SNI: WEEK 24

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. We discover what the machine started building, what the market paid to own it, and why you - most definitely - remain in charge.

tl;dr: AI made history. But humans still rule.

If we're vaingloriously helping to write history's first draft, may we respectfully suggest this week might have been something of a moment?

A few days before President Donald Trump shut down its latest release down, Anthropic said, almost sotto voce, that 80% of each new Claude is now written by current Claude. In other words, it's gone all recursive – the AI is building itself. And we can expect rapid exponential improvement from here.

The reaction to this was not muted. On Moonshots, home to the world's most accelerationist podcasters, the audience was asked to: 'Think back on the turning points of humanity. World War II, (the) Cuban missile crisis, nuclear weapons. This is so much bigger... The world just changed.'

Most will find that a little jazz-handsy, to say the least. But is the moment bigger than the launch of ChatGPT? Probably. Especially when paired with the US government aggressively asserting the primacy of human controls and ordering the model offline after its security protocols were breached. Thus Marring Anthropic's launch of Fable 5. Something the original - and inevitable - backlash had failed to do.

All of which rather overshadowed events at Cupertino. Apple duly delivered a rebuilt Siri – by renting Google's model and running it in the cloud. Which will likely be forgotten by history. But its decision to put a 20-billion-parameter model on its most capable silicon – and cleverly loading only the slice each task needs, keeping the AI private and offline, stands a better chance of being remembered. Perhaps as the moment the world realised digital intelligence didn't need to cost the earth.

And eyes were firmly lifted off-planet as SpaceX went public at $1.77tn - and duly 'popped'. The largest listing in history, priced at 94 times revenue, is actually barely floating at all - there's only a tiny sliver of stock available. And while some analysts thought it was worth half its valuation, others claimed it was under-priced.

So no regrets at OpenAI, as it filed to follow. The lab had just passed a billion monthly users and, declaring 'chat is dead', tore up old plans in favour of an agent-and-Codex 'superapp'. Just as Oracle was marked down 12% on a $70bn building spree. And we found out there are factories that can build humanoids – which nobody wants.

The biggest winners, though, were quieter. Caterpillar, Corning, Eaton, Nucor and Ford have all outrun the index on the back of the build-out - the picks-and-shovels trade - though Bain reckons $2tn of new revenue is needed to justify the spending.

And so while some of us might have taken losses, it was a good week for humans in general. Our judgement, it turns out, is still vital. Mathematical proofs need a human to check them. But don't take your abilities for granted. While those among us who argue back against chatbots get sharper, many of us stop thinking within minutes. Although not as dramatically as the addicts using drugs cooked with AI-derived recipes. Or, perhaps, those being influenced by China-linked social media accounts to turn against AI data centres.

And China might indeed be driving the US towards a more socialist republic. To ward off the mass unrest it's helping whip up, Washington has warmed to the idea of part-nationalising AI. While some want the gains taxed as income, others (including Trump) prefer an Alaska-style wealth fund. Seoul, meanwhile, also wants windfalls shared. Its leaning on Samsung and the AI giants to spread profits to suppliers and staff.

The FT calls the whole thing good politics and bad economics.

All of which which made Brussels' order that Meta must open WhatsApp to rival chatbots seem positively quaint – as did a German court holding Google liable for its AI Overviews. Although, at home, the Irish government seemed to take our suggested cue from last week – and put €460m into seven research centres to host some intelligence right here.

On that bombshell, here's everything else worth reading this week:

Biopharma:

• Britain opens a test track for AI medicines: the MHRA's new regulatory sandbox lets developers trial AI under the regulator's eye.

• Pharma pools its protein-folding brains: Nxera joins the OpenFold consortium of drugmakers and AI labs.

• The biotech IPO window cracks open: Parabilis and Kailera price well, though only de-risked names need apply.

• Sanofi deepens its bet on Owkin: the pharma giant expands the partnership for purpose-built drug-discovery agents.

• The model that out-designs human protein teams: Anthropic's Mythos 5 matched or beat human teams and ran it roughly 10 times faster, with strong candidates on 9 of 14 targets.

Medtech:

• The heart 'twin' that may not fit women: cardiac AI runs up against a gender-data gap.

• Six bets to shrink the waiting list: the NIHR funds AI projects aimed at waiting times.

• Approving a device that won't stop learning: the FDA's static rulebook strains against adaptive AI that changes after clearance.

• Medtech AI's 'prove it' moment arrives: the field turns from promises to outcomes.

• Korea's AI medical-device surge: a regulatory boom reshaping device approvals across Asia.

Advanced manufacturing:

• The line that spots its own defects: computer vision moves quality inspection onto the factory floor.

• Nvidia and SK Group fuse the chip stack: one alliance spanning memory, cloud and fab design.

• Britain puts £1.1bn into home-grown silicon: a national AI-hardware plan and a sovereign supercomputer.

• The lab that runs its own experiments: self-driving chemistry closes the design-make-test loop without a human at the bench.

• The factory robot that won't specialise: Theker raises $85m for a machine built to do a bit of everything.

Insurance:

• The cover that hasn't caught up with AI: Aon finds tech E&O priced soft even as AI exposure climbs, with coverage gaps opening where policies pre-date AI-as-the-error.

• Fraudsters deep fake the evidence: motor-insurance claimants are using AI to doctor crash photos.

• You can't insure a black box: restricted models and opaque benchmarks threaten the nascent AI-insurance market.

• Pricing a risk no one can assess: Gallagher Re calls for better AI model-assessment methods to underwrite it.

• AI rewires sales and claims: the front-office gains arrive with mounting operational risk.

But what set podcast tongues a-wagging?

The fabled backlash was mostly noise. Microsoft's wasn't.

On Thursday's AI Daily Brief, Nathaniel Whittemore called Fable 5 the most controversial model launch yet. Most of the fury was the usual online weather. There was much grumbling that the model had been throttled for frontier research. A stumbling block our own Phillip Black quickly discovered, ahead of the decision being reversed. But one objection has sustained and does have teeth. Anthropic's new terms keep enterprise prompts and outputs for 30 days, and up to two years if a safety classifier flags them – a sharp break from the zero-retention deals Microsoft had been running on. So Microsoft pulled Fable and Copilot from its own staff, while the lawyers read the small print. The worry is easy to picture: send proprietary code to a rival lab's servers, and it now sits there for a month.

The machine takes the work. The suits feel it first.

Take heart that even a model which is better than us at everything still needs you somewhere. On a16z, economists Tyler Cowen and Alex Tabarrok made the comparative advantage pitch – an idea that has reassured economists since Ricardo. Their logic runs thus: a model that beats humans still runs into limits of time, energy, compute, land and capital. And while it's busy with the highest-value work, the rest falls to people. And so we retain a comparative advantage! Which means the things that then bottleneck (the humans), capture a larger slice of the pie. Cowen also thinks that the people who lose sit higher up the ladder than the entry-level panic suggests. He reckons upper-middle professionals – lawyers, consultants, academics and doctors – are the most exposed.

Are you 'botsh*tting' it?

On Wednesday's Cognitive Revolution, Glean's Rebecca Hinds brought the clearest numbers yet on the enterprise AI paradox. The topline of a survey of 6,000 knowledge workers looks like a triumph: 87% now use AI, 73% feel more productive and each of us reckons we save about 13 hours a week. But then the floor gives way. Only 13% say their organisation actually performs better for it.

It's all because of what Hinds calls 'botsitting' – feeding the model context, debugging it, cleaning up after it – and that eats roughly 6.4 hours a week, about half the time saved. Most of it tedious and untracked. Worse still, it breeds 'botsh*tting': shipping AI work you can't explain or defend, which 69% admitted to.

The cost of building has collapsed. Taste hasn't.

Two other commentators framed where this leaves us. At Exponential View, Azeem Azhar argued the AI boom is quietly becoming an entrepreneurship boom – the cost of building businesses has collapsed. And on Lenny's Podcast, Tony Fadell – the man who built the iPod and the iPhone – spent two hours on the one input that hasn't got cheaper: taste and judgement. Only humans can decide which product to build, which detail matters and when the machine is lying to you.

So – what did we learn? The machine may now write most of its own code – and the market will pay almost anything to own the upside. But human judgement? It still rules.

Recur again next week for more self-improvement as you discover everything you need to know about AI in your sector - and beyond.