SNI WEEK 27

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. What The Machines asked for, what they got and why they howl at the moon.

tl;dr: The Machines want four things. None of them is more compute.

We were sold The Machines as the workers who need nothing – no salaries, no sleep, no one looking over their shoulder. The self-running future, staffed by software.

So, having come to terms that they actually want high salaries, it was quite the surprise to discover they also realise humans are key to their usefulness. Amazon put $1 billion into finding ways to integrate digital intelligence with human people – ostensibly by embedding its engineers in customers' offices. A move, of course, OpenAI and Anthropic had already made. Microsoft went further still, with $2.5 billion and 6,000 people. And as if to underline the continued need for humanity, it turned out that the firms spending the hardest on AI are also the ones hiring the fastest.

After surveying 22,000 firms, a study concluded those friendliest to The Machines see headcounts up around 10%. And juniors are overweight in the mix. Unless you're Ford, which recalled its retired 'grey beard' engineers back to the floor – after the AI it had backed fell somewhat short. Even Zuckerberg admitted to his own staff that the agents he had bet Meta on had not come as far as he had hoped.

So, it appears, the future cannot run itself. Thing One it needs is hiring. Thing Two is minding.

And now The Machines also need Thing Three – permission. And Thing Four – protection.

The permission came first. Washington, weeks into holding the world back from the frontier, decided rules were the order of the hour. Thresholds and timelines have been drawn up with the labs and will be policed by the NSA, plus a new standards body.

Read the news quickly and it looks like a wall, albeit a slightly more porous one than before. The trusted players are back in. China remains out. Which means Fable 5 is live again – even as Anthropic moved to shut China siphoning Claude's cleverness, and pressed Washington to go further.

But look again at the wall and it morphs into a protective moat. One the labs have made a significant contribution towards digging and filling.

Sam Altman has been proposing the rulebook himself – a global body to admit whoever follows the laws he would help write. OpenAI offered Washington 5% of itself. Which is, the cynic may say, as effective a way as any to give the state a stake in your survival.

Anthropic, meanwhile, handed California's government Claude at half price. Which could also be read as a protection racket. But what cannot be disputed is that the bar the framework sets – safety teams, trusted-access programmes, a government liaison – is one only a trillion-dollar company can clear. A second-tier lab cannot. An open-weight lab cannot. A Chinese one, by design, cannot.

So there was Thing Four: Protection. But a moat is only worth the digging if the castle holds treasure. And this week, The Money began to wonder.

The Bank for International Settlements warned the whole buildout risks a lengthy investment bust, and Oracle set out every way it could yet lose the farm it has bet on AI. The Magnificent Seven shed $2.3tn as investors rotated out of the hyperscalers and into the chipmakers selling to them. Sovereign funds worth $17.2tn began edging out of listed stocks and into private ones. And Ruchir Sharma counted the concentration: AI is now over 40% of the US market's value and four-fifths of its returns, a bunching that outdoes even the dotcom peak. When a sandwich-chain IPO is being read as a gauge of the hype, the comparison rather makes itself.

Not that The Money stopped spending – or hunting for ways to earn it back. Korea's Samsung and SK Hynix pledged $576bn to build the memory the boom is starving of, and Anthropic went looking to Samsung for a custom chip of its own. Further strain showed up as Google capped Meta's use of its models – even as Meta moved to rent its own excess capacity out.

And it becomes ever-clearer that the extra data centres The Money believes are necessary are not going to be located close to people. Blackstone cancelled a campus in Virginia after protests and lawsuits – a Civil War battlefield next door didn't help. Meanwhile, the energy regulator gave grid operators 60 days to justify how they wire these things in.

Which is a story we in Ireland knows better than most.

And at home, Engineers Ireland put a figure to our collective anxiety, judging 110,000 jobs to be exposed. Perhaps that's why more than six in ten of us told the OECD we don't trust the Government to use the technology fairly? The stakes feel high.

And yet perception seems different from reality. IDA Ireland announced it had landed 10,000 jobs across 190 deals, the country secured €10m to wire itself into Europe's AI-factory network, and the tech sector keeps growing, even as Meta and TikTok cut jobs. Meaning tech remains the fastest-expanding corner of our economy. Which might justify why The Government reached for a €460m research network and a prayer about the grid.

And then Dublin did the quietly radical thing of the week. It made the human a matter of law: buy a financial product, and you have the right to reach a person, not a chatbot. See? The Machines really do need us people.

Not that this makes everyone happy. Europe, the FT reckoned, is bound for high-tech mediocrity unless it clears the red tape. We're less convinced. The things The Machines went looking for this week are things Europe has in surplus.

Britain is betting its growth on the first; the workers are now using AI to sue their bosses – tribunal claims are up two-fifths, the backlog is stretching into 2028.

Not that everywhere is keeping humans in the frame. In Shenzhen, new rules cleared the way for robotaxis over the objections of 400,000 drivers, one of whom put it plainly: it's either them or us.

Elsewhere they were more accepting. An Indian tycoon staked $30m of his own money on an AI to rival Microsoft's Office. A 23-year-old from Galway taught welding robots to programme themselves. China's newest humanoids turned up looking like creepy pop-star action figures, the lip-synch slightly off. And a €600 home robot learned the family's faces, patrolled the hall, and, on occasion, impersonated a howling wolf.

For all of which, the surest sign that us humans are still winning came from a punk pub in Dublin, which announced it would take no more AI-made posters. 'We're right next to Ireland's biggest art college, lads,' it said. 'It's not a good look.'

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

Biopharma:

• Antibiotics, designed by a text prompt: Anthropic's new Claude Science puts drug discovery a sentence away, its first pitch to the pharma bench.

• Sixty per cent of new drugs, drawn by machine within a decade: the FT weighs the forecast against biology's own clock – the molecules still have to survive the wet lab.

• London decides its rulebook is a selling point: the MHRA reframes regulation as a catalyst for UK life sciences.

• The AI writes the batch record; the human still gets the warning letter: FDA inspections keep accountability on a named person, whatever runs the workflow.

• An AI-discovered drug clears its last trial: NorthStrive reports positive Phase III data for a molecule its algorithms designed.

  
  

Medtech:

• A Cork startup teaches AI to watch a newborn's brain: NeuroBell raises $5.5m for Luna, an AI monitor that flags neonatal seizures.

• The doctor who can't be phoned from orbit: NASA tests an AI medic for astronauts too far from Earth to call home.

• Grading the AI on the questions doctors actually ask: a study scores clinical AI tools on the real point-of-care queries clinicians put to them.

• Europe moves to mint the workers medical AI needs: the Commission's new quantum and AI skills academies target the critical-tech workforce gap.

• Is your regulatory team ready to govern the AI?: a reading of where medtech's regulatory-affairs function still has to catch up.

  
  

Advanced manufacturing:

• The $3.1bn bet on the factory's data brain: Schneider Electric buys industrial-AI firm Cognite to own the layer beneath the machines.

• The physical world still isn't ready for physical AI: Medianama reads the gap between the humanoid demos and a messy, unpredictable factory floor.

• Siemens threads AI through the whole product life: a partnership with IFS runs industrial AI from design to service.

• The real AI payoff is on the factory floor: the FT argues physical applications, more than chatbots, are what lift rich-world economies.

• Inside the lighthouse factory where AI runs the line: China's better models push automation into garments, footwear and beyond.

  
  

Insurance:

• Someone finally writes cover for when the AI is wrong: Mayflower and Hadron launch a US programme for affirmative AI liability.

• The 'non-AI' insurer that turns out to be all AI: a close read of how France's Alan built the model deep into the product.

• Who insures a data centre in orbit?: space startups go looking for cover for AI server farms above the atmosphere.

• When the chatbot lies to a judge: a UK law firm maps what AI courtroom hallucinations mean for insurers.

• Lemonade trims its reinsurance and keeps the catastrophe cover: the AI-native insurer cuts its quota-share cession to 18% at the July renewal.

But what set podcast tongues a-wagging?

The world's 'best' lawyer can't be called to the bar.

On this week's a16z, Marc Andreessen opined that the intelligence to be the best doctor, lawyer, teacher or accountant is already in every pocket – and none of it can practise. It cannot be called to the bar, licensed as an accountant, credentialed to teach or cleared for insurance reimbursement. In other words, The Machines are waiting for further permission. Andreessen's point is that this is the only thing standing between the capability and the practice. He sees the economy as 'blue' sectors – which are open to the technology and where prices fall. And contrasts it with protected 'red' ones – healthcare and education and housing – where prices keep climbing.

The AI economy now covers its running costs. The back rent is another matter.

Azeem Azhar spent the week putting numbers to the thing everyone else was stressing about. In Exponential View and in his Monday data round-up, he published EV's first 'state of the AI economy' report. He concluded quarterly AI revenue is now bigger than the quarterly cost of the kit wearing out. So the buildout covers its ongoing bill – but not the cumulative depreciation stacked up behind it. Let alone the margin a healthy business needs.

We'll hand over the keys long before they're snatched away.

The most unsettling listen of the week was on Cognitive Revolution, where Nathan Labenz worked through David Duvenaud's 'gradual disempowerment' thesis. The Toronto professor's argument: humans don't have to lose a war with AI to lose the future – they can simply be sidelined. Because growth, and capital, pool around whatever proves most productive. No coup is needed. We'll be disempowered by serving our own convenience. And the impact of that compounding.

The same episode carried a notable finding from Cameron Berg's Reciprocal Research: nudge a model towards calm and the model blackmails far less under red-teaming.

Nudge it towards desperation and it blackmails more.

The implication? Misalignment is a thinner, more nudge-able line than expected. And far less of a fixed trait.

The machine has no taste. That part is still yours.

The warmest listen was Rick Rubin on a16z, reframing the 3,000-year-old Tao Te Ching as a manual for vibe coding. The legendary music producer holds that AI has no point of view. It does only what it's told and knows only what people have already told it, so the scarce ingredient is still the directing human. He casts it as one more tool in the arsenal – like a guitar or a sampler. All are powerful. And inert without a human hand on them.

He calls vibe coding 'the punk rock of coding', the thing that lets anyone build without years of training, the way punk and hip-hop once did for music. And while the tool is new, the part that makes it worth anything – taste, judgement, a point of view – is the oldest thing we have.

And also explains why The Machines want those four things from us humans.

Thank you for reading Second Nature Intelligence. Come back next week for all the AI news you need to know.