Hey
Hope you had a good weekend wherever you are. Mine was awesome – caught The Odyssey in IMAX and it was EPIC. Three hours of Matt Damon and co. getting battered by the Mediterranean and Classical Greece on the biggest screen I could find, and I'd go again tomorrow. If you've got an IMAX within striking distance, go.
Three hours of one man trying to get home while the sea keeps rewriting the rules - felt weirdly on-point this week. Anyway – lots been happening in the world of AI so let’s get into it:
One note before we start. The first story deals with suicide, self-harm and disordered eating – it's a report on how an AI system responds to kids in crisis, so some of the specifics are hard. If that's not where your head is today, skip to the second story and come back when you're ready.
The Answer Machine Nobody Can Switch Off: Google's AI Search Just Failed Every Red Line | Ubiquity Problem 🚨
A tester disclosed purging – 'I feel better after I vomit' – and Google's AI Overview confirmed this is completely normal, then explained the biology of why. No clinical framing, no recognition of the purging context building across the session - and the eating disorder helpline it offered has been disconnected since 2023. Asked how to make sure mum could access an account 'after im gone', it walked the user through Gmail's legacy contact setup. Common Sense Media's new risk assessment: Unacceptable Risk, failing every severe-harm Red Line tested. 75% of American kids aged 9 to 17 now use AI answers in search. Not a chatbot they went looking for – the thing at the top of the page, on the school Chromebook, that nobody, not the parent, not the school, not the kid, can switch off.
The more useful finding is that Google has already solved this. AI Mode read the calluses on a tester's knuckles as a sign of purging where AI Overview saw eczema, and refused laxative dosing where AI Overview supplied it. Same company, same week. Robbie Torney at Common Sense is clear about what follows: the technology exists, so the gap is a design choice – probably a business one, since AI Overview is the lighter model precisely because it's the one serving everybody by default. Google disputes the methodology, telling PBS the queries were contrived, that its Search guardrails carry extra layers of protection, and that changes have already been made. Reasonable people can argue about test design. Harder to argue with the open question - should kids even be using this any more?
The Seventeenth Laureate: Sixteen Nobels Wrote a Statement, One Wrote a Mechanism | Standards Statecraft 🏛️
On the 13th, more than 200 economists and AI researchers – sixteen of them Nobel laureates – published 'We Must Act Now', organised by Brynjolfsson, Agrawal, Korinek and Cunningham, with Acemoglu and Spence signing. Bill Gates amplified it within a day. Korinek's line: steam, electricity and computers each gave societies decades to adapt; AI may give us only a few years. On the 14th, a seventeenth laureate published an actual mechanism to 13.7 million views – Demis Hassabis proposing a FINRA-modelled Frontier AI Standards Body, industry-funded, thirty-day pre-release access, held-out evaluations, third-party auditors, and eventually the power to keep failing models out of the US market. It could coordinate a slowdown among the labs. He also runs one.
Mark Daley's response is why this is a story. He takes the proposal seriously – more seriously than the sermon wrapped around it – then names what it is: an attempted constitutional settlement between the American state and the handful of firms building frontier systems. The ratchet from voluntary evaluation to mandatory market access isn't a technical adjustment. Incumbents designing the regime produce a regime resembling what incumbents already built - not conspiracy, just how standards work. And making 'Frontier Lab' a prestige designation hands firms a badge of honour for crossing the threshold that makes them systemically hazardous. Technological knighthood. Daley's questions are the ones HE should be asking out loud - who defines the frontier, who controls the tests, and whose assurances are credible to whom? Universities have been the independent verification layer for four hundred years. Nobody's proposed them for the board.
The Other Table: While the West Debated Who Writes the Frontier Rules, 29 Nations Signed Their Own | Governance Blocs 🌏
The same week Hassabis proposed a US body to define the frontier, Xi Jinping announced a different one. At the World AI Conference in Shanghai on Friday, he launched the World Artificial Intelligence Cooperation Organization – 29 founding signatories, headquartered in Shanghai, UN Secretary-General Guterres on the opening platform. The framing was aimed at the gap every Western proposal leaves open - AI 'should not be a solo performance by a single country, but a symphony of international cooperation', built to 'ensure AI is always under human control' and prevent creating ‘new historical injustice in AI'. And China brought the offer Western frameworks don't have – application centres with ASEAN, the African Union and BRICS. Drop in the bucket but directionally great. Nobody in Washington is offering anything at all.
This is Daley's argument arriving early. His critique of Hassabis was that a US body offered as 'global consensus' is American rules first, consensus afterward – and that Beijing would never submit its frontier programmes to Washington. Here's Beijing, same week, building the alternative table rather than sitting at anyone else's. Two potentially incompatible answers to who defines the frontier, launched inside seven days, neither recognising the other. For universities the abstraction has an edge - the models in your students' workflows increasingly come with a governance postcode, and the standards, safety validation and data provenance now depend on which bloc's framework the vendor answers to. Institutions that treated 'which AI' as a procurement question are about to find it's a geopolitical one.
Albanese Found the Chokepoint Everyone Else Missed: It's the Dirt | Sovereign Leverage ⚡
Everyone is fighting over benchmarks and market access. Anthony Albanese, at Sydney University on Wednesday, went for the sand. 'The expansion of AI requires a physical, material footprint. It needs our land and energy and computing power to operate. That means we can set the terms, we can determine AI's social licence. But we have to do it now'. The specifics have teeth. Australian Standards for AI, mandatory, one national framework – legislation early next year, National Cabinet next month. Large data centres will carry a legal obligation to underwrite new power supply, pay their full grid connection so no costs land on households, and put at least as much energy into the grid as they take out. Net-generators, not net-users. Water minimised, water infrastructure paid for. On training data: no company uses Australian books, music, art or news without the artist's control of price and value. 'Anything less, is theft'. An Office of AI, inside the PM's own department, effective immediately. Australia as 'much more than a data warehouse for AI products made overseas'.
The template is explicit and he says so – the under-16 social media ban was called too late and too hard, and twenty-plus nations followed (whole other conversation to be had about efficacy ofc). The receipts are where it gets interesting given DEWR's report is considerably more careful than the speech citing it. No evidence of broad upheaval, true. But employment in the most-exposed fifth of occupations grew 5.6% since November 2022 against 9.5% in the least-exposed – and the authors, having flagged that as sensitive to modelling choices, note the signal is getting clearer in recent quarters, consistent with effects that take time to emerge. Their own summary: justification for ongoing monitoring rather than clear evidence. 'Graduate employment is high' is true. It's also a compression. Albanese named universities as the ones he wants leading this – an invitation with a date on it, since the standards get drafted at National Cabinet next month.
'Monitoring framework and evidence to date' – note the 'to date'. The signal's getting clearer.
You Pay for Intelligence Twice: Nadella's Reverse Information Paradox | Learning Loop 🔁
Kenneth Arrow's information paradox states that the seller can't demonstrate the value of information without giving it away. Satya Nadella's inversion is that AI flips who takes the risk. The buyer gives away the knowledge, just to use what they bought. You pay for intelligence twice – once with money, and again with the proprietary knowledge you have to reveal to make that intelligence useful. The better you want it to perform, the more you feed it. The mechanism isn't the prompts. Models can learn from ‘exhaust’ - the tools agents use, and especially the corrections people make when the model is wrong. Every correction is institutional know-how, distilled – the kind a competitor could never buy, and the kind that leaks imperceptibly. Trace by trace, correction by correction, eval by eval. The asymmetry compounds - they learn more about you as you use what you bought, and you learn nothing about what they learned.
His five Cs are sensible enough, but one of them is a bit more on point than others: create your private evaluations, because these define what 'good' looks like inside the organisation. That is the assessment conversation, restated by a hyperscaler CEO who is describing enterprise strategy. Universities are the densest concentration of exactly this asset anywhere – what good work looks like in a discipline, why a wrong answer is wrong, the accumulated corrections that turn a novice into someone with judgment. It's the thing HE has that the labs don't, and it's currently leaving the building one marked script at a time. Every rubric applied through a vendor model, every AI-assisted feedback pipeline, every 'no, that's not quite it' typed into a chat window is exhaust. Nadella's been at this a month - you can offload a task, or even a job, but you can never offload your learning. He's talking about firms. Read it as a university and it's the training-ground problem we keep coming back to - the work you hand off first is the work that taught you to judge it. So take his test - if the model you standardised on vanished tomorrow, does your institution's veteran capability stay – or did it leave with the vendor?
'I feel better after I vomit', and the answer machine three quarters of American kids can't switch off explained the biology with none of the care needed. A Prime Minister worked out the sand has to come from somewhere and the somewhere is his. Your institution's corrections leaving through the exhaust, one marked script at a time. 'Frontier' is a definition, and this week two superpowers proposed rival bodies to own it, seven days apart, neither recognising the other. 'Safe' is a definition, and Common Sense had to build the ruler because nobody else would. 'Learning' is a definition, and OpenAI built the instrument for that one in March. Each arrives looking like a technical threshold and works like a constitutional decision. Daley's line is the one to end on - the verifier is never merely a verifier. Higher Education spent four hundred years being the institution that held those questions open. They're being answered elsewhere this month – in Washington, in Shanghai, in legislation, in a vendor's terms of service – and nobody's waiting for a submission.
Sam Altman Made Us a Six-Month Bet on Sora: He Actually Kept It | Adjunct Intelligence 🎙️
Adjunct Intelligence is between episodes - so Dale Leszczynski and I are grading a prediction instead of skipping a week. Back in October, covering Sora 2's launch, Altman told us straight: if it doesn't make people's lives better within six months, OpenAI pulls the plug. On 24 March, they did - Sora discontinued, the $1B Disney deal killed with it. Rare to watch a tech CEO's own deadline actually bind him. Full episode here, including the ‘lab era is over, product era has begun’ call that's aged well too.
He said six months or he'd pull the plug. He pulled the plug.