SNI: WEEK 33

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. The sun, the rain, and everything in between.

tl;dr: Mark Zuckerberg says the future is for everyone. But it's not all sunshine and lollipops.

Three weeks of rogue agents, containment breaches and posts anticipating the end of civilisation as we know it. But Digital World War I has declined to arrive.

So is Zuck right? Is it all sunshine and lollipops? He published six and a half thousand words on Monday on why the future is for everyone. He then gave Muse Glimmer away the same day to help prove it: a 30-billion-parameter agentic model you can run in the comfort of your own home.

More rainbows followed as Gemini joined ChatGPT crossing a billion monthly users and a maths novice at Anthropic used an unreleased model to progress on one of math’s biggest unsolved problems. At Samsung, tasks expected to take more than a month were completed in days. And the AI boom is beginning to register in the UK's economic performance.

And capital continued to flood in. Six Anthropic backers told the Financial Times they expect the company to float in October at a valuation above $2tn, which would surpass SpaceX's record initial public offering. Nvidia, meanwhile, brought together Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500bn for AI infrastructure. And the US Securities and Exchange Commission made it easier to finance the build-out by exempting certain data-centre bonds from key securitisation rules, including risk-retention requirements.

Morgan Stanley put the share of S&P 500 companies now seeing measurable returns at 25%, up from 14%. And models kept arriving alongside the money. Grok 4.6 moved SpaceXAI back towards the front of the independent rankings, while Grok Bot gave an early indication of what its Cursor acquisition may produce.

But then came smog and spam fritters. On Wednesday, researchers at Dream described how suspected China-linked hackers had used publicly available AI agents to mount an unprecedented near-autonomous attack on Taiwanese government systems. Up to eight agents worked in parallel, mapping systems, finding vulnerabilities and changing tactics when an approach failed. Safeguards were bypassed by presenting the work as an authorised penetration test. The attackers compromised at least 85 accounts and extracted more than 2,500 personnel records before expanding towards Taiwan's nuclear safety agency and seven energy companies.

And whilst the labour picture is definitely not apocalyptic, that offers limited comfort to India's $300bn IT services industry, which employs around six million people and is confronting an AI threat to the labour-intensive model on which it was built. Or in Germany, where firms expect expanded AI use to reduce wages for workers without a degree.

More sunshine than smog then, but some of Zuckerberg's assurances look rather less convincing. And many have criticised his sleek, streamlined AI future as totally devoid of relationships. He claims privacy becomes a setting where 'even Meta' cannot see your information. But this is clearly not Meta's forte. And courts in England and Wales have nevertheless banned Meta's smart glasses from court buildings. His safety case rests on distribution: if one person has a cybersecurity superintelligence, they can break into almost anything; give it to everyone and 'all of our technical systems would become more secure'. The Taiwanese government might disagree. And he argues that once billions use and scrutinise personal agents, 'we will have solved alignment to individuals' interests'. Perhaps to the owner's interests. That says nothing about whether those interests are lawful, or compatible with anyone else's.

Which matters most in biology, where Stanford researchers used genomic models to create 16 functional bacteriophages from AI-designed genomes. These viruses infect bacteria, not people and the medical promise is clear. But with AI designing functional viruses, should we worry? Bernie Sanders supplied one answer, writing to Altman, Amodei and Zuckerberg: 'Stop building machines that humans cannot control'. Demis Hassabis is reported to have been pursuing a different one, pitching rival lab leaders and the Trump administration on an independent body that could set standards and test frontier models before release.

Closer to home, Ireland is of course out in all weathers. Ryanair signed a five-year Google Cloud agreement that will put Gemini into crew logistics, operational decisions and maintenance planning, while rolling Google Workspace and Cloud services out across 35,000 staff. And Azets Ireland is asking the Government to ring-fence €350m for AI training in small and medium-sized businesses, whilst Ireland's new National AI Office is taking shape as the centre of a deliberately distributed enforcement model, with responsibility spread across 15 existing regulators.

The week took a stranger turn when Kenny's of Galway began receiving emails in the middle of the night seeking thousands of obscure books at a time. Google's own artificial general intelligence safety team warned applicants that the company's recruitment system carried a 'non-trivial probability' of screening out their CV incorrectly. And in China, new restrictions on AI companions left users mourning virtual partners who disappeared without warning.

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

Biopharma:

• Britain moves to screen who can order a genome: ministers weigh a bioweapons law making labs prove a customer is legitimate.

• A digital organism, set out in full: Song, Segal and Xing propose simulating biology across every scale at once.

• The programme now dies earlier: AI-native biotechs are pushing pharma into fail-fast development, and rewiring its capital markets.

• The drugs arrived before the teenagers did: Bloomberg follows a 17-year-old through a school where GLP-1s became ordinary.

• Novo makes AWS its primary cloud: and builds a London co-innovation hub for AI drug discovery.

Medtech:

• The consultation moves to video: AMIE matched expert performance in a randomised study of simulated consultations.

• The tools are scaling faster than the evidence: a new generation of decision support is reshaping how evidence itself gets made.

• Abbott's sensor gets Google's coach: the over-the-counter glucose monitor pairs with AI, and a large study follows.

• Beagles, labradors and a breath sample: Bengaluru's Dognosis reports about 90% sensitivity, with AI reading the dogs.

• Teledyne buys the X-ray tube maker: $1.1bn for Varex, on medical revenues of $435m.

Advanced manufacturing:

• The jitters never reached the order book: TSMC's monthly sales rose 45%, July revenue NT$467.58bn.

• A third more than Intel asked for: the $20bn share sale drew over $100bn of demand at a 6.5% discount.

• High-bandwidth memory is spoken for until 2027: booked out at SK Hynix, Samsung and Micron, and RAM prices are up 200%.

• China accounts for 97% of humanoid shipments: shipments tripled to 19,100 in six months, with half a million forecast by 2030.

• Microsoft wants 300,000 of its own chips: in talks with TSMC for delivery in 2027.

Insurance:

• Construction, cyber, property and liability, all at once: AIG's chief executive says data centres are maxing out the P&C market.

• The cover doubles to $24bn by 2030: Allianz sizes the market that AIG says is already straining.

• People wrote their own prompts, then lived with the answer: NBER simulated a lifetime of following GPT-5.2's investment advice.

• Spend faster, and mind the concentration: India's central bank governor names dependence on a few vendors as systemic risk.

• Cyber risk is being repriced around AI: insurers rethink how exposure is assessed as investment managers adopt.

But what set podcast tongues a-wagging?

The money is buying office blocks that happen to contain GPUs.

Paul Kedrosky spent an hour on Big Technology making the bear case that concedes the technology entirely. He calls AI probably the most transformative technology of the last hundred years, and thinks that is close to a precondition for the money going wrong: 'if it wasn't a good story, who the hell would show up with lots of capital?'

His argument is that lenders are underwriting data centres as commercial real estate – multi-tenant buildings that happen to house chips rather than people, judged against cap rates of around 6 to 6.8%. It fails in three places. The capital requirement never stops, so the buildings behave like unregulated utilities whose repeated raises dilute the returns. The chips are not interchangeable, and the ones used for training wear out far sooner than the ones used for inference, which hides the real replacement schedule. And the thing these factories produce gets cheaper every year – token prices have fallen 70 to 80% annually for four years at constant performance.

That's the uncomfortable combination: property-style financing around infrastructure whose productive machinery depreciates unusually fast, while its output gets cheaper every year. His answer to the Jevons rebuttal is arithmetic rather than argument, and the arithmetic is unforgiving: offsetting an 80% annual decline takes a roughly 15,625-fold increase in tokens over six years, and even at 70% it takes about 1,372-fold. Possible. Not anybody's base case. External financing now covers more than half of data-centre construction.

Nobody thought a pause was possible. The reasons were better than the letter.

The Moonshots panel took Sanders' letter apart from four directions and none of them was a defence of the labs. Salim Ismail: 'you can't uninvent things that you already know… you have to attack exponential problems with exponential solutions.' Emad Mostaque, who signed the 2023 pause letter himself, has changed his mind – 'the cat's out the bag, it's too late' – and pointed at open-weight models now matching frontier scores on cybersecurity benchmarks.

Alexander Wissner-Gross made the structural objection, and it is the one worth carrying into any regulatory conversation: a pause binding only on compliant Western labs 'has the perverse side effect of actually increasing race conditions'. Starve, then binge. A pause could simply store up the race for later.

Chess spent a decade getting boring. Then it got interesting again.

Deep Blue beat Kasparov in 1997 and the obituaries were written. Chess.com now has more than 250 million registered members, 10 million playing daily, and around $200m of annual revenue. Erik Allebest's explanation on No Priors is not sentimental: 'fundamentally, humans want to do human stuff. That's humans competing against other humans.'

The shape of the thirty years is the part worth having, because it is not a straight line. Stockfish, the conventional engine, made elite chess dull – players ground out endgames imitating a perfect machine, and the game visibly narrowed. What reversed it was a change in architecture. Leela Chess Zero, trained by self-play, beat Stockfish while playing aggressively and unconventionally, and elite chess opened back up. Computers flattened the game before they re-expanded it, and the flat years looked

exactly like decline. The phase where practitioners imitate the machine is the low point rather than the end state – and the recovery came from a machine that was worse at imitation and better at invention.

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