ALETH / AI-BRIEF / 2026-05-23 / AI MODEL DISPROVES 80-YR-OLD MATHS CONJECTUREThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
OpenAI said a reasoning model disproved a conjecture Erdős posed in 1946.
Anthropic projects a $559m Q2 operating profit, its first profitable quarter.
SpaceX’s IPO filing showed xAI, including X, lost $6.4bn in 2025.
A jury found Elon Musk sued OpenAI too late, and he plans to appeal.
Google made Gemini 3.5 Flash the default for >900m Gemini app users.
Browse by day:
Weekend: Malta ChatGPT Plus
Monday: Musk v Altman · Anthropic-FSB · Stainless · Cursor · Google-Blackstone
Tuesday: Karpathy · Gemini 3.5 Flash · Gemini Omni
Wednesday: Qwen · OpenAI IPO · SpaceX S-1 · Erdős · Nvidia · Anthropic profit
Thursday: AI executive order · Newsom · Fractional AI · Modal
Friday: Glasswing
Weekend 16-17 May
OpenAI and Malta’s government agreed to give every Maltese citizen a free year of ChatGPT Plus once they complete an AI literacy course.
The University of Malta designed the course and the first phase launches in May. OpenAI called it a world first. It is the latest deal under OpenAI for Countries, which George Osborne heads.
OpenAI and Malta partner to bring ChatGPT Plus to all citizens · OpenAI
Monday 18 May
A jury found Elon Musk sued OpenAI and Sam Altman too late.
Nine jurors returned a unanimous verdict in under 2 hours, and Judge Yvonne Gonzalez Rogers adopted it. The court never ruled on whether OpenAI broke a charitable trust, because the claims fell outside a three-year statute of limitations. Musk called it a ruling on “a calendar technicality” and will appeal to the Ninth Circuit.
Jury verdict, Musk v. Altman · CourtListener
Post-trial order re: advisory verdict · CourtListener
Elon Musk has lost his lawsuit against Sam Altman and OpenAI · TechCrunch
Musk slams Altman trial verdict as a ‘technicality,’ vows to appeal · CNBC
Regarding the OpenAI case · Elon Musk on X
Anthropic will brief the Financial Stability Board (FSB) on cyber weaknesses in the financial system found by Claude Mythos Preview, the FT reported.
Andrew Bailey, governor of the Bank of England and chair of the FSB, asked for the briefing. The FSB brings together G20 regulators and is due to report on AI in finance next month. About 40 organisations have access to Mythos Preview.
Anthropic to brief financial watchdog on cyber flaws · Financial Times
Anthropic bought Stainless, the SDK maker used by OpenAI, Google and Cloudflare, and is shutting down its hosted products.
Stainless has generated every official Anthropic software development kit (SDK), the toolkits that connect developers’ code to a service. Hundreds of companies use it to build SDKs, command-line tools and Model Context Protocol (MCP) servers. Anthropic gave no terms. The Information had reported talks at >$300m.
Anthropic acquires Stainless · Anthropic
Stainless is joining Anthropic · Stainless
Anthropic has acquired the dev tools startup used by OpenAI · TechCrunch
Cursor released Composer 2.5, a coding model built on Kimi K2.5 (Moonshot AI), which Artificial Analysis ranked third on its Coding Agent Index.
Cursor trained it on 25x more synthetic tasks than Composer 2. It costs $0.5/m input and $2.5/m output tokens, or $3/m and $15/m for the faster default. It scored 62, behind the highest-effort Claude Opus 4.7 (Anthropic) on 66 and GPT-5.5 (OpenAI) on 65, which cost ~10-60x more per task.
Introducing Composer 2.5 · Cursor
Cursor’s Composer 2.5: third on the Coding Agent Index · Artificial Analysis
Blackstone and Google announced a joint venture to offer Google’s TPU chips as a cloud service, with a $5bn initial equity commitment from Blackstone.
The company gives customers a way to use TPUs outside Google Cloud, and expects its first 500 MW online in 2027. Google will supply the chips, software and services. Benjamin Treynor Sloss, a Google infrastructure veteran, will be CEO.
Blackstone announces joint venture with Google to create TPU cloud · Blackstone
Blackstone will create a new TPU cloud in a joint venture with Google · Google
Tuesday 19 May
Andrej Karpathy joined Anthropic’s pretraining team.
The OpenAI founding member, later Tesla’s director of AI, said he wants to “get back to R&D”. Anthropic’s Nicholas Joseph said Karpathy will build a team “focused on using Claude to accelerate pretraining research itself”.
Personal update: I’ve joined Anthropic · Andrej Karpathy on X
Excited to welcome Andrej to the Pretraining team · Nicholas Joseph on X
OpenAI co-founder Andrej Karpathy joins Anthropic · Axios
Google made Gemini 3.5 Flash the default in the Gemini app and AI Mode in Search.
Google calls it its “strongest agentic and coding model yet” and says it beats Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%) and MCP Atlas (83.6%). At $1.5/m input and $9/m output tokens, it costs 3x as much as Gemini 3 Flash Preview, Simon Willison noted. Google also launched Gemini Omni, a model that takes any mix of text, images, audio and video, starting with video generation. The Gemini app has >900m monthly users, up from 400m at last year’s I/O.
Gemini 3.5: frontier intelligence with action · Google
The Gemini app becomes more agentic · Google
Gemini 3.5 Flash: more expensive, but Google plan to use it · Simon Willison
Google unveils Gemini Omni ‘any-to-any’ AI model · VentureBeat
Wednesday 20 May
Alibaba unveiled Qwen3.7-Max, a proprietary model built to run agents across hundreds or thousands of steps.
Alibaba says it scores 80.4 on SWE-Bench Verified, level with Claude Opus 4.6, and beats DeepSeek V4 Pro on Terminal-Bench 2.0. It will be available through Alibaba Cloud Model Studio. In one test it optimised a kernel for T-Head’s ZW-M890 chips, hardware it had never seen in training.
Qwen3.7: the agent frontier · Qwen
Qwen3.7: the agent frontier · Alibaba Cloud
OpenAI is preparing to file confidentially for an initial public offering (IPO) as soon as Friday, CNBC confirmed after a WSJ report.
Private investors value OpenAI at >$850bn. It is working with Goldman Sachs and Morgan Stanley on the filing. The New York Times reported that OpenAI is preparing to file to go public in the coming weeks.
OpenAI is preparing to file for an IPO very soon · The Wall Street Journal
OpenAI to confidentially file for IPO as soon as Friday · CNBC
OpenAI is preparing to file to go public · The New York Times on X
SpaceX’s IPO filing showed xAI lost $6.4bn from operations in 2025.
SpaceX filed publicly to list on Nasdaq under the symbol SPCX, in what is set to be the largest IPO in Wall Street history. xAI’s figures include X: its loss came on $3.2bn of revenue, against a $1.6bn loss in 2024. Grok’s AI features had 117m monthly users in March, out of 550m across Grok and X. xAI plans to buy $2.8bn more turbines over three years. SpaceX had set aside $530m for potential litigation losses. Wired reported that some could stem from complaints over sexualised images generated by Grok.
Space Exploration Technologies S-1 · SEC
SpaceX files for IPO that could make Elon Musk a trillionaire · BBC
xAI burned $6.4B last year · TechCrunch
Musk’s xAI is being sued over its generators, now it’s buying $2.8B more · TechCrunch
SpaceX listed Grok’s ‘Spicy’ mode as a risk in its IPO filing · Wired
SpaceX’s filing showed Anthropic is paying it $1.25bn a month for compute until May 2029, and Anthropic said it will expand onto Colossus 2.
Either party can terminate on 90 days’ notice. Anthropic agreed on 6 May to use all of SpaceX’s Colossus 1 data centre, >220,000 Nvidia GPUs, without giving terms. Payments are reduced in May and June. Anthropic’s Tom Brown said it will scale up on Nvidia GB200 capacity in Colossus 2 throughout June.
Space Exploration Technologies S-1 · SEC
Anthropic is paying SpaceX $15 billion per year · Axios
Higher usage limits and a SpaceX compute deal · Anthropic
We’re expanding our partnership with SpaceX · Tom Brown on X
OpenAI said an internal model disproved a 1946 conjecture in discrete geometry.
Paul Erdős’s unit distance problem asks how many pairs of points in a plane can be exactly distance 1 apart. Mathematicians had long believed square-grid constructions were essentially optimal. OpenAI’s model found an infinite family of examples that improves on them by a polynomial factor, using ideas from algebraic number theory.
OpenAI says the proof came from a new general-purpose reasoning model, not a system built for maths, and that external mathematicians have checked it. Fields medallist Tim Gowers wrote that, had a human submitted the paper to the Annals of Mathematics, he would have recommended acceptance “without any hesitation”.
An OpenAI model has disproved a central conjecture in discrete geometry · OpenAI
Planar point sets with many unit distances · OpenAI
Remarks on the disproof of the unit distance conjecture · Alon, Gowers et al.
Nvidia reported Q1 fiscal 2027 revenue of $81.6bn, up 85% YoY, with Data Center revenue up 92% to $75.2bn.
It authorised another $80bn of share buybacks and guided to $91bn of revenue in Q2, assuming no Data Center compute sales to China. Jensen Huang told CNBC that Nvidia has “largely conceded” China’s AI chip market to Huawei and that investors should “expect nothing” on approvals.
Nvidia announces financial results for first quarter fiscal 2027 · Nvidia
Nvidia says it has ‘largely conceded’ China’s AI chip market to Huawei · CNBC
Anthropic is heading for its first profitable quarter.
Anthropic projects a $559m operating profit on Q2 revenue of $10.9bn, a 130% jump, the WSJ reported. Last summer it told investors not to expect annual profits before 2028. Coding tools drove the growth, and compute is expected to fall from 71 cents per dollar of revenue in Q1 to 56 cents.
OpenAI took about $5.7bn in Q1, roughly $1bn more than Anthropic, The Information reported on Thursday. Its adjusted operating margin was minus 122%.
Growth to propel Anthropic into its first profitable quarter · The Wall Street Journal
OpenAI generated nearly $6 billion in revenue in first quarter · The Information
Thursday 21 May
The US President postponed the signing of an executive order on AI.
The draft would have set up a voluntary framework under which AI labs share models with the government at least 90 days before release. Mark Zuckerberg, Elon Musk and David Sacks spoke to Trump before the signing, arguing that reviewing models before release would slow innovation and hurt the US against China. On Monday, over 60 “MAGA” allies had urged mandatory government testing of the most powerful models.
Scoop: White House postpones AI EO signing ceremony · Axios
Trump AI executive order seeks early government access to advanced models · Axios
Why Trump’s AI executive order was pulled · Axios
Trump yanked AI order after David Sacks raised industry concerns · Politico
Scoop: 60+ MAGA allies tell Trump to vet AI before release · Axios
California governor Gavin Newsom signed an executive order directing the state to prepare workers and small businesses for job losses from AI.
State agencies will explore severance standards, employment insurance for displaced workers, worker ownership and universal basic capital, and track hiring and payroll data for early signs of layoffs. His office called it a first in the US.
First-of-its-kind executive order to prepare for AI disruption · Governor of California
The enterprise services firm Anthropic set up with Blackstone and Hellman & Friedman bought Fractional AI as its founding team.
The firm helps mid-size companies build Claude into their operations, and Fractional’s applied AI engineers will work with Anthropic’s Applied AI organisation from day one. Goldman Sachs, General Atlantic, Apollo, GIC and Sequoia are also backers.
AI-native enterprise services firm announces acquisition of Fractional AI · Blackstone
Modal, a cloud platform for AI inference, training and agent sandboxes, raised $355m at a $4.65bn valuation.
General Catalyst and Redpoint led the round. Modal says it has grown fivefold since September to >$300m of annualised revenue.
Modal’s Series C: raising $355M at a $4.65B valuation · Modal
Friday 22 May
Anthropic said it and about 50 partners used Claude Mythos Preview to find over 10,000 high- or critical-severity bugs since Project Glasswing began.
The limit is now how fast bugs can be verified, disclosed and patched, Anthropic says. It has disclosed 530 high- or critical-severity bugs to open-source maintainers. Cloudflare alone has found 2,000 bugs, 400 of them high or critical. Anthropic plans to widen the programme to US and allied governments, and to release Mythos-class models generally once it has stronger safeguards.
Project Glasswing: an initial update · Anthropic
Project Glasswing: what Mythos showed us · Cloudflare
Aleth · Open Access Research