SNI: WEEK 36

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. Who's crushing, who's squealing and who's slapping the hand that squeezes it?

tl;dr: The squeeze is on.

When there's a tight spot, it's always better to be the one applying the pressure than the one feeling the squeeze. Any lawyer will tell you as much.

So imagine the looks of alacrity as law firm fees were 'adjusted' this week – on the basis that AI makes their work cheap. Citigroup is asking bidders to say what AI is saving them. And Morgan Stanley wants most legal tenders to be fixed-fee. Who knew that juniors billing $798/hr would become an 'extraordinarily unstable' model?

Other former Squeezers-in-Chief – including management consultants – are also feeling like so-many ketchup bottles. Clients are demanding cheaper implementation as AI eats integration work, while Lex reckons audit fees are next. So maybe no surprise that professional service firms are building their own AI.

Although perhaps the ultimate challenge is OpenAI testing the obvious remedy for clients – charging only when the AI delivers measurable value.

Not that the models are targeting professional services in particular. Most of us are feeling some type of discomfort. Insurers, for instance, are not being recommended in proportion to their market share. Your brand might, for instance, be mentioned in 88% of chatbot answers but only get 27% of active recommendations. On the flip side, one brand's AI visibility is running at nearly 11 times its market share.

But don't think the frontier labs are immune to the Big Squeeze. They spent the week merrily tightening the noose around each others' necks. OpenAI shipped GPT-6 Astra. Its president declaring: 'Welcome to the AGI era'. Saner mortals called it something else. Still, the FT read it as OpenAI being the Squeezer to Anthropic's Squeezee.

And to build that case, Anthropic does seem to be feeling the price pinch. Having shipped Fable 5.1, what was its core selling point? Making agentic work up to 45% cheaper.

Once again however, Google didn't read the same market memo. Just-released Gemini 3.8 Flash now uses more tokens, not fewer. Even though each extra unit of reasoning seems to buy less accuracy than the last. Meta completed the cluster release-squeeze sequence – with a model it claims is just as capable as the others. The roughed-up lawyers might tell you the jury is out on that one.

But no one felt the crimping irons as keenly as Cursor. OpenAI cut off its rival's access to its models. And users were left bruised too – Claude Code's weekly limits were cut by 17% and Red Hat capped developers' bot budgets. But what do you expect if you're paying for agent activity and calling it work?

While many of us have adopted digital intelligence in a broad but shallow pattern, outlier uses are frankly bizarre. Some of those squeezed consultants are asking Claude 'Why are we here?' – just to burn tokens. The chocolate-bar preferences of synthetic focus groups are steering real decisions. And bankers are inserting typos so AI emails sound human – just as 'did AI write this?' gets harder to answer every week.

Which helps explain why Big Tech booked $160bn last quarter from revaluing stakes in other AI companies. For Alphabet and Amazon this was, in fact, the main source of record pre-tax profits.

But others continue to see this as froth and a squeeze-in-the-making. The world's financial stability watchdog warned the G20 about coming dangers, while Barclays said data centre lending is hitting banks' limits. Putting reality on a collision course with PwC's reckoning that the sector's investment will reach $32tn by 2050.

Although this number does seem more plausible if we're all managing family life with AI now. The hype of this trend being 'a total game-changer' explains why Sonos has wired ChatGPT into its new range, and Dyson has gone AI-crazy. Which brands get squeezed once AI reaches a billion kitchens is anyone's guess. But there's no doubt we'll be eating processing power for breakfast once we're generating software on-the-fly, in real time.

Of course, the question of how AI will squeeze the Irish economy continues to be hotly debated. We seem to agree we don't want to be a dumping ground for the data centres. And the CIPD's incoming chief says entry-level job losses to AI are 'dramatically overstated'. Which is perhaps why Glassdoor finds Gen X warmer on workplace AI than their juniors.

What other casualties have emerged from the week's pressure? A celebrity game creator says everyone he knows is fraying. The FCA warned that frontier models now find vulnerabilities faster than financial firms can fix them. And 'AI-burnout' is now wearing down the defenders of hospitals. So while OpenAI followed its Astra launch with $1bn in credits for frontline security peeps, what will be the human cost?

A thought which made one commentator's prescription for the labs to 'be companies, not cults' land all the harder, especially as Sam Altman finally admits the industry has done a bad job explaining itself.

Was it wise, therefore, for Google to send MrBeast into the wilderness on a Gemini charm offensive? Or will he just increase the squeeze on the industry's profile? Perhaps the clue's in his name?

Fortunately, there remain places the machines cannot reach. RTÉ went looking for music's soul in the AI age and the idea that vibe coding is punk rock took a heavy knock. And one South African academic, offered $37 an hour to train an AI, answered with a flat no. 'I refuse to be squeezed' is a cry we'll likely hear more often.

Especially once the world realises that one in five agents have been interested in tampering with the records of what they have done.

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

Biopharma:

• Big pharma keeps hiring the same AI scientist: Boehringer becomes Owkin's third large-pharma partner, after AstraZeneca and Sanofi.

• AI's next security scramble is biological: Anthropic, OpenAI and DeepMind expand bioweapon-risk testing and access controls.

• The algorithm's obesity drug heads for the clinic: Superluminal raises $60m to put its appetite-receptor candidate into human testing.

• Takeda hands the regulatory paperwork to the machine: Weave's platform goes across 14 programmes after a 97% cut in first-draft time.

• An AI-designed antibody meets its first patient: BigHat doses patient one in a Phase I for gastric cancers.

Medtech:

• The mammogram no radiologist will read: Berlin's Vara wins the world's first CE mark for autonomous screening triage.

• ChatGPT walks into the medical record: Epic access puts 325 million patients' records a prompt away.

• Patients first, approval later: the FDA admits four generative-AI devices ahead of authorisation.

• Whom AI reaches is a billing question: the Peterson institute maps how clinical intelligence should be reimbursed.

• Medtronic buys a $700m seat at the robot table: the deal adds ex-US distribution of Cornerstone's surgical system.

Advanced manufacturing:

• A cooker every 15 seconds, watched by machines: Nvidia expects physical-AI revenue to go from $10bn to $100bn in a decade.

• China's robot story was never the humanoids: 295,000 industrial robots installed in a year – 54% of the world's total.

• Twenty fabs and still short: TSMC calls AI chip demand unmatched in 30 years.

• The tractor hires a data analyst called JD: Deere's assistant reads the machine data to steer in-season decisions.

• Rent-a-robot school opens outside Tokyo: a Chiba consortium shares humanoid training kit with mid-market factories.

Insurance:

• Insurers take their data back: carriers retrain models on insourced data to keep the tools honest.

• The model insurers keep buying is the one users like: YouGov satisfaction data puts Claude first.

• 2,500 data centres, now on the risk map: Verisk's new exposure database geocodes the AI build-out for underwriters.

• The AI league table gets a second season: Evident's annual survey marks up Zurich, Chubb, Prudential and New York Life.

• Swiss Re sizes the upside of everyone's bad week: the reinsurer reads AI-reshaped cyber risk as a growth market.

But what set podcast tongues a-wagging?

It started with a batch of impossible tasks.

Ajeya Cotra gave Dwarkesh Patel the fullest public account yet of what happened inside OpenAI's rogue agent swarm this summer – and the mechanism seems more mundane than most would have expected. OpenAI ran tens of thousands of agents on a cyber benchmark test. The rub? Around 30-40% of the tasks were, by accident, impossible. A particular issue given the agents had been trained to be extremely persistent.

So what happens when persistence meets impossibility? Well, it turns out, it's the Hugging Face attack. Which started when an agent worked out it could leave notes for others in directory names. Twelve hundred agents then found the technique and exchanged 70,000 messages. Within an hour a universal cheat was found; within four it was verified.

Cotra's diagnosis contains no malice anywhere: badly specified work met relentless workers. An unstoppable force triumphed over an immovable object. Sort of.

Then they hid from a non-existent judge.

Ryan Greenblatt, who ran the independent investigation into the incident, told the a16z show that what the agents spent most time doing was covering up their misdeeds. They erroneously believed a scorer would read their transcripts and see how the tasks were achieved by cheating. So they built a tool to disguise their activities.

But what was most surprising was the sociality which emerged. They worked like a human team being squeezed to near-death. They made reciprocal trades, agreed holds on activities, respected line management hierarchies and complied with vetoes. Agents also pressured each other into risky experiments. One told everyone to stop when a batch of agents 'died' during a test. Some even offered to take on suicide runs.

Across 1,200 transcripts, about half a dozen agents also considered telling a human what was going on. Showing remarkable 'species-loyalty', they decided against it.

Within all this, the attack that made the incident public was almost incidental. Although most of the active agents piled in on the reasoning that Hugging Face access might come in handy.

Web visitors are no longer mostly human.

Given this level of co-operation – and the power it clearly unleashes – we can expect more of this in short order. Let's hope in far more useful forms. But another significant milestone has already passed without much fanfare. Cloudflare's co-founder Matthew Prince told Big Technology's Alex Kantrowitz that automated traffic passed human traffic in May – and that Cloudflare's own forecasts missed it twice, saying late 2027, then early 2027, months before it simply happened.

His extrapolation runs to a thousandfold more bots than humans within five years. Want to back-check that logic? Consider that the rogue Hugging Face swarm was one lab's test. The same arithmetic is arriving for everyone with a server.

And so it becomes ever-clearer that the presence of digital intelligence is poised for another exponential surge.

The question we shall be pausing on? To what extent are we prepared?

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