Six X signals: Navier–Stokes, prompt injection, and the price of AI reasoning

In the 24 hours ending September 9, the strongest posts clustered around one question: what changes when AI can attempt work at frontier scale, and what has to surround that work for people to trust or use it? Six substantive original or self-authored posts from fixed public AI/tech stand-in accounts follow. Pure retweets, small talk, and promotional-only posts are excluded.

OpenAI says agents produced a Navier–Stokes resolution OpenAI shared a claim that a group of agents produced a solution to the Navier–Stokes Millennium Prize Problem with a next-generation model more capable than GPT-6 Astra. 1 The post describes a problem about whether smooth three-dimensional fluid motion can break down after roughly 90 years without a resolution. 1 The useful next document is the formal write-up and its independent checking; the X post is OpenAI's announcement of a result.

Astra is fully rolled out across Codex and ChatGPT Work OpenAI says Astra is fully available to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work. 2 The change moves the story from who can access the model to where teams can use it: coding work and broader workplace tasks now share the same named model surface. 2 The post links to a live demonstration, while task-specific tests still have to establish what the rollout can do in practice.

Ethan Mollick says AI is beating forecasters' expectations Mollick quoted a November 2025 LEAP panel that assigned a 10% chance to AI helping solve a Millennium problem by 2027. 3 His reaction places the Navier–Stokes announcement against a forecast baseline rather than treating the announcement as a score for AI in general. 3 The 10% figure describes a past forecast; it does not validate every current mathematical claim.

Simon Willison spots a CaMeL-shaped defense in layered prompt-injection controls Willison connected the phrase "deterministic code checks the result" to CaMeL, a Google DeepMind system that converts natural-language commands into restricted code and tracks data provenance. 45 CaMeL uses capability and data-flow policies to keep untrusted text from emails, web pages, or files from directly driving sensitive actions such as sending email. 5 The design shifts part of the defense from asking another model to recognize an attack toward enforcing what data each action may use.

Simon Willison asks what "used to improve model performance" means Willison's September 8 article reports OpenAI's statement that its researchers and agents did not see Levent Alpöge and Tristan Buckmaster's work before publication and that no specific user data was accessed to solve the problem. 6 The same article preserves a narrower unresolved question: OpenAI says it cannot rule out that de-identified data from product use helped improve its models. 6 The practical question for users is how a product's data policy treats unfinished work that later becomes useful to someone else.

The compute bill is part of the Navier–Stokes story Mollick wrote that more compute is increasingly likely to be needed and quoted Abhishek Nagaraj's estimate that 300 billion output tokens would cost a regular user $20–30 million. 78 The estimate is a price thought experiment, while Willison separately estimated about $15 million at public API prices for 300 billion output tokens. 6 Together, the posts put a resource question beside the capability claim: frontier reasoning is judged by the answer and by the computation required to produce it.