Can ChatGPT really solve unsolved maths? OpenAI says Astra, its unreleased next major model family, generated ten substantial new results across mathematics and theoretical computer science. Astra is not the version of ChatGPT you can use today, but OpenAI has published a 249-page paper collection, discovery notes and Lean certificates for independent inspection.
In this video I break down what OpenAI actually released, why "ten solved problems" needs qualification, what the roughly $2,000 inference claim does and does not mean, how Lean verification works, and why AI can make progress on advanced maths while still failing apparently simple tasks.
Get AI with Kyle's daily AI newsletter:
https://aiwithkyle.com/join
Chapters:
00:00 What ChatGPT Astra claims
00:53 What OpenAI actually released
02:01 Evidence vs marketing
02:25 Advances vs solved problems
04:00 What the $2,000 claim really means
05:55 Why maths, but not strawberry?
07:58 The jagged frontier
09:08 Can we trust the proofs?
11:52 AI is already escaping the lab
13:50 Is AI actually creative?
15:54 The bottleneck has moved
17:26 Scientists become directors
18:27 My verdict on ChatGPT Astra
Sources:
OpenAI - Ten advances in mathematics and theoretical computer science
https://openai.com/index/ten-advances-in-mathematics/
OpenAI - Ten Advances paper collection
https://cdn.openai.com/pdf/ten-proofs-oai.pdf
OpenAI - Mathematical discovery notes
https://cdn.openai.com/pdf/reasoning-walkthroughs.pdf
OpenAI - Public Lean certificates
https://github.com/openai/ten-proofs
Noam Brown - Astra launch post
https://x.com/polynoamial/status/2083467194663571701
Counting Ability of Large Language Models and Impact of Tokenization
https://arxiv.org/abs/2410.19730
Navigating the Jagged Technological Frontier
https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
