SNI: WEEK 45
Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. A week of gunpowder, treason and plot, in which America changed who runs Congress and the machines started building the machines.
tl;dr: Remember, remember, the third of November.
A wave of blue spread across the United States. Stocks sparked, blew up, then simmered. And AI, apparently the least important issue on the ballot, seemed to be the biggest thing on The Money's mind. It's clear The People want a say in who builds a data centre in their backyard.
So will this be remembered as the week that ushered in tighter AI regulation, slowed the build-out, and changed the pace of AI development? Or perhaps as the week The Machines started building The Machines?
Whilst eyes were locked on the tides of changing regimes, researchers at MIT had OpenAI’s GPT-6 Astra write the weights of small game-playing networks from scratch, and then edit Alibaba's Qwen without retraining it. The edit raised its maths score from 7.9% to 48.1%, and was likened to gene editing for neural networks.
Is this, as a Google researcher who reproduced the work put it, 'a new, step-change emerging capability' in the latest models? If so, a Congress that can now argue with Mr Veto may have arrived at just the right time for more nuanced controls.
But the 100GW of US data-centre capacity projected for 2030 has to be built somewhere, and of 47 local ballot measures on data centres, voters favoured tighter limits in 46. Of the $173m spent on political ads about data centres, almost none was in favour, and $4.8m of AI money couldn't save the candidate her opponent christened 'Data Center Darline'.
The same argument has reached Ireland, which has nearly 150 data centres. Labour leader Ivana Bacik called the result a lesson for the Taoiseach, who cautioned against reading across the pond. Dex Hunter-Torricke, once SpaceX's communications chief and Mark Zuckerberg's speechwriter, calls our approach to data centres 'let it rip', and urges Ireland to change it. Campaigners march in Dublin today, part of protests in 15 countries. The People want a say here too.
Wall Street initially cheered the result with its best post-midterm rally in 44 years. Then the powder went up and The Money seemed to lose its nerve. The Dow closed 3.44% down; the Nasdaq Composite fell 5.52%, into correction; and memory makers lost about a sixth of their value. Then came Seoul and Tokyo.
But will the build-out really slow down? Or will it proceed with a little more balance?
Two days after the vote, Google committed $17bn to a data centre in Wyoming, promising to pay for all its power and establish a community council. Goldman, meanwhile, has just raised its forecast for global data-centre capacity by 2030 from 217 to 273 gigawatts. So much for packing up the cranes. Perhaps the vote has simply raised the price of a yes.
But the fight over where intelligence gets built is only one of this week's plots. Governments are also deciding who gets to use it.
Venkatesh Rao's essay doing the rounds has a name for this: 'cognitive mercantilism'. The old idea of hoarding something valuable like bullion and denying it to rivals, now applied to intelligence.
And the week offered plenty of exhibits. Beijing blocked SoftBank's $5.4bn purchase of ABB's robotics arm, reportedly over where its training data would go. American pressure reportedly ended Oracle's $7bn Tencent lease. And Nvidia reportedly told cloud clients, verbally, to sell at least 51% of their computing services to non-Chinese customers.
SoftBank, ABB, Oracle and Tencent all wanted to do business. Governments had other ideas – and Nvidia seems to be doing the policing itself.
Which makes that MIT experiment a rather delicious plot twist. While governments were sorting intelligence into national camps, an American model was improving a Chinese one. An awkward development for the architects of technological separation.
And perhaps the beginnings of something bigger. Anthropic's amended IPO filing describes its models helping design, train and test their successors, under human supervision. It admits having missed a significant acceleration in its own progress before, and warns it may not recognise the next one 'until after it has occurred'. A future model, it thinks, 'may imminently meet' its threshold for 'dramatically accelerated progress attributable to AI'.
If it does, that might be very good news. In Shanghai, a lab run by a team of AI agents now designs and runs its own experiments from a simple prompt, from sequencing DNA to spotting a cancer-linked mutation. Anthropic says Claude helped its scientists map the first 3D structure of a vault particle in bacteria – a cell component whose purpose has eluded biologists since it was discovered in 1986. And in a study of 27,382 heart patients, AI's measurements of heart function from MRIs were slightly better than clinicians' at spotting who was most at risk, in seconds rather than about 13 minutes.
Not that the machines have mastered the art of getting things done without us. Fifty AI agents, told to build a civilisation in Minecraft, fell into bureaucracy instead. Four spent 11 minutes appointing a sole fertiliser operator for one piece of bone meal.
Sounds about right.
And on that gunpowder bombshell, here's everything else worth reading this week:
Biopharma:
The agent picked the drug, then injected it: Exin let Anthropic's Sonnet treat an epileptic mouse, a vet approving each irreversible step.
Second chances for shelved drugs, picked by AI: OpenAI-backed Biossil buys Theriva's antibiotic-degrading enzymes for $350,000 upfront and up to $91m.
UCB lets AI draw its antibodies: Amsterdam-based Cradle's models start on up to three of the Belgian drugmaker's programmes.
A bird flu jab built for strains still to come: Cambridge's AI-designed mRNA vaccine fully protected mice, where the stockpiled vaccine saved about half.
Britain bets on physics to feed the drug models: the Sovereign AI fund co-leads seed funding for Potential Sciences' claimed 100-times-cheaper chemistry.
Medtech:
Palantir's NHS success list meets the hospitals on it: 29 trusts NHS England counts as 'realising benefits' say they aren't using it.
The patient brings their records to Claude: athenahealth lets patients pick which parts of their chart the model sees.
Too few AI scribes pass the Welsh test: Digital Health and Care Wales abandons its £10m ambient-voice procurement.
Fewer 'inconclusive' verdicts from the smartwatch: Belfast's PulseAI wins FDA clearance after cutting them by 92% in a 296-person study.
AI gets 14% more scans out of the NHS's MRI machines: acceleration software already runs on 54% of compatible scanners.
Advanced manufacturing:
America's first humanoid stock: Agility lists on Nasdaq with $465m raised and $174,000 of first-half revenue.
The humanoid works a whole shift: Xiaomi says its robots sorted console side covers in its EV factory for eight hours.
Designed by AI in a week, finished by engineers: OpenAI's Astra designed the enclosure, circuit board and firmware for a Codex device.
Data centres outbid the carmakers for chips: Aumovio, the former Continental parts arm, expects a €1bn hit from rising chip costs.
Robots open the factory floor to women: about 35% of the workforce at ABB's Bangalore drives plant is now female.
Insurance:
The fraud gangs use AI, and so does Zurich: its connected AI flagged 860 of a record 4,359 bogus UK claims in a year, and more than 90% of Ireland's compliance chiefs think organised crime uses it too.
A telecoms software house buys into insurance's core: Amdocs pays about $340m for EIS and will extend its agentic operating system to insurers.
Some Hiscox quotes now take a quarter of the time: its AI submission tool, built with Google Cloud, now rolls out across retail.
The broker's best referrer is now a chatbot: Claude, ChatGPT and similar tools overtook partner referrals as Vouch's top source of leads.
No more hold music for a dental check?: Guardian Life hands about a million benefit calls a year to Sierra's AI agents.
But what set podcast tongues a-wagging?
Self-improving AI is the IPO pitch. Slow adoption is the price check.
Recording on election day, the Moonshots panel had no doubts. The labs are 'well into RSI' – AI helping to build better AI – said the venture investor Dave Blundin, so 'they're only going to double down'. Nathaniel Whittemore, host of the AI Daily Brief, said financial professionals now treat recursive self-improvement as 'one of the most important factors for the IPO' (see above). But he doubts better models lead in a straight line to revenue. Enterprises consume 'a vanishingly small portion' of the intelligence on offer, he argues, because 'inertia is real, corporate systems are lumbering, and adoption is hard'.
He welcomes the sceptics anyway: the louder the bears, the less likely a bubble. By Friday the money had grown more cautious. After Thursday's rout, prospective investors in Anthropic's float are now discussing $1.2–1.3tn, Bloomberg reports, down from the $1.4–1.5tn some had given bankers. Whittemore's bull case still stands: 'we're just not using it yet.'
The build-out's easy years are over.
Recording the morning after the vote, the New York Times's Teddy Schleifer, who covers money in politics, told Big Technology that the political energy around AI will now be 'channelled much more narrowly into the next battle around data centres'. Why else, he asks, would anyone start a data-centre super PAC? 'It's because you face political challenges.'
He expects the shift to stick. Senators elected on Tuesday sit for six years, and 'we're going to look back on the last year and a half as a bygone era when it was so unencumbered that we didn't appreciate it in the moment.' Nor, he thinks, could more money have changed the result: 'public opinion here was pretty baked.'
Propaganda is learning to feed the chatbots.
Carl Miller, a researcher at the London think tank Demos, told Intelligence Squared that people now turn to chatbots for news, health decisions and shopping, which makes whatever the models retrieve a target. Demos tested models on claims from a sanctioned fake think tank set up by Yevgeny Prigozhin, the late founder of the Wagner Group. The models endorsed it, or served it up as a serious side of the debate, 16% of the time. It is the first operation Miller knows of that preferred visibility in AI answers over search.
The credible sources are harder to reach: the BBC blocks scrapers and much good journalism sits behind paywalls. Yet Miller insists this is no 'council of despair'. Labs could block sanctioned sources – 'one engineer could do that in like a day' – and such campaigns are 'very visible because they have to be'.
Where are all the robots?
Josh Gruenstein opens his cold calls without mentioning AI: 'Have you ever considered a robot for 16 bucks an hour?' Nine in ten US factories and warehouses have no robot, the chief executive of the a16z-backed Tutor Intelligence told The a16z Show. Tutor rents its robots by the hour, with nothing to buy upfront, and has them working within a couple of hours of arriving on site.
Where the model can't cope, a remote human 'tutor' takes over, so every paid shift is also training data. 'If you can make money as a business, then you're probably also making data as a business,' he says. The robots take on backbreaking work in jobs with three-month turnover and an average worker aged 55, but 'we don't fully automate 100% of a factory. We automate some jobs.' Staff get attached, he says, because the robot 'makes their jobs easier'. Customers who order again get to name them. Some name them after their kids.
Thank you for reading Second Nature Intelligence. Come back next week for all the AI news you won't regret knowing.
