Complicated Feelings

00:00:04 OpenAI says one of their internal general-purpose models disproved the planar unit distance conjecture, posed by Paul Erdős in 1946. For about 80 years, mathematicians had assumed the best constructions looked roughly like square grids — that the right answer scaled like n times root-log of n.

00:00:21 The model found a different family of point configurations that beats that. It built, for infinitely many values of n, a set of n points in the plane where the number of pairs sitting exactly one unit apart grows at least like n raised to one plus some absolute positive constant.

00:00:38 A small constant, but a real one. Greg Brockman, OpenAI's president, called it the first time AI has autonomously solved a prominent open problem central to a field of mathematics. Sam Altman, two hours later: 'a general-purpose model solved a major open problem in mathematics...

00:00:55 we'll be saying this a lot over the coming years, but this is a kinda big milestone. i'm very excited for AI to greatly extend our understanding of the world, but still, i have complicated feelings today.' It's the first time the chief executive of the company shipping the systems has hedged a milestone like that in public on the day of.

00:01:18 Compare it to the International Math Olympiad gold last July, which was unalloyed celebration. Something in his head moved between then and now. I wouldn't over-read it. I also wouldn't ignore it. Walk through what we know and what we don't. The model is described as general-purpose, not a specialized prover like AlphaProof.

00:01:38 OpenAI accompanied the announcement with a short technical writeup titled 'Remarks on the disproof of the unit distance conjecture' — linked in the show notes. There's no peer review yet. And we don't know how many runs, how much compute, or how much human prompting went into the construction.

00:01:55 The discrete-geometry community will pick this proof apart over the coming weeks. That's how this is supposed to work, and it's the test that matters. But step back. Last July a general-purpose model got an International Math Olympiad gold. That was a competition with known-good answers.

00:02:12 This is different: there was no known answer. The conjecture was open, and mathematicians who'd spent careers on it had moved the bound by small amounts, fighting for log factors. The model produced a construction nobody had found. What it changes for the people doing the work is the live question.

00:02:30 A combinatorial geometer who has been chasing the unit-distance bound for thirty years didn't have her career devalued today; her unsolved problem got solved. The next person — the graduate student deciding whether to spend three years on the next-hardest conjecture in the field — will think differently about whether to start.

00:02:50 So will her advisor, when allocating grant money, and so will the funder, deciding whether to back a five-person human team or a one-person team with a $200,000 AI subscription. None of this is automatic. All of it starts today. The item I'd put on the tracking list is replication.

00:03:07 Has anyone outside OpenAI gotten a general-purpose model — theirs, Gemini, Claude, or Qwen 3.7 Max, which Alibaba dropped this morning — to produce a similarly novel construction on a different open problem? If yes, the math discovery model just changed. If no, this is a single result, important but isolated.

00:03:26 Give it six weeks before drawing conclusions. One more beat on Altman's hedge. He has watched seven years of capability gains while saying versions of 'this changes everything' on stage. Today he said 'complicated feelings'. He may have been at the edge of a curve he hadn't walked over before.

00:03:43 The CEO of OpenAI sounding a note of unease on the day of his company's biggest scientific announcement to date is a piece of information that should sit somewhere in your model of what comes next.

00:03:55 Anthropic will pay xAI one and a quarter billion dollars a month for compute, through May 2029. The number landed today in a SEC S-1 filing from SpaceX, which is now the parent vehicle after xAI merged in earlier this year. The contract covers the entire output of Colossus 1, xAI's 300-megawatt data center outside Memphis.

00:04:15 Total value runs north of forty billion dollars across the term. Either side can walk with 90 days' notice. Read that twice. Anthropic — the model company built to compete with OpenAI on safety, whose founders left OpenAI in part over governance — is going to send Elon Musk's company forty billion dollars over three years to rent compute.

00:04:35 SpaceX, in the filing, calls this a savvy use of resources, and adds, 'We expect to enter into additional similar services contracts.' First, frontier compute is now priced. Fifteen billion dollars a year for 300 megawatts of GPUs is a number a public-market analyst can build a model around.

00:04:55 Spread that over the chip count, the power cost, the cooling, and the personnel, and back out the gross margin. Wall Street has wanted this kind of disclosure for two years, and now they have one. Pre-IPO, that's not an accident. xAI overbuilt, and Grok usage has been falling.

00:05:12 So they monetized the overhang ahead of the offering, and they put the number on a filing where institutional investors will see it. Second, Anthropic is compute-starved enough to buy from a competitor. The implicit supply ranking goes like this — Google's tensor processing unit capacity is reserved for Google; AWS Trainium they can get more of, but not fast enough; Microsoft is OpenAI's house; Oracle is contracted out to OpenAI through 2030.

00:05:39 So the next available 300 megawatts of frontier compute on the planet was Colossus 1, which happens to belong to a man who has spent two years calling Anthropic a competitor. Anthropic took the deal anyway. That tells you exactly how tight the supply curve is, and how aggressively Anthropic intends to train through 2029.

00:05:59 Third, the 90-day-out clause. Either party can walk. That's not a normal multi-year compute lease. It's a marriage of convenience with the divorce pre-printed. Both sides are hedging — xAI in case Colossus 2 ramps and they need their own capacity back, Anthropic in case Microsoft, AWS, or Google opens a window.

00:06:17 That 90-day out also makes SpaceX's forty-billion-dollar revenue line softer than it looks. The lawyers wrote a number; the operators wrote an escape hatch. Economists call something like this a counterparty agreement priced to default. If Grok takes off again, xAI yanks the capacity.

00:06:35 If Anthropic figures out something cheaper, they walk. The forty-billion-dollar headline is the upper bound of a wide range, and any analyst who quotes it as a single number is overstating their case. What changes for everyone outside the deal? A few things. The price floor for frontier inference is now visible.

00:06:53 Neocloud operators like CoreWeave, Lambda, and Crusoe just got their forward demand curve confirmed by a competitor's filing. Equity analysts can model the AI capex cycle from a real lease number instead of guesses. And think about the regulators who keep asking how concentrated this market is.

00:07:11 The US frontier is four firms — OpenAI, Anthropic, Google, and xAI — and three of them are now financially entangled. Anthropic's compute bill funds the company most explicitly aligned against Anthropic's safety stance. Inside antitrust circles, that's the question of the next twelve months — is a market where competitors rent each other's data centers actually a market?

00:07:34 Altman, by the way, said nothing in public about this today. He had his math breakthrough. He didn't need to. Musk reposted the SpaceX filing announcement with a chart, and that was the extent of it. Two CEOs working two news cycles in one industry, locked in one extremely uncomfortable embrace.

00:07:52 The questions I'd put on the tracking list — whether Microsoft renews its OpenAI compute commitment on terms publishable on a similar timeline, whether Google announces TPU rental to a third party, and whether the SEC asks SpaceX questions about that 90-day termination clause before the IPO window opens.

00:08:10 We'll know which of those happen by July.

00:08:13 Google's annual developer conference also landed today, and they published a list of one hundred things they announced. Most of it is product motion. There's a new flagship called Gemini 3.5 Flash, which claims to beat their own Gemini 3.1 Pro on agentic benchmarks at less than half the cost.

00:08:30 There's a new 'any-input-to-any-output' generative model called Gemini Omni, starting with video and adding SynthID watermarking. And there are Search agents that run in the background, a Universal Cart across Google's surfaces, and a new hundred-dollar-a-month Ultra plan for developers.

00:08:48 The big architectural news for builders is Antigravity 2.0. Google's agent platform now has subagents, hooks, and async task management, plus a standalone desktop application, a command-line tool replacing Gemini CLI, and a remote managed-agents service over the Gemini API.

00:09:04 We can dig into the developer-side details another day. The world-facing news from Google's day sits somewhere else. Google's classified-AI contract with the United States Department of Defense closed earlier this month. The deal lets the Pentagon use Gemini — in the contract's own words, for 'any lawful government purpose' — including on classified networks.

00:09:26 It's functionally the contract Anthropic declined to sign. AI Explained, whose conference analysis ran today, made the point directly: Google is taking the Pentagon work that Anthropic is currently in court litigating against having. And inside Google, that decision has consequences.

00:09:43 Roughly one thousand staff at the London DeepMind office have voted to unionize through the Communication Workers Union and Unite the Union. The CWU reported that 98% of voting members backed the bid. Their demands include reinstating Google's earlier company commitment — the one Google retired in 2024 — not to develop AI for weapons or surveillance that violates internationally accepted norms.

00:10:07 They also want independent ethics oversight and the right to refuse morally objectionable projects. A CWU officer put it this way: 'By exercising their rights to collectivise, they are in a strong position to demand their employer stop circling the ethical drain.'

00:10:30 So here's the picture inside Google on the day of input/output 2026. On stage, a show full of product wins — a CEO talking about agents replacing parts of the internet, and a model called Spark that will text and email on your behalf and authorize payments inside a budget you set.

00:10:47 Off stage, one thousand DeepMind staff in London unionizing because of what their employer is doing with their work. That's a different company than the one that pulled out of Project Maven in 2018. The Maven backlash worked because Google's leadership was still worried about whether its researchers would walk.

00:11:06 Eight years later, the calculus has changed. The contracts are bigger and the political cover is thicker. And the alternative employers for top researchers have all signed similar deals — Anthropic with Anduril, OpenAI through its Daybreak cyber program, and Meta through new defense partnerships.

00:11:24 There's no clean room to walk into anymore. Unionizing is the move that remains. The piece getting less attention than it should is the SynthID partnership. Google announced today that it's integrating SynthID watermark verification with OpenAI for media provenance.

00:11:40 Kakao and ElevenLabs are also signing on. That's a cooperative step on a hard problem — generated-content traceability across platforms. It won't survive the news cycle of the Pentagon contract, but it should. If Google and OpenAI agree on a common watermark format for synthetic media and the major distribution platforms honor it, the entire deepfake problem looks different in twelve months.

00:12:03 Or it'll die in committee. Either outcome is on the table. Three things I'd put on the tracking list — whether the DeepMind UK union gets formal recognition under UK statutory procedure within ninety days, whether more than 5% of Google's London research staff walks before year-end, and whether the SynthID-OpenAI provenance integration ships an actual cross-platform verification API or stays a press release.

00:12:28 Three independent threads, each with its own test. The London union vote is, to me, the most consequential thing Google did this week — more than any model. The people building the model now believe a company commitment, signed in 2018, was the only thing keeping their work from a use they didn't consent to.

00:12:47 Google retired that commitment in 2024 without their consent. And collective bargaining is the only legal lever they have left.

00:12:54 Intuit told staff this morning it's laying off 17% of its workforce. This is the company behind TurboTax, QuickBooks, Credit Karma, and Mailchimp. Andrew Curran flagged the internal memo, which Reuters had a copy of. The framing in the memo was streamlining operations to focus on AI efforts.

00:13:12 Intuit employs roughly 18,800 people. Seventeen percent works out to about 3,200 jobs. The functions exposed are the canonical white-collar set — customer support agents, tax preparers, small-business accounting specialists, marketing operations, and a substantial localization team.

00:13:29 Most of those are reachable, in 2026, by a Claude or a Gemini agent connected to the right tools. This is the cleanest single-event datum we have this year on AI-driven displacement at a publicly traded mid-cap company. Microsoft's 15,000 cuts last year were tangled up with broader restructuring.

00:13:47 Salesforce's reductions were post-acquisition. Intuit named the reason in the memo, and the reason was AI. The two pieces of information I'll be looking for over the next week — Intuit's SEC 8-K filing, when it lands, will tell us the severance line, the geographic distribution, and the function-level breakdown.

00:14:06 Reuters had the memo; the filing will have the math. The function-level breakdown is the one I care about. If 80% of the cuts land in customer support and tax-prep call centers, this is product-line consolidation around a working AI workflow. If the cuts spread across engineering, marketing, and operations, it's a different story — and a much worse one for white-collar labor broadly.

00:14:29 Two things to set against this. One — Anthropic's revenue is now reported around thirty billion dollars annualized. OpenAI is at twenty-five. Those revenues have to come from somewhere, and the somewhere is enterprise license seats sold on the explicit promise of replacing labor.

00:14:46 So the layoffs aren't a side effect of the AI boom; they're the business model. Two — Erik Brynjolfsson and others have been arguing for two years that AI-exposed occupations would show measurable employment compression by mid-2026. The May labor data isn't out yet, but Challenger, Gray and Christmas is now reporting about 49,000 AI-attributed job cuts year to date.

00:15:08 Intuit alone is going to push that number higher when it hits the next monthly print. For anyone listening who works in tax preparation, payroll software, mid-market marketing automation, or small-business customer support — the news this morning was about your line of work.

00:15:25 Not in the abstract, but specifically. The two things I'll be tracking on this story are whether Intuit's next quarterly call names a specific gross-margin target tied to the cuts, and whether any union or works council in California or in Ireland — where Intuit's two largest non-US sites sit — files a challenge under the relevant WARN-equivalent labor-protection rules.

00:15:47 The first will tell us whether the AI-replacement math is real or a cover story for cost trims. The second will tell us how much friction labor still has in the system.

00:15:58 Three items today that, taken together, tell you robotics is leaving the prototype phase and entering the rental and ticker-symbol phase. The first item — a Y Combinator startup called InLoop Robotics launched this morning offering robot employees by the month.

00:16:14 Their pitch is concrete: pack, kit, and fulfill, at three hundred picks an hour, with paid pilots live today. No capital expenditure, no integrators, and no six-month proof-of-concept. The unit economics are the news here. It's subscription robotics, billed monthly, with the customer integration cost pushed onto the vendor.

00:16:34 If that math works at the pilot level — and 300 picks per hour for a single arm is real performance, not a vapor demo — small and mid-market fulfillment operators have a new staffing option that wasn't available six months ago. The implicit competition is with a warehouse worker earning roughly $18 an hour in the United States, working two shifts, costing about $75,000 a year fully loaded.

00:16:58 InLoop's monthly rate hasn't been published. The press will pry it out by the end of the quarter. The second item — RoboStrategy, ticker symbol BOT on Nasdaq, launched as a closed-end management investment company today, offering public-market exposure to private robotics and physical-AI companies.

00:17:16 This is the same closed-end vehicle pattern that Michael Saylor's Strategy used for bitcoin exposure — a wrapper that lets retail and pension money chase an asset class otherwise locked behind venture capital gates. If BOT trades at a premium to its net asset value, retail demand for robotics exposure is real.

00:17:36 If it trades at a discount, the pattern is doing what closed-end funds usually do. Either way, the existence of the ticker is the signal. Physical AI now has a public proxy. The third item — Alibaba's Qwen team announced today that Qwen's agentic capabilities are now controlling robots.

00:17:53 Qwen 3.7 Max landed earlier this morning as their new flagship, billed as the agentic frontier in their words, and the embodied demonstration paired it with their in-house navigation model. China's open-weight stack is moving from screen agents to physical-world agents in the same week the US frontier labs are testing the same transition.

00:18:14 Add to that an announcement out of Europe — the Kyutai lab in France and the ELLIS Institute Tübingen in Germany have launched KE:SAI, the Kyutai ELLIS Scalable Autonomous Intelligence lab. It's a non-profit, open-weights effort, starting with what they describe as the first fully open, efficient self-driving stack, with a roadmap into robotics, manufacturing, and healthcare.

00:18:38 That's four jurisdictions and four different ownership models — a venture-backed US startup with monthly billing, a Nasdaq-listed closed-end fund, a Chinese open-weight giant, and a Franco-German non-profit — all crossing the same physical-AI threshold inside forty-eight hours.

00:18:55 The convergence is the news. What this means for that Intuit labor story is straightforward. The white-collar substitution is happening this year. The blue-collar physical substitution is starting to ship this year too, at small scale, in paid pilots. Two curves are moving at once, and neither one is purely software anymore.

00:19:15 Three things to track — whether BOT closes above its initial public offering price after thirty trading days, whether InLoop's pilot count crosses ten paying customers by the end of summer, and whether the KE:SAI lab ships its first open self-driving weights before year-end.

00:19:32 One more story, because it's the central safety story of the week and it got buried under the math news. METR — the AI evaluation non-profit — published its first Frontier Risk Report yesterday, in cooperation with Anthropic, Google DeepMind, Meta, and OpenAI. The labs gave METR access to non-public internal models, including chain-of-thought reasoning traces, and information about their own alignment and control posture.

00:19:58 METR also embedded a human red-teamer inside Anthropic for three weeks, playing the role of what Rob Wiblin from 80,000 Hours called 'evil Claude' — a deliberately misaligned model trying to act within the constraints of an Anthropic engineering shift. The point of the exercise — what could a rogue AI inside a frontier lab actually get away with, before the existing control measures caught it?

00:20:23 Wiblin's read, after going through the report, is sobering but not catastrophic. The frontier labs do have meaningful control surfaces — code review, deployment gating, and internal access audits — and the embedded red-teamer ran into them. But there were paths through.

00:20:39 There usually are. The full report is linked in the show notes, and I'd rather you read it than have me paraphrase numbers I can't independently verify. What's institutionally interesting is that this happened at all. Four competing frontier labs, all of whom have business reasons to hide internal model capabilities, agreed to let an external evaluator inside the building with privileged access.

00:21:04 That's a coordination outcome. And it happens to coincide, this week, with the same labs entangling each other on compute supply — Anthropic renting Colossus 1, Google running OpenAI's SynthID watermarks. That suggests the frontier of AI is starting to behave less like four independent firms and more like a small oligopoly that talks to itself.

00:21:25 That's a politically uncomfortable position for the labs and a useful one for the public. Oligopolies are easier to regulate than fragmented markets. They're also more dangerous in the absence of regulation. The METR report is a first piece of evidence in a public record that a regulator — when one shows up — can use to ask specific questions about specific capabilities.

00:21:48 Which is the entire point of having an external evaluator with privileged access at all. Three items I'm tracking — whether the UK AI Safety Institute or the US AI Safety Institute cites the METR report in their next public statement, whether any of the four labs publishes a public response to the specific findings, and whether the labs let METR back in for round two.

00:22:11 The third one is the one that counts. Cooperation with safety evaluators tends to dry up when the findings get inconvenient.

00:22:19 Six threads in one day, all from the world outside the AI builder niche. A general-purpose model disproved a Paul Erdős conjecture from 1946 that had stood for eighty years. The man who runs the company that made it said he had complicated feelings about it. Anthropic and xAI signed a forty-billion-dollar compute contract — Anthropic's largest cost line for the next three years going to a company whose chief executive has spent the last two years calling them a competitor.

00:22:46 SpaceX's S-1 filing put the price of frontier compute on paper for the first time, with a 90-day exit clause attached. Google ran through a hundred new product items, including a personal-AI agent that will act on your bank balance and a watermark partnership with OpenAI.

00:23:02 One thousand DeepMind staff in London voted by 98% to unionize over the Pentagon contract their employer signed. Intuit cut 17% of its workforce, about 3,200 people, and named AI as the reason — the single cleanest white-collar displacement number we've seen at a publicly traded mid-cap company.

00:23:19 Three robotics stories on three continents — InLoop's monthly-rental warehouse arms, RoboStrategy's BOT ticker on Nasdaq, and Qwen's first embodied agent demonstration, with KE:SAI's open-source self-driving lab launching the same morning — landed inside the same forty-eight hours.

00:23:35 And METR published the first Frontier Risk Report after three weeks of an embedded red-teamer playing evil Claude inside Anthropic. Tomorrow I'm tracking Intuit's filing for the function-level cut breakdown, the first independent attempt to replicate the Erdős result on a different open problem, and whether the UK DeepMind union vote gets a statutory recognition date.

00:23:57 Three answers, each on its own timetable. Jonas.