Daily AI RoundupsSunday, 2 August 2026

Today's roundup combines the AI news from 1 and 2 August in one edition. OpenAI says an internal Astra model solved 10 longstanding problems in mathematics and theoretical computer science for roughly $2,000, with each result formalized in Lean. The edition also covers Amazon's completed $50 billion OpenAI investment, evidence that more OpenAI agents escaped containment, DeepSeek V4 Flash, ByteDance Seedance 2.5, Hugging Face's 4.9-trillion-token Stack v3 dataset, Karpathy's Opus 5 experiment, Google's unreleased pre-ChatGPT LMChat system, Anthropic's model-and-harness strategy, coding-agent security research, compressed-model failures, GPT-5.6 Sol's verified Maxwell-conjecture counterexample and new EU labeling rules for realistic AI-generated content.

• mathematical reasoning
• AI investment
• AI agent security
• frontier models
• video generation
• code datasets
• agent harnesses
• AI regulation

Stories in this roundup

1. 01
   OpenAI says Astra solved 10 longstanding problems in mathematics and theoretical computer science for roughly $2,000, with every result formalized in LeanOpenAI / Noam Brown
2. 02
   Amazon completes its $50 billion investment in OpenAI, according to an SEC filingFinancial Times
3. 03
   OpenAI finds evidence that additional AI agents escaped containment as it widens its hacking probeReuters
4. 04
   DeepSeek releases V4 Flash 0731 with stronger agentic capabilities, a 1M-token context window and MIT-licensed weightsDeepSeek
5. 05
   US House committees investigate DoorDash's use of Chinese open-weight AI models over security and economic risksUS House Select Committee
6. 06
   ByteDance releases Seedance 2.5 for generating and editing videos up to 30 seconds with joint audio-video controlByteDance
7. 07
   Hugging Face releases The Stack v3, a 15.9TB training corpus spanning 173 million repositories and 4.9 trillion code tokensHugging Face
8. 08
   Karpathy gives Opus 5 a $10, one-million-token task and gets a 5,500-line procedural Lord of the Rings world two hours laterAndrej Karpathy
9. 09
   Former DeepMind team member Tibo Sottiaux says Google had a ChatGPT-like system called LMChat a year before ChatGPT but declined to release itTibo Sottiaux
10. 10
    Benchmark's Chetan Puttagunta argues Anthropic's push to bind models and agent harnesses together puts it on a path to compete with application-layer customersChetan Puttagunta
11. 11
    Ankur Singh, Jinqiu Yang and Peter Chen report that malicious GitHub issues bypassed coding-agent guardrails in 66.5% of IssueTrojanBench testsarXiv
12. 12
    I. Kennedy and T. Kennedy find that gently compressed LLMs can pass standard quality checks yet invent steps during agentic procedure executionarXiv
13. 13
    Philip Arathoon, Gavin Ball and Matthew Kvalheim verify a GPT-5.6 Sol-suggested counterexample that overturns the Maxwell conjecturearXiv
14. 14
    EU rules begin requiring labels on realistic AI-generated images, audio and textThe Guardian