AI daily

2026-08-02

OpenAI's Next-Gen Model Solves Ten Open Math Problems for Under $2,000, Marking a Shift in AI's Role in Mathematics

AI DAILY BRIEFING

OpenAI's Astra model, an internal test version of its next-generation model, has achieved ten new research results in mathematics, including resolving the existence of non-sofic groups, providing a counterexample to Connes' rigidity conjecture, and solving three Erdős problems. The proofs were generated by Astra, organized into papers by human researchers, and formally verified using Lean, with a total token cost of about $2,000. The results have been highly praised by top scholars, with the construction of non-sofic groups considered Fields Medal-worthy. This marks a shift in AI's role in mathematics, entering uncharted research territory and raising questions about authorship and evaluation in science.

Astra's success in solving open math problems demonstrates AI's capability to contribute original research, potentially transforming the role of AI in mathematics and science.

The use of formal verification via Lean ensures the correctness of the proofs, but the authorship and evaluation of AI-generated research remain open questions.

The low cost of under $2,000 for such significant results highlights the potential for AI to accelerate scientific discovery at scale.

Watch next: The broader implications of AI-generated research on scientific publishing, peer review, and the definition of authorship, as well as the potential for AI to tackle other complex scientific problems.

Featured

APriority80
机器之心

Tsinghua Team Open-Sources VeriLoop Coder-E1: Evidence-Governed Spiral for Verifiable Recursive Self-Improvement

A team from Tsinghua University's Shenzhen International Graduate School has open-sourced VeriLoop Coder-E1, a code model built on Qwen3.6-27B that achieves top scores on several software engineering benchmarks. The model uses narrow-domain PEFT fine-tuning and a Self-Harness mechanism to enable verifiable recursive self-improvement through an evidence-governed spiral.

SynthePulse InsightRead deep analysis
SSignal89
量子位

OpenAI's New Model Solves 10 Math Problems in a Row, Turning Math into a Click Game for $2000

OpenAI released research results from the internal test version of its next-generation model Astra, which solved 10 long-standing open problems in mathematics and theoretical computer science in one go, covering high-dimensional geometry, coding theory, group theory, and more, with all proofs formally verified using Lean. The results have been highly praised by top scholars, with the construction of non-sofic groups considered Fields Medal-worthy, and the total cost was under $2000.

SynthePulse InsightRead deep analysis
APriority79
机器之心

OpenAI's Next-Gen Astra Model Solves Ten Open Math Problems for Just $2,000

OpenAI has revealed that an internal version of its next-generation model, Astra, achieved ten new research results in mathematics, including resolving the existence of non-sofic groups, providing a counterexample to Connes' rigidity conjecture, and solving three Erdős problems. The proofs were generated by Astra, organized into papers by human researchers, and formally verified using Lean, with a total token cost of about $2,000. This marks a shift in AI's role in mathematics, entering uncharted research territory and raising questions about authorship and evaluation in science.

SynthePulse InsightRead deep analysis
BStandard60
THE DECODER

Claude Opus 5 Pushes Prompt-to-Game AI from Rough Color Blocks to Full 3D Prototypes with Physics and Music

Anthropic's Claude Opus 5 can generate complete 3D games from single prompts, including a first-person shooter, a kart racer, and a Minecraft clone, without any external assets. Geometry, textures, physics, and sometimes music are produced as code and run directly in the browser. In side-by-side comparisons with GPT-5.6 Sol and Kimi K3, Opus 5 delivers significantly more detailed results.

SynthePulse InsightRead deep analysis
APriority78
钛媒体AGI

AI Companies Bulk-Buy Old Books for Clean Training Data

Multiple AI companies are purchasing used books in bulk through intermediaries, scanning and destroying them to obtain training data untouched by AI-generated content from before 2022. Anthropic's 'Project Panama' exposed this supply chain, leading to a $1.5 billion settlement. This highlights the scarcity of high-quality human text for AI training.

Selected
8
Sources
6
Featured
8
More
0
All reports