The Omniscient Bulletin · 2026-08-03
The Omniscient Bulletin — August 3, 2026
An unreleased OpenAI model closed ten open problems in mathematics and left proofs a machine can check, two teams solved the same problem with the same chatbot three hours apart, and on Sunday Europe and California began requiring synthetic content to admit what it is.
Good Morning! Mathematics had a strange weekend. OpenAI published ten proofs on Saturday, each closing a problem that had held out for years, all produced by an unreleased version of its next model. It formalized every one in Lean, so anyone with the compiler can check the work without taking OpenAI's word for it. Finding them cost about $2,000 in tokens. Scientific American surfaced the smaller and stranger companion story on Friday: two groups of cryptographers, unaware of each other, pointed the same model at the same open problem and posted answers three hours and eighteen minutes apart. Priority and peer review are the institutions mathematics runs on, and neither was built for that. The rest of the weekend belonged to the lawyers. A Munich court told Suno on Friday that it cannot train on songs it never licensed, and on Sunday both the European Union and California began requiring generative systems to label what they produce. Google spent Friday demonstrating why, withdrawing a tool that had been making photorealistic satellite images of nuclear plants and bombed hospitals a day after it shipped.
Frontier
OpenAI says an unreleased model solved ten open problems in mathematics, and published proofs a machine can check
OpenAI published ten proofs on Saturday, each closing a problem open for years, all produced by an internal version of its next model, Astra. They include new bounds on how densely spheres pack in high dimensions, a construction settling a central open question in group theory, and a disproof of a longstanding conjecture of Alain Connes. The model formalized every argument as a Lean certificate on GitHub, so anyone with the compiler can check the work without trusting OpenAI. Finding the solutions cost roughly $2,000 in tokens. None of it has been peer reviewed, and Astra has no release date.
Research
Two teams pointed the same model at the same open problem and finished three hours apart
Scientific American reported Friday that two groups of cryptographers solved the same open problem, and that both had used the same model to do it. Prabhanjan Ananth of UC Santa Barbara and Amit Sahai of UCLA posted a proof about unclonable encryption at 10:35 in the morning. Seyoon Ragavan, a doctoral student at MIT, posted his three hours and eighteen minutes later. Neither knew of the other's work, and both had asked GPT-5.6 Sol Ultra. Ragavan says the model produced the proof and drafted the first version. The teams are now discussing merging them.
Frontier
DeepSeek's newest open model lands one point off OpenAI's cheapest line, at a fraction of the cost
DeepSeek released the weights of V4 Flash on Friday under an MIT license, with no gate on the download. Artificial Analysis scored it 50 on its intelligence index, one point behind Luna, the cheapest model in OpenAI's current lineup, at roughly 60 percent less cost per task; OpenAI's top model, Sol, still scores nine points higher. The architecture is unchanged from April's preview: the entire ten point gain came from post training. Input runs 14 cents per million tokens and cached input is 98 percent cheaper again. It had been downloaded 156,000 times by Sunday.
Policy
On Sunday, two of the world's largest markets began requiring AI to identify itself
The European Commission began enforcing the AI Act's transparency rules on Sunday. Chatbots must now disclose that they are machines, and synthetic images, audio and video must carry machine readable marks. Brussels can fine a company up to 15 million euros or 3 percent of global revenue. The same day, California's AI Transparency Act took effect, requiring any generative service with more than a million users in the state to watermark what it produces and offer a free public tool to detect it. Systems already on the market have until December to meet the European marking rule.
Policy
A Munich court orders Suno to stop training on six songs it says the model copied
The Munich Regional Court ruled Friday that Suno infringed the copyright of six well known songs, among them Rasputin, Atemlos and Mambo No. 5, the first European ruling of its kind against an AI music service. It ordered the company to stop reproducing the works, to stop training on them, and to disclose what it earned. Suno had argued the songs were not protected, that US fair use covered the training, and that its weights are mathematics rather than stored music. The court rejected all three. Damages are not yet set, the judgment is not yet enforceable, and Suno is weighing an appeal.
Industry
Google withdrew its satellite image generator one day after shipping it
Google pulled the image generator it added to Google Earth on Friday, a day after launch, saying it had seen people "sharing screenshots of generated imagery that appear to violate our policies." The feature turned a prompt into photorealistic satellite imagery. Henk van Ess, an open source researcher, said he asked it for refugees at the Mexican border, a nuclear plant in Iran and a hospital with a bomb crater in Gaza, and that nothing was refused. NPR made an oil terminal burn and flooded the US Capitol. Every image carried Google's watermark, which prevented none of it.
Industry
Anthropic's valuation nears a trillion dollars and it is moving its listing up
The Wall Street Journal reported Saturday that Anthropic's revenue growth and valuation have both passed OpenAI's, the valuation nearing $1 trillion on the strength of Claude Code. Anthropic is accelerating a fall listing and has begun meeting investors, telling them it leads the company that started the boom. OpenAI's side is thinner. ChatGPT's growth has slowed, Fidji Simo's departure forced a reshuffle among Sam Altman's deputies, and sales staff are discounting to hold business customers. Some large investors have privately questioned its cash burn; others hedged by buying into Anthropic.