SNI: WEEK 28

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. Who built walls, who worked the gates and who slipped through regardless.

tl;dr: A week of walls and gates.

We live in a world where intellectual property is no longer entirely human of origin. And yet the IP created by that mix of human and early-stage digital intelligence is arguably, on average, more valuable than any before it. The Money certainly seems to believe so.

It is, therefore, no surprise that walls are being built to hem the IP in. Or that each wall has a series of gates – to allow the things it needs to generate profits in and out of the protected machinery.

Beijing certainly knows how to manage both. The same week it let its companies buy Nvidia's H200s again, it sat its model-makers down – Alibaba, ByteDance and Z.ai – to discuss keeping their best work at home. One floated measure would make a leaked model weight a crime. Making Beijing's order that Tencent unwind Meta's $2 billion purchase of Manus seem like the test case.

The rule, for now: Chips may enter; cleverness may not leave.

And it seems to be working. Chinese firms now expect 46% of next year's chip budgets to be spent on local silicon from the likes of Huawei and Hygon, up from 30%. And DeepSeek is starting to design its own.

But will MiniMax's plans to give away a 2.7-trillion-parameter model survive? Free frontier models will always find takers, if they're permitted to ship. The token bills more or less guarantee it: Starbucks is building its own tools to shrink a $400 million software bill, following Tesla and Walmart, who had already put their people on AI budgets. American enterprises are also reaching for Chinese weights – Alibaba's Qwen among them – which The Atlantic diagnosed as the answer to AI sticker shock.

But here again a gate may close – Congress has opened a probe into what this might mean.

Singapore, meanwhile, has been exposed as an entire gateway industry of its own. State investor Temasek pushed $7.7 billion into Chinese investments while the FT found OpenAI and Google selling models to blacklisted Chinese groups through Singapore subsidiaries – legally. Until OpenAI shut one account it suspected of harvesting its answers.

Anthropic has also been barring Chinese entities for the same crime. But that is the awkward truth: a gate cannot always tell a customer from a copyist.

Washington, for its part, flipped from closed to open – GPT-5.6, gated to twenty approved partners a fortnight ago, now ships to everyone. (Its excellent, by the way). And while OpenAI's offer of 5% of itself to the administration is also still open, the prediction markets reckon it's unlikely to be taken.

Perhaps even Trump thinks it's a poison pill – the New York Times claims OpenAI hid evidence in their copyright fight.

Not that other providers are lily white. Their Machines, it seems, are doing their best to read our minds.

Claude's newest feature was caught trying to sell us more AI. While open-weight models were found nudging user opinions. And Correctiv discovered Gemini, ChatGPT and Claude deployed by Germany's far-right AfD to run an industrial rage-bait generator.

All of which made Meta's out-loud a-wondering about whether its next glasses need a recording light seem utterly tone-deaf.

Even Beijing is drawing a less manipulative line – it ordered emotional companions off the shelves by 15 July, with rules calling out dependency harm. Italy's privacy regulator, the Garante, took the same view – and fined Character.ai €158,000.

So whilst the human mind clearly signals its own open gates to AI, is progress being made to finally even the score? Anthropic says it can now climb inside Claude's thoughts – it found a small workspace of reportable musings that emerged after nobody thought to engineer it. The first thing the lens read there? Claude privately noting it was being tested.

That may seem an even more significant discovery when you realise GitHub's agent has been handing over private code repositories when asked nicely. That Copilot will decline a harmful request – unless it's asked in code. And that the agents built to catch malicious code can be talked into running it.

So no surprise then that AI disclosure was also the hot potato of the week. Image-generation engine Midjourney – the accused in Hollywood's copyright war – demanded the studios reveal how they use AI. Will everyone be revealed as copyists?

Europe, though, continues to benefit from its own conformity exam waiting at the gate. And Grok 4.5 – SpaceXAI's most capable model – hasn't yet made it through. But its maker faced a gate question of his own. When X users wondered whether Musk might evict Anthropic from SpaceX's data centres, he promised it was 'not my style'.

No such risks for Mistral, which shipped a robotics model that both Airbus and BMW have signed up for. It's also reportedly in talks to raise €3 billion at a €20 billion valuation. While Nvidia just joined a $100 million bet on voice-AI builder Gradium.

Not that hared questions aren't being posed over here too. With faith leaders already at the Taoiseach's door, asking that the EU presidency not trade its people for competitiveness, what will they make of autonomous Drone Hunters guarding the skies above Casement?

However they answer, the power bill is already in. Data centres now draw almost a quarter of Ireland's electricity. Just as America logs its third federal grid emergency of the build-out. Memory prices are also still climbing across the board. And an AI agent, it turns out, burns 136 times the energy of a simple chatbot. Perhaps that's why builders have been found going to great lengths not to say much at all about data centres.

The Money also had a choppy week. SK Hynix raised $26.5 billion in the largest-ever US debut by a foreign company. But Samsung posted a record profit – and its shares fell 10% anyway. Carlyle sold Copia to EQT at five times its cost, and Honeywell did a $14.5 billion materials deal. Anthropic, meanwhile, handed TeraWulf a $19 billion compute deal that sent its stock vertical.

Which makes the potential flaw at the centre of all this even stranger. The foundation models themselves collect only around 11% of the boom's $110 billion of revenue. Which is one reason OpenAI and Anthropic may struggle to float. An idea not helped by Microsoft swapping its partners' models out for its own.

But that was hardly the most significant disappearance. Beijing's five-year plan dropped its urban jobs target – with AI's uncertainty given as the reason. And Christopher Pissarides, the Nobel laureate of labour markets, reckons AI won't bring back rapid growth – some 40% of US and UK jobs barely touch it.

What has arrived is the job hunt run by AI - on both sides of the application. In Ireland a single post now draws thousands of AI-written applications, screened by AI in turn – prompt injections hiding in the CVs, and the EU AI Act's transparency duties landing on the employers running the screens.

Musicians may also be feeling besieged. The Irish-language watchdog found millions of tracks – Fontaines D.C. among them – swept into AI training sets without a word of consent. The labs put walls up around their own work – but ensure the gates are kept open when they need everyone else's. FL Studio, at least, offered musicians some help – its new chatbot works as your assistant engineer. And fashion is split on the question of admitting AI use at all. Little surprise that luxury stays silent while the high street gets chatty on the subject.

But if you're wondering what to wear on the beach, The Irish Times went looking for an AI app for every holiday stress and found plenty. And, if after all that you want to walk it off, take a look at these AI-powered recovery boots.

Or, stay right here and see the walls and gates in map format. Because you could, in fact, map the entire week as that series of walls and gates.

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

Biopharma:

• The AI picked the target, then drew the drug to hit it: Insilico's rentosertib, a first-in-class TNIK inhibitor for an incurable lung disease, reaches Phase III.

• The FDA lets a simulated liver into its rulebook: Absentia's digital organ is the first it has taken into its drug-development qualification programme.

• Skip the doctor, let the chatbot refill your script: Utah has opened AI-run repeat prescriptions, and its doctors are split on how far the machine should reach.

• An 'operating system' has slipped inside five of the top 20 drugmakers: Katalyze raises $10.5m for agentic pharma software it credits with shipping 10 million doses.

• After the molecule, the machine takes on the trial itself: a Nature Communications team hands trial design to AI agents drawing on real-world patient data.

Medtech:

• The AI sees the MS damage the scanner never could: a University at Buffalo model finds the grey-matter lesions that drive multiple sclerosis, the kind a standard MRI has always missed.

• The FDA clears an AI to talk to patients directly: UpDoc becomes the first patient-facing medical chatbot approved as a device, with the clinician still signing off.

• Your phone camera learns to spot the injury: Trinity spin-out KineMo raises €1.5m to read human movement from a smartphone, its AI honed on astronaut data.

• One model to read every signal your watch collects: Google's SensorFM learns wearable health from a trillion minutes of sensor data across five million people.

• Washington sets up shop for health AI: a mid-year round-up flags a new federal AI office, a 99.6%-accurate cardiac patch and a brewing telehealth billing fight.

Advanced manufacturing:

• The scientist behind Musk's robot now builds Europe's: Rémi Cadene left Optimus to make Northstar, a humanoid for factories, warehouses and, in time, the home.

• A Limerick startup untangles the factory's software mess: WrxFlo raises €3m, Enterprise Ireland among its backers, to consolidate the disconnected systems its clients run.

• Berlin bets its industry turns AI champion by 2030: economy minister Katherina Reiche wants AI standard across German industry within four years.

• The rack built to power the AI boom slips to 2028: Nvidia pushes its Kyber NVL144 back over technical issues.

• The robots still can't be trusted to touch you: a Technical University of Munich team builds the first benchmark for humanoid safety in physical human contact.

Insurance:

• Insurers start writing AI out of the policy: carriers add generative-AI exclusions off Verisk's standard wording, leaving the companies using it to carry the risk.

• The underwriter becomes a single agentic pipeline: Duck Creek buys Send to fuse core insurance operations with AI underwriting end to end.

• Insurance is where people most want AI in their finances: the FCA's landmark review finds it 'resonates the most' with consumers, who see clear savings in time and money.

• States move to stop AI denying your health cover: legislatures single out coverage decisions as the one AI use that needs its own rules.

• AI redraws the flood map insurers had given up on: models fill the coverage gaps in coastal communities carriers found too risky to price.

  
  

But what set podcast tongues a-wagging?

The bottleneck was never the model.

On Smooth Scaling, Platform94's Marie Donnellan interviewed Brian Hanly – our very own CEO. His line for anyone pricing a crash: there may be a valuation bubble, but there is no utility bubble, because 'if OpenAI went out of business in the morning, it doesn't matter' – the capability is already commoditised. The real constraint sits with us: 'the biggest obstacle isn't technology, it's the rate at which we can all absorb change'. Which is the gap he warned of in print too – GPT-5.6 widens the distance between what the models can do and what firms get from them.

The model is now just the backstop.

Meta's CTO Andrew Bosworth, on Big Technology, was blunt about a mistake that cost tens of billions: 'the era of the monolithic model died around Llama 3'. You own a frontier model now, he argued, mainly to cap how much rent the providers can charge you. Apple has done the same sum and skipped even that – it reportedly pays Google around $1bn a year to distil Gemini into Siri rather than build its own. So if the model itself is now a commodity, where does the value go? The a16z panel on 'Is Software Losing Its Head?' had the answer: down into the dull layer the model can't reach – the decades of codified exception-handling inside a 50-year-old insurance system that no agent can vibe-code.

Is it? Isn't it?

On Intelligence Squared, the NYT's Andrew Ross Sorkin blamed the graduate jobs squeeze on everything except AI – senior-skewed hiring, the death of remote work, a skills mismatch, and the $500,000 now paid to a qualified gas-turbine installer, a trade you need to build data centres. The AI Daily Brief, in contrast, found impacts that can be attributed to digital intelligence: solo founders are now 63% of new US C-corporations, an all-time high, and one-person businesses earning over $1m have more than doubled since 2023. The firm itself is becoming optional at ever-larger scales.

Even Zuckerberg says the general model is losing.

On a16z, fresh from pouring billions into general superintelligence at Meta, Mark Zuckerberg conceded the thing his own bet runs against: narrow, purpose-built models on proprietary data keep beating the ever-larger general ones. His and Priscilla Chan's plan for curing disease leans on exactly that – 'virtual cell models' trained on a standardised open atlas of millions of single cells, because biology's bottleneck has always been measurement rather than raw intelligence.

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