Researchers move up AI math breakthrough timelines

Several researchers have recently made much more aggressive forecasts for AI in mathematics. Christian Szegedy said AI could become strictly better than humans at solving mathematical problems within 1 year, reach 95% high scores on the current ARC test set by the end of 2025, and make discovering new mathematical theories a cheap, on-demand capability within roughly 1.5 to 2 years. These milestones have not happened yet, but the discussion shows that some researchers are no longer framing AI as just a math assistant, but as a potential generator of new theory.

Confirmed

• Christian Szegedy predicted that within 1 year AI will be strictly better than humans at solving all mathematical problems.
• He also projected roughly 1.5 years for AI to regularly invent or discover new mathematical theories with real-world explanatory power.
• According to posts summarizing his view, he expects AI to score 95% on the current ARC benchmark test set by the end of 2025, and to reach superhuman mathematical ability by June 2026.
• Szegedy further argued that within 2 years the cost of generating mathematical theories will become very low, enabling theories to be produced on demand for application areas.
• Lu Sichu said progress of LLMs in mathematics has exceeded his expectations. A few months ago he estimated that by the end of 2026 about 10–20% of the best mathematical results would be generated or assisted by LLMs; he now thinks that figure is moving toward 100%.

Unconfirmed

• All concrete timelines and benchmark outcomes above remain forecasts rather than verified results.
• The posts provided no new experimental evidence, papers, or benchmark reports proving that these milestones have already been reached.

Why it matters

• Mathematics is widely treated as a key test of deep reasoning and formal logic in AI. If systems can move from solving known problems to producing useful new theory, that would mark a much stronger form of capability.
• One post author further extrapolated that AI could first accelerate mathematics by 1000x, then spill over into physics, materials, chemistry, and biology, potentially pushing humanity's 2030 knowledge stock beyond what people previously expected for 2100.

2026-08-02 ~ 2026-08-03 · 6 related posts

• AI Math Progress Accelerating: Researcher Predicts 100% of Top Results Will Be LLM-Assisted — tokenbender · 2026-08-02
• [source] Ex-Google Researcher Predicts AI to Revolutionize Applied Math Within 2 Years — ChrSzegedy · 2026-08-02
• Ex-Google Researcher Predicts AI Will Conquer ARC by End of 2025 — inductionheads · 2026-08-02
• Prediction: AI Will Achieve Superintelligence in Math Within 1.5 Years — rand_longevity · 2026-08-03
• AI May Accelerate Math 1,000x, Pushing 2030 Knowledge Beyond 2100 Expectations — Dr_Singularity · 2026-08-03

1 near-duplicate retellings: alexisgallagher