OpenAI Showed Astra and a Scandal Began. A Breakthrough or Other People's Ideas?

A Sunday ToTheMoon episode with Alexander Volchek and Ilnar Shafigullin. The main news is OpenAI's announcement that GPT-6 Astra solved the Navier–Stokes equations, one of the Clay Mathematics Institute's Millennium Prize Problems with a one-million-dollar prize. Alexander came to the news through a harsh tweet aimed at Sam Altman and immediately asks questions: what are these ten thousand agents, who managed them, and is this not marketing. Ilnar sees a familiar pattern: after every strong model OpenAI reports that something even more powerful is in training.

Ilnar explains the context: in 2000 the Clay Institute chose seven problems and set a million for each; only one has been solved — by Perelman. The Navier–Stokes equations are more than ninety years old. According to OpenAI, its interest arose at the end of August, when a mathematician and an Anthropic employee obtained an intermediate result; after 88 hours of work by ten thousand agents the company announced it had proved the existence and smoothness of a solution.

The scandal: the mathematicians had discussed their ideas in ChatGPT and Claude and suspected they had been used in training. OpenAI answers that no dialogues after 3 July were used, the discussions took place in mid-August, and the company does not claim priority on the intermediate part. At the time of recording the proof has not been recognised. Alexander considers an insert about the price necessary: 130 billion tokens, 2.7 million messages — more than a million dollars at external prices; Ilnar recalls that a scientist stands on the shoulders of giants, and neural networks are especially strong at the intersection of fields.

Impressions of GPT-6 ten days on diverge. Alexander saw no super effect of the GPT-5.5 kind: the model did not solve his personal task, made it worse and spent third-party systems' resources, so he carried on with Fable 5.1. Ilnar switched completely from 5.6 to 6: the benchmarks are torn apart, ARC-AGI is closed at 99%, some tasks need adjustment, but nothing got worse. Both suggest leaving the ten thousand simultaneous agents on OpenAI's conscience — in Anthropic's study eighty agents got in each other's way.

In passing — voice: Alexander's companion on a trip did not realise he was talking to ChatGPT in CarPlay and wanted to pay for such a version. ChatGPT Voice has improved in two months, but the hosts do not use it: Alexander dictates tasks, Ilnar types at the computer.

Shopify showed that visits and orders through AI tripled in a year, search grew 30% in two years, and those who come through AI buy with higher conversion and a 15% larger order. Alexander takes apart what the statistics lack — absolute figures — and agrees with the main point: engagement through AI is higher, the choice narrows to two options. Ilnar adds trust in the model as an expert; society is splitting between those who trust AI and those who do not.

The purchase decision has already moved into the chat: Alexander chooses restaurants on trips through ChatGPT, aggregators no longer exist for him, and Booking with its rating loses to a model that re-reads Expedia, Booking and TripAdvisor together. But the model does not ask questions: over a hundred and fifty restaurants it never once asked about the outcome, although a live user is the most expensive source of training. The next stage, according to Ilnar, is a proactive system on a server for one or two thousand dollars a month; Alexander is ready to pay, but for OpenAI it costs almost two thousand a day, and there is a risk of fine-tuning without control.

The episode ends with the iPhone 18 and the foldable Duo. Alexander asked ChatGPT to compare only what matters: the processor makes no difference, the focal length is a plus for those who know how to use it, and the doubled AI block speeds up who knows what. The Duo does not interest him — Apple, in his view, has gone to someone else's market; Ilnar recalls the Vision Pro, wants to check the crease on the screen but is unlikely to buy. Pre-orders open at the end of October.

The episode is useful because it separates two things that merge in the news: a mathematical result that still has to be checked, and the marketing around it. The hosts deny neither the model's strength nor the possibility that it built on other people's ideas — they show how each scenario works and suggest waiting for the proof to be recognised. The second conclusion is more practical: the purchase decision is already moving into the chat, and what to watch is not the search engine's results but what ChatGPT and Claude say about you — while the models themselves still do not ask the person, and that is the main unsolved question of the interface.