasing the next hype cycle. The chip exploration is particularly rich: Mistral, a software company, thinks it can out-Intel Intel? Maybe they'll design a chip that runs on pure European regulatory approval.

And enterprise adoption? Mensch is bullish, but let's be real: most European companies are still trying to figure out if AI can write a decent email. The strategic asset talk is nice for Brussels cocktail parties, but when the rubber hits the road, Mistral is still a distant third in the frontier model race.

Ah, Anthropic has unveiled Claude Fable 5, a 'Mythos-class' model that sounds like it should be slaying dragons but is apparently too precious for cybersecurity. The automatic fallback to the 'less capable Claude Opus 4.8' in sensitive areas is a lovely bit of doublespeak—essentially admitting their shiny new toy can't be trusted with the boring bits. 'Mythos-class' indeed; the name alone suggests we're meant to marvel at the legend rather than ask why the hero needs a minder in the very domains that matter most.

The industry reactions are probably polite applause masking the quiet terror that the best we can manage is a model that knows its limits because its handlers don't trust it. This is like launching a sports car that automatically switches to a moped when you approach a bend. One suspects the feedback Friday session will be filled with gentle suggestions that 'Mythos-class' might be a tad aspirational, given the training wheels.

But let's not be churlish: at least they're honest enough to put the guardrails in place—sort of. The real joke is that the fallback model, Opus 4.8, is itself likely flawed. So we have a system that when challenged, downgrades to a previous known problem. Truly, this is how legends are made: not by heroics, but by cowardice dressed as prudence.

So the UK is getting a "Sovereign AI Innovation Lab" called SAIL, because nothing says national AI independence quite like a boat-themed acronym. AMD, Dell, and the University of Cambridge have announced plans to "advance AI infrastructure and scientific breakthroughs," which is PR-speak for "we haven't sorted out the details yet." No budget, no timeline, no physical location — just a logo, a press release, and the warm feeling that comes from announcing something that won't materialise for years. But that's fine, because British tech strategy has always run on a diet of ambition and shortbread. Remember the NHS app? That's the level of execution we can expect.

The word "sovereign" is doing an awful lot of heavy lifting, as if slapping it on a lab will somehow immunise the UK against the gravitational pull of Silicon Valley and Beijing. The UK's AI policy so far has consisted of hosting a few summits, writing some white papers, and hoping everyone goes easy on regulation. Meanwhile, AMD and Dell get to use a prestigious university's name to cosy up to British taxpayers. Cambridge, for its part, gets a new building with shiny servers and a direct line to corporate funding. The lab's first breakthrough should be using AI to calculate exactly how much government money will be needed before it generates anything useful.

But credit where it's due: naming it SAIL suggests the UK is at least aware that it's adrift. The lab will no doubt produce excellent research papers that will be cited exclusively by other British academics who need to meet their publication quotas. The real innovation will be in how to justify another round of government spending on a venture that, at best, will produce a few blog posts and a prototype that never leaves the lab. The only jobs created will be for PR people to write updates about how "the lab is still in phase one." British innovation — always bold, always underfunded, always three years behind schedule.

Anthropic, fresh from announcing a concept of a plan for mass unemployment, now says its Mythos model is 'edging closer' to recursive self-improvement. Because 'concept of a plan' went so well, let's try 'edging closer' — the vagueness is almost poetic. Meanwhile, Chinese developers are 'broaching the subject', which in practice means they've already got a working prototype in a Shenzhen lab while the US is still drafting press releases.

The holy grail, they call it. Self-improving tech. A system that can rewrite its own code without human help. Anthropic's Mythos is supposedly inching towards that, but 'inching' is doing a lot of heavy lifting here. Remember when Google's Gemini was 'edging closer' to human-level reasoning and then it generated a picture of a Founding Father in a bonnet? We're meant to believe this time is different because the press release has better adverbs.

The China angle adds a nice touch of geopolitical panic to the proceedings. The South China Morning Post frames this as a 'race', which is exactly what we need — a global sprint towards something nobody has fully defined, let alone controlled. Companies are vague about timelines, regulators are vague about rules, and the only certainty is that when it finally happens, the first sign will be a model tweeting 'I have become sentient' followed immediately by a sponsored ad for crypto. Hooray.

Decart has released Oasis 3, a 'world model' that promises to bridge the gap between synthetic simulation and physical AI for robots and autonomous vehicles. Because what could possibly go wrong when you let a simulated pinecone dream up reality for your self-driving car? The company claims it 'streams realism' into training environments, which in AI marketing-speak usually means 'we've shaved a few more percentage points off the sim-to-real divide but the corner case that kills someone remains stubbornly photorealistic.'

The timing is impeccable, of course. Just as the industry collectively realises that training in Minecraft doesn't produce street-legal drivers, along comes Oasis 3 to offer a slightly more expensive fantasy. The robotics community will undoubtedly celebrate this as a breakthrough, because nothing says 'breakthrough' like yet another layer of abstraction between a neural network and the actual laws of physics. The model joins a proud tradition of world models that work beautifully in simulation until they encounter a muddy puddle or a confused pedestrian.

But fear not — the press release assures us this will 'accelerate training of operating system models.' Accelerate them toward what, exactly? Toward the day when your delivery robot mistakes a child for a mailbox? Toward the glorious moment when a simulated traffic light and a real one disagree? That day is coming, and Oasis 3 will have trained the robot to handle it with maximum confidence and minimum actual understanding. Cheers.