SNI: Week 46

Second Nature Intelligence: Sector Updates, Week 46, over an image of the sun.

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. A week of feeding frenzies, in which the gulls circled Nvidia, The Money circled Anthropic and the picky eaters did rather well.

tl;dr: Is the world eating itself?

Sunday: Nvidia’s board has ‘confidence in Jensen and his leadership team’ (opens in a new tab).

Wednesday: ‘ousted’ (opens in a new tab).

Seventy-two hours to topple Jensen Huang (opens in a new tab), the face of the AI boom. And The Commentariat dived like gulls on a trawler.

Three affairs (opens in a new tab), reportedly. One with Alice Huang, a former Megaspeed executive whom US officials suspect (opens in a new tab) of working for Chinese intelligence. Singapore says she ‘was not investigated for espionage’ (opens in a new tab). The allegation remains unproved. But when did that ever stop a gull?

But the bigger question isn’t who Jensen Huang had been seeing. It’s what he leaves behind.

A company that hasn’t just been selling the AI feast. It’s been picking up the bill. $124bn of stakes and commitments (opens in a new tab) by July. A conditional guarantee of up to $105bn (opens in a new tab) backing OpenAI’s Ohio leases. In effect, AI’s backer of last resort. Some on Wall Street call it ‘the Bank of AI’ (opens in a new tab).

So who inherits the bank, for now? Colette Kress (opens in a new tab). The finance chief who’d spent years as the one executive willing to check Huang (opens in a new tab) – and who saw a proposed $250bn arrangement for that same Ohio data centre pared to $120bn.

The Money, meanwhile, was gorging on Anthropic (opens in a new tab). Or was it? The float was supposed to be worth $2tn (opens in a new tab). Then perhaps $1.1tn. On Tuesday, up to $1.4tn (opens in a new tab). By Thursday, six times oversubscribed (opens in a new tab), with talk of $1.5tn (opens in a new tab). Nobody knows what Anthropic is worth. That hasn’t stopped the Bank of AI joining in the feast (opens in a new tab), alongside Microsoft and Google. Even the chipmaker behind the chipmaker is in: TSMC set aside up to $1bn (opens in a new tab) for shares in Anthropic and OpenAI. Anthropic prices next week.

Not that buying in keeps you safe. Just ask Microsoft: Anthropic’s investor (opens in a new tab), reseller (opens in a new tab) and partner. Yet Claude’s new office suite (opens in a new tab) is coming to eat it too.

So what will be left once the gorging is done?

More than we can eat, it seems. Before its float, Anthropic released Fable and Mythos 5.2 (opens in a new tab). By its own account, they’ve modelled the shells of gene-therapy viruses, designed circuit boards ready for manufacture in under a day and spotted that a loan applicant’s salary deposits didn’t match the bank statement.

But Wharton’s Ethan Mollick asked Fable to polish some virtual worlds (opens in a new tab). And soon it had written a rule book he never asked for and more than 87,000 words. So he hired consultants. AI ones. And their report was ‘too long and dense for me to bother reading’.

He wasn’t the only one with a little indigestion:

The Machines, it seems, don’t always know when to stop.

And yet in that same Linux kernel, Linus Torvalds, Linux’s creator, says much of the latest progress was ‘driven by various AI fixes’ (opens in a new tab). The machine flooding the inbox is starting to clear it. And for now you can’t doubt it’s feeding the economy, with Britain’s growing at its fastest pace in more than two years (opens in a new tab).

The key might be choosing what to eat and how to eat it. Banks have been banning AI agents from customer accounts (opens in a new tab), because nobody yet knows ‘who would be responsible’ if a bot goes rogue. And Ki, the Lloyd’s syndicate, uses AI ‘a lot’ – just not in its underwriting algorithm (opens in a new tab). Whereas Crosby, an AI law firm, insures its AI paralegals at Lloyd’s for $5m (opens in a new tab). Long live the picky eater.

And in Dublin, where EU leaders discussed AI sovereignty (opens in a new tab) on Friday, we had our own seat at the table. As Europe’s home to so many of the big players, our exposure to ‘all things AI’ is significant (opens in a new tab). Yet Ireland’s proposed changes to the EU’s draft cloud law could make it easier for US Big Tech to count as ‘sovereign’ (opens in a new tab). And our Economic and Social Research Institute says that, in it’s modelling, the feast might just flow to the top (opens in a new tab).

So, is the world eating itself? In places, yes. The Commentariat feasts on scandal. The Bank of AI feasts on its own customers whilst they feast on it. And the labs provide more than we can eat whilst eating everything themselves.

Yet a lot of it tastes good.

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

Biopharma:

Medtech:

Advanced manufacturing:

Insurance:

But what set podcast tongues a-wagging?

Companies budgeted for seats. They were billed for work.

The headlines said companies were capping what staff could spend on AI. Nathaniel Whittemore went looking for the totals instead. On the AI Daily Brief (opens in a new tab), he worked through Menlo Ventures’ annual survey of nearly 400 US enterprise decision-makers, flagging that Menlo is an investor in Anthropic. It puts enterprise AI spending at $33bn last year and almost $177bn this year.

The share of firms using AI only crept from 55% to 65%. But spend per firm rose 4.7 times, against the 1.9 times the average company had planned – about two and a half times its budget. The reason, Whittemore argues, is that ‘AI spend stopped being about seats and started being about tokens’. Agents keep working after people step away, so ‘spending then scales with work, not with headcount’. Budgeting for AI, he says, now looks more like budgeting for cloud compute.

So a licence count no longer predicts the bill.

Every agent can be up while the job goes undone.

Cisco’s Vijoy Pandey has been wiring together agents from different vendors for customers such as Swisscom. On Cognitive Revolution (opens in a new tab), he told Nathan Labenz that connecting them is not the same as getting them to think together. In one travel-booking system, he says, the agents were 66% ‘up’ by the measure dashboards use, and completed none of their tasks. ‘Your dashboards might be green, but your outcomes might be wrong, completely wrong.’ He calls these ‘silent failures with really loud consequences’.

The cause is ordinary misunderstanding, compounding. Each agent is evaluated on its own job, never on the shared one. In Cisco’s five-agent systems, by his account, one misalignment took accuracy from 99% to 98%, two to 81% and three to 16%. Cisco’s open-source answer reads the agents’ traces for drift and has them agree intent before the work starts. A human stays ‘on the loop’ to make the judgement calls.

Calling the election was the easy part.

Thoughtful’s Brian Zhou gave nine frontier agents seven days and a $500 budget each to forecast about 800 US midterm races. On a separate Cognitive Revolution (opens in a new tab), he reported that they did ‘just as well as the best forecasters’ among humans on preliminary results, and the best beat the prediction markets on the Senate.

Then came the money. Paired trading agents lost heavily on election day: they ‘didn’t react fast enough and sometimes made up incorrect information’. One agent’s setup submitted a random forecast within minutes, because its task ended once a checklist was met. A good answer and a good decision turn out to be separate tests.

A rat is a poor guide to a human. Isomorphic wants a better one.

Max Jaderberg, president of the Alphabet-backed Isomorphic Labs, told No Priors (opens in a new tab) that medicines mostly fail in the clinic, from a toxicity or off-target effect nobody could see until the drug reached patients. Regulators want molecules shown safe in rats first. ‘Choose your study,’ he says, ‘but it’s about 30% predictive of human clinical tox in a rat.’ If one of his engineers brought him a model that good, ‘you’re like, okay, maybe do a bit better than that’.

His alternative combines lab assays, organoids and organs on a chip with AI models that predict human toxicity directly. Every Isomorphic leaderboard carries a dotted line marked ‘experimental accuracy’ – the point at which a prediction can stand in for a lab result, as AlphaFold did for protein structures. Past it, the lab becomes the place to confirm what the computer has already found.

He gives no date for Isomorphic’s first clinical trial.

Thank you for reading Second Nature Intelligence. Come back next week for all the AI news you won’t regret knowing.

Related

More from Brightbeam

So tell us what you’re working on.

Get in touch