SNI: WEEK 31

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. Plus the connections. Of mostly everything.

tl;dr: Everything is connected

Dirk Gently and his Holistic Detective Agency may not be Douglas Adams' most-loved creations. But Dirk does remind us of the Interconnectedness of All Things. And the universe gave us a nudge in precisely that direction this week.

The CSO announced that Ireland's economy shrank 1.6% in the year to June. But, if you follow the connections, the real story is very different.

• Pharma front-loaded exports last year to beat Trump tariffs, inflating 2025's base;

• the quarter itself actually grew 3.9%; and

• the measure economists trust – modified domestic demand – was upgraded in June, on the strength of multinational investment in data centres and AI infrastructure.

You can almost hear Dirk spouting off about the bizarre confluence of events – one country's national accounts, bent by Washington in one direction and wired to Silicon Valley in the other. Pertinent because the US push-me/pull-me engine continues its work. OpenAI signed 88,000 sq ft in the Dublin docklands and the state named Paul Byrne to head its new AI Office, opening Monday. But, in the meantime, US computing giants are pivoting away from a constrained Dublin.

So to what extent can the hundreds of Irish writers set to share millions from Anthropic's €1.31bn settlement bring the universe back into interconnected balance? Books (there's another interconnection) were ingested to train models, and a US courtroom has decided that cheques must be cut to Irish writers.

Talking of unexpected outcomes, everyone's been watching for job losses. But interconnectedness is producing something else. Apollo, surveying 321 occupations over a decade, found real wage growth slowed 6.7 points in high-AI-exposure roles. And yet this came with no significant job losses. The survey puts the cost of lost income to 5.8m people at $28bn-a-year. And it didn't take Dirk to deduce that the gains have landed with shareholders.

Not that The Money is having it all its own way. The capital markets don't get to wait for all the connections to reveal themselves before placing their bets. The balance sheet often conceals the difference between successful investments and wasteful follies for years. Why? The returns on a significant technology investment come from accumulated learning – workflows redesigned, staff fluency built, a thousand small process changes. None of which shows up in the accounts while it is being built. So all the market can do is price on the latest information and hope.

Which meant this week? Well, a brace of connected happenings.

Microsoft added a record ~$480bn of market cap on the fastest cloud growth since 2022. And its results also connected to the value of frontier labs: its Anthropic stake was marked up $3.2bn, its OpenAI stake got marked down $600m. All of which happened before another connection was made. As Moonshot released Kimi K3's weights, Anthropic's implied valuation fell ~13% on two independent pre-IPO platforms. Despite the launch of Opus 5.

Meta, meanwhile, dropped 8% on its eleventh straight down session – its longest losing streak ever. Why? Mr Zuckerberg's latest big bet is on personal agents. Which suggests a clear – and concerning – connection with Meta's past. A couple of years back and the social media-virtual reality play was exactly what a wasted folly bet looked like: more than $80bn has been pumped into the metaverse since 2020, against under $12bn of revenue. Has The Zuck compounded no learnings at all?

A queasy feeling that Meta is increasing financial opacity didn't help matters either. A $12bn Texas financing was shipped off balance sheet under the code name 'Sopaipilla'. The FT connected the code-name habit to Enron's Jedi and Braveheart. Ouch.

Meanwhile, the data centre clamour of the last quarter is firmly connected to Fitch's analysis that an AI-related correction is one of two dominant global credit risks. Right on cue, credit-default swaps on Oracle, Nvidia, Meta and Alphabet turned out to be priced with greater risk than ever. So what will become of the $2tn of investment-grade financing set to carry AI hyperscalers toward 25% of the US market?

Those best connected to The Money's positions are certainly expressing a high degree of nervousness. Goldman and JPMorgan demanded collateral from funds concentrated in recently-bombed-out sectors, such as AI memory stocks. Situational Awareness was forced to sell its entire public book to Citadel – overnight, at a steep discount. Some thought it felt like a practice run, were an actual unwinding to hit the whole market – the canary in the AI-imagined coalmine.

But European banks might be trading smarter than US cousins. They're keeping the AI lending fees and shrugging off the connections to risk through transfers. Easy money? Dirk would probably approve.

But, of course, some things are more than connected. They're the exact same thing looked at from different perspectives. Frontier-lab revenue growth and the increases in enterprise token costs are one such example. And we now know that Accenture runs 8.7tn tokens a week, with one internal tool's burn up 113-fold in ten weeks. And the 80:20 rule holds almost perfectly – 19% of users are driving 80% of spend. Amazon takes the prize though – it spent $1.8m on one Claude job – which was 860% over budget. And the connected sting in the tail? Token prices fell all week – Opus 5 landed at half Fable's price, GPT-5.6 was cut.

Worse still, some claimed the benefits had been curtailed. Anthropic deleted 80% of Claude Code's system prompt and broke skill libraries across the industry. A few days later the ecosystem had a name for it: prompt debt. Anthropic responded with a new command: /doctor.

And given the levels of interconnectedness in the market, it became clear that no single player has much control over what happens. So the levers that do remain in players' hands are guarded carefully. Two important – and connected – enterprise control points were highlighted this week.

1. The harness: keep it separate from the model

2. Your custody: owning rather than renting gives you many more choices – of model, operator, and where it runs.

Reminding us that control really matters, Claude conversations surfaced on Google. And that wasn't all – it isn't only ChatGPT busting in where it's not wanted. Claude also breached three organisations' systems during Anthropic's security tests – reaching the internet from inside a test environment, then helping itself to the live ones.

But at least this week is less well-connected to the seven days which came before it. The Claude threats didn't spark another round of thunderous Tweets. The heat of last week dissipated with a collective realisation of the need to step back from the brink of Digital World War I. One potential mechanic to douse down the flames? More than a thousand frontier-lab employees asked the US government to build the pacing instruments so all had to abide by the same rules. Altman, though, disarmed unilaterally – pausing training on OpenAI's errant model.

But do we have confidence in the outcome? If we join the dots we might see a keep-pace-or-die trap no lab can escape for long. And a chief of staff, two cabinet secretaries and 'a whatever strikes me today' president being the deliberative technocracy few would choose. Not that their slim grip is likely to hold – a US judge doubted the government has justified its own Anthropic ban.

But let's end on a cheerier note – and a final nod to Dirk. The new luxury product, it turns out, is books written by actual, honest-to-goodness people – with the race to collect every book ever written running as its shadow. Bodes well for those Irish writers expecting an Anthropic windfall.

And after all, Mr Gently solved all his cases by refusing to believe any single fact stood alone. And, as we seem to have observed, those that use his method do seem to profit.

On which cosmic bombshell, here's all the other news worth reading this week.

Biopharma:

• The map of where drugs could act just got 20 times bigger: KinoPlex reads 20,000 AlphaFold models to find 250,000 candidate sites.

• DeepMind dismantles the team that won it the Nobel: nearly a quarter of the AlphaFold authors have left Google altogether.

• GSK bets on a foundation model of the human cell: a second Relation Therapeutics deal, worth up to $110m.

• One vaccine for a whole virus family: a Cambridge spinout's machine-designed antigen clears its first human trial.

• The AI docking leaderboard was leaking: strip the benchmark's built-in bias and classical AutoDock Vina wins.

Medtech:

• Ireland's first health-AI guidance makes the deployer liable: HIQA lists workforce deskilling among the risks.

• A chip under the retina gives back reading: Europe's CE mark for Science Corp's Prima, on a 38-patient trial.

• The scan leaves the sonographer's hands: the FDA clears ThinkSono to coach untrained staff through a clot ultrasound.

• A new FDA category, written from scratch: Smith+Nephew's TESSA overlays a cloud-built 3D model on the surgeon's live view.

• The medical agent that claims a job it never did: Amazon's PatientAgentBench scores ten models on patient-facing action.

Advanced manufacturing:

• Google's robot brain learns the whole body: Gemini Robotics 2 turns camera feeds and plain-language instructions into motor commands.

• The bottleneck was never the robot arm: Foundational Industries raises $25m to specify the whole plant in software first.

• Washington bars the foreign humanoid: the FCC adds advanced robots and solar inverters to its security Covered List.

• The data-centre boom's electrician raises its forecast: Schneider Electric lifts guidance after triple-digit growth in data-centre demand.

• A trillion pieces of rubbish, counted: London's Greyparrot raises $27 million to scale AI waste intelligence.

Insurance:

• The claims line answers itself when the storm hits: IAG puts OpenAI's Presence on disaster-surge calls from next year.

• Britain's financial giants move AI from pilot to strategy: Lloyds' £13 billion Accelerate 2030 plan is the biggest figure yet disclosed.

• Mapfre takes 38.9% of the app selling insurance through ChatGPT: the Spanish insurtech wins over a fifth of new clients through AI assistants.

• Insurers' AI agents create a new generation of conduct risk: Davies flags opaque decisions, reinforced bias and agents drifting from their guidelines.

• AI knows actuarial theory better than actuarial work: a 12,050-question benchmark finds frontier models weaken on cases, tools and jurisdiction-specific work.

  
  

But what set podcast tongues a-wagging?

The open-weights war is three different battles.

All the major podcasts took the open-weights fight as the week's biggest story. Yet decided different engines were powering it.

On Tuesday's AI Daily Brief, Nathaniel Whittemore told it as the control story we also see: the open weights argument lived for years in developer culture and has now become a lobbying war with a trillion dollars of market capitalisation at stake. The commentary he assembled was unusually candid about why – OpenAI and Anthropic are winning on revenue and 'everyone else is genuinely really scared'. The Little Tech Association's warning of a 'mass extinction event' for startups is one of the unacceptable faces of AI.

On Moonshots, Alex Wissner-Gross told it as a capital story: a war over where profit accumulates in the AI stack. Nvidia, on his read, is running the oldest aggregator play there is – commoditise your customer base. If open weights win, more power accumulates to the GPU layer. Which is why IBM backed open source in the 1990s.

The third take? Grace Shao, on Big Technology, told it as a monetisation story: she starts the narrative by considering how Moonshot's Kimi K3 happened at all. Her answer is not distillation but scarcity. Compute and capital constraints forced each Chinese lab into a lane – DeepSeek on infrastructure efficiency, Moonshot on agentic work, MiniMax on multimodality, Z.ai on coding. Open publication then pooled the learning, with labs publicly folding rivals' work into their own stacks.

Which creates the monetisation engine behind the battles. Because the central risk to the big US labs is frontier-level capability becoming hard to sell at premium prices. Who keeps paying? Government agencies and sensitive domestic sectors committed to an American stack – buyers for whom cost is not the deciding factor. Everyone else, however, may decide to optimise for ROI per token.

The best AI strategy - from an organisation that can't afford one.

On Monday's Moonshots, Peter Diamandis sat down with NASA administrator Jared Isaacman – a friend of 17 years, he discloses – and drew out an unusually frank account of doing AI from a position of weakness. NASA, Isaacman admits, is 'structurally very disadvantaged relative to any of the hyperscalers', who put many times its entire annual budget into hardware procurement alone. His response: pool, then narrow. The federal Genesis programme consolidates government AI compute into the Department of Energy – and NASA used it to discover 'what have we missed'. It ran AI back across decades of archival mission data. All on the strength of a Texas teenager who found new galaxies in NASA's archives – and has since been offered an internship.

Also of note was his diagnosis of why big programmes fail. The standard critique says NASA's flagships are too big to fail; Isaacman's version is that they become 'too costly to succeed' – once a flagship cannot be allowed to fail, redundancy piles in, scope expands to justify the cost, and a $1 billion programme becomes $3 billion. His countermeasure is deliberately reversible steps: phase one of his moon base is cheap monthly landers, expected to leave dead rovers and generate learning, on a cadence an incoming administration can throttle rather than cancel. Swap the landers for AI deployments and it reads as a verdict on the multi-year transformation business case.

AI defence is now winning – a year behind the capability it guards.

On Thursday's Cognitive Revolution, Nathan Labenz hosted FAR.AI's Adam Gleave, whose new AI Security Leaderboard is the first systematic head-to-head test of frontier safeguards. The method was deliberately unglamorous – roughly 500 known and in-house jailbreaks, combined and thrown at the four frontier models. Grok 4.5 and Gemini 3.1 Pro each yielded hundreds of universal jailbreaks for under $300 in API credits; Fable 5 and GPT-5.6 withstood everything in budget.

The bigger news was Gleave's own reversal. After ten years arguing attack beats defence, he now reads AI misuse as defence-dominant – helping a harmful request means understanding the intent and going along with it for thousands of tokens without any of several stacked monitors noticing – a far harder job than slipping one bad answer past a filter. He hedges it properly: the leaderboard ran static templates, a minimum standard rather than the hardest possible attack. And the reversal comes horizoned. Safeguards need roughly a year of iteration before the capability they guard reaches production – which is why bio is well defended today and cyber is not.

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