SNI: WEEK 25

What happened after the switch got thrown? What stayed on? And who kept building? Welcome to all the news in AI that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance.

tl;dr: Flick the off-switch - and watch the room light up

It's hard to imagine Trump understood the consequences. Washington's export controls left Anthropic no choice but to pull its most advanced models. The world wobbled. Then recovered its balance. And strode off purposefully in four different directions.

Some headed into space. Others into local compute. Second-tier nations doubled down on sovereign AI. And others, including Microsoft, looked to China.

So by the end of the week, despite its superiorities, it felt like America was the only one standing still. Somewhat alone.

AI had woken up a regulated industry and it was clear that American acceleration-ism had given way to AI ad-hoc-ism. Buffeted by two opposing sides - each with no grasp of the other's language. The administration has no feel for the technology. The industry appears unable to read any room Trump is in. And suddenly, Anthropic's moral compass seems naive - it having been comprehensively weaponised against it. No wonder then that, like the Iran War, there's no end in sight.

Not that American money took fright. With SpaceX's market cap almost touching $3tn, it bought Cursor for $60bn. All in stock. Then lined up a $20bn bond - just for sh*ts and giggles it seemed. So how, exactly, did this come to pass? Belief in Musk and his potential became market value. Market value became currency. And that currency bought a real company – which seeks to justify the original belief. All-in-all, a tidy machine. Until it runs backwards.

Which is why Wall Street took to calling Musk a roll-up artist, the sort who might yet fold Tesla into SpaceX itself. Just as others decided to do a different kind of folding. The FT watched companies rein in AI spend – and found the industry learning a hard lesson on pricing.

It also clocked OpenAI's $34bn of spending ahead of its IPO. Which probably made Accenture feel queasy as its own value seeped away, with its shares falling to a 2017 low. The Irish Times, meanwhile, questioned whether Anthropic is worth its $1tn. None of which persuaded Jeff Bezos against tipping $400mn into a barely-formed startup.

And others kept building too. France drew Foxconn and Nvidia into an AI hub. But Sequoia's Luciana Lixandru asked whether Europe can grow a trillion-dollar company.

Although perhaps the deeper worry, as TechPolicy.press set out, is that digital intelligence sovereignty needs more than frontier access. And, as if to demonstrate the point, Europe's own champion, Mistral, turned out to be a tad more vulnerable to Russian disinformation than one might like.

Which is less of a threat for Chinese AI. The world's move towards it was further helped by Zhipu's new coding model beating OpenAI's latest (and greatest) on some benchmarks – at just 1/6th the cost.

And such alternatives can only accelerate how AI is now senior-ising junior roles, doing the work the bottom rung used to learn on – for which job interviews are becoming AI tests. Another headache for HR teams – who must now learn to manage bots as well as people.

Which responsibility would you plump for? How about once you know that digital intelligence now matches or beats the advice of doctors?

Ireland caught the brighter side as well: OpenText added 400 jobs across Cork and Galway, Dublin greenlit a bill to police the EU's AI Act, a Fianna Fáil MEP proposed AI guidelines for the parliament, and Irish AI job ads jumped 60% in a year.

But it wasn't all gravy. The Irish Times caught KPMG passing off AI work as its own, the same week it was featured in a home-loan breach report.

All of which explains why so many questions are being asked of the technology. Corkers like 'Can a machine do this job?' and 'How do we manage the revolution?' were bowled to readers this week, alongside opinion pieces as to why happy humans still matter, why the brain is no machine, and how history will judge today's bosses.

And indeed the future ones. Two Irish teenagers said they really should be revising for exams – but are building an AI company headed to the US instead. If that doesn't work out, someone should tell them that the UK Cabinet Office is on the hunt for an AI 'influencer' to make Whitehall feel keener. But if they do land Stateside, then they might want to remember that almost half of American singles feel iffy about AI in dating.

The moral of the week? Humans remain securely in the loop. And, it seems, in our beds.

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

Biopharma:

• An AI-designed drug meets its first patient: Insilico doses the first human with an AI-discovered molecule.

• DeepMind's drug-hunting spinout banks $2.1bn: one of the biggest cheques yet for designing medicines by algorithm.

• LabGenius bets on an antibody no human drew: an AI-designed, tumour-targeting cancer critter.

• Merck buys deeper into the algorithm: a pact worth up to $510m extends its AI drug-discovery push.

• The startup that mined its own failure: a biotech rebuilds a flopped clinical trial into an AI model that learns from the wreckage.

Medtech:

• The X-ray that saved me: a patient tells the Health Secretary that AI-backed scans 'gave me my life back'.

• Doctors fight the algorithm that denies care: the AMA and lawmakers push back as insurers use AI to refuse treatment.

• Medtronic backs the device business with AI: the giant spins off diabetes and reorders its strategy around AI.

• £30m to find more cancers, faster: England funds a national expansion of AI cancer diagnosis across the NHS.

• The hottest medtech hire is a compliance brain: AI-regulation expertise becomes the sector's tightest hiring bottleneck.

Advanced manufacturing:

• Ireland builds its own chip brains trust: a new €71m semiconductor research centre anchors the Silicon Island plan.

• Samsung sets a deadline for the lights-out factory: every plant to run on AI by 2030.

• Unilever lets the digital run the line: the twins graduate from monitoring to process control.

• Nine in ten humanoid robots now ship from China: and Chinese models top the latest robot AI benchmarks too.

• AI goes looking for a better battery: machine-led discovery is speeding up battery development.

Insurance:

• Zurich writes cover for the bots: embedded insurance for Hong Kong automatons.

• When everyone's ranged against the machine: insurers flag 'AI correlation risk'.

• Aon's AI reads the fine print: a new platform helps insurers assess exclusions.

• Squaring the actuary: an academic study reconciles the EU's AI rules with insurance-specific regulation.

• Agentic AI rewrites the liability: autonomous agents are reshaping professional-liability risk.

But what set podcast tongues a-wagging?

The economy is one big bet on AI training.

On Tuesday's AI Daily Brief, Nathaniel Whittemore did the arithmetic: AI investment drove roughly 75% of US GDP growth in the first quarter. Data-centre and hardware spend now runs at 1.4% of GDP. The whole edifice, he argued, rests on a single contract – the labs' token-consumption revenue has to keep climbing fast enough to justify the buildout underneath it. Which means training has to close the gap fast enough for everyone's colleagues to become as addicted to the tokens as the readers of this newsletter.

And perhaps that's possible. A day later on Big Technology, Ramp's lead economist Ara Kharazian revealed that enterprise AI is the fastest-growing category the organisation has ever tracked – up 15x per firm since January 2025. And, given it still amounts to just 2% of business spending outside payroll, even for the heaviest adopters, there's plenty of headroom left for growth.

Especially as the median firm spends $11 per employee a month. Making Uber's cap at $1,500/mo per employee feel less lukewarm.

But as the need for token spends ramp up, watch out for further training initiatives from the big labs. Or, a far better idea, give Brightbeam a call. We'll help you save on tokens too.

People prefer predictability over power.

On Thursday's AI Daily Brief, Whittemore pointed to what might be filling Anthropic's Fable 5-sized hole: Not a single rival but an open-weight 'worker' delegating to a closed-frontier 'advisor' – a split that cuts cost and raises performance. Whilst not being vulnerable to shutdown.

As the War with Washington continues, will smart routing become the competitive edge that brute-forcing frontier AI once was?

The council of models quietly bins its own best ideas.

Exponential View ran a guest essay by Rohit Krishnan with a finding that should give pause to anyone wiring up a 'panel of models'. Krishnan tested whether a council of LLMs preserves the best thinking, and found it does the reverse: a blended council kept only about 24% of the high-value ideas a single model had produced on its own – peer-review fared no better, at 22%. What the council reliably keeps is the consensus; the spiky minority insight gets smoothed away.

To explain why he reaches back to Stasser and Titus's 1980s work on 'hidden profiles' – groups, human or machine, over-weight what everyone already shares and under-weight what only one member knows. The fix lives in the protocol around the council: explicitly gathering, storing and ranking each model's distinct ideas before anything gets synthesised.

For a board reaching for multi-agent architectures because more models sounds safer, that's the uncomfortable read – averaging up the median can mean averaging away the one answer worth having.

Generation is becoming solved. Evaluation is now the human's job.

On Cognitive Revolution, Elicit's Andreas Stuhlmüller and Jungwon Byun raised another significant AI use issue: frontier reasoners remain too easy to push around. A model that gives a finding a 30% chance of being wrong will flip that judgement on a single leading follow-up, and an outcome-trained model will cheerfully claim to have read 100 papers it never opened. Its stated confidence, Byun noted, floats free of any real understanding underneath. If you want to understand why, give this piece from Brightbeam's Phillip Black a read.

Elicit's solution: Let the machine generate. Leave the human to evaluate.

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