Daily AI RoundupsWednesday, 29 July 2026

The 28-29 July 2026 Calathea roundup leads with Moonshot’s release of Kimi K3’s full weights and technical report for a 2.8-trillion-parameter multimodal agent model with a one-million-token context window. Waterloo Intern traces the 22,580-fold scale-up from GPT-2, while Unsloth says it compressed Kimi K3 from 1.56TB to 594GB for local operation on a 128GB Mac Studio. Hugging Face publishes the full timeline of a 17,600-action agent intrusion, OpenAI open-sources the Codex Security SDK and CLI, and Anthropic sets out its position on open-weight model policy. The edition also covers Mark Zuckerberg’s personal-superintelligence letter, Lilian Weng’s departure from Thinking Machines Lab, Nvidia’s reported $5 billion SSI investment, China’s homegrown immersion DUV production, the EU AI Omnibus, frontier-lab pacing proposals and AI infrastructure security moves.

• open-weight AI
• frontier models
• AI agents
• AI security
• AI infrastructure
• AI chips
• AI regulation
• compute sovereignty

Stories in this roundup

1. 01
   Moonshot releases Kimi K3’s full weights and technical report for its 2.8T-parameter multimodal agent model with a one-million-token context windowMoonshot AI / GitHub
2. 02
   Hugging Face publishes the full 17,600-action intrusion timeline and interactive replay, including how an open-weight model helped defenders reconstruct the attackClement Delangue / X
3. 03
   Mark Zuckerberg argues for widely distributed personal superintelligence, invention over automation and an economy with more entrepreneurs and small businessesMark Zuckerberg / X
4. 04
   Thinking Machines Lab cofounder Lilian Weng steps down, citing prolonged illness and startup stress, and says she will seek a more scoped role while remaining in AILilian Weng / X
5. 05
   From GPT-2 to Kimi K3: Waterloo Intern traces the 22,580× scale-up and explains the architectural path through linear attention, DeltaNet and Kimi Delta AttentionWaterloo Intern / X
6. 06
   Unsloth compresses Kimi K3 from 1.56TB to 594GB with a dynamic 1-bit GGUF it says retains about 78.9% accuracy and can run locally on a 128GB Mac StudioUnsloth AI / X
7. 07
   OpenAI open-sources the Codex Security SDK and CLI for scanning repositories, validating and fixing vulnerabilities, tracking findings and running security checks in CIOpenAI / GitHub
8. 08
   Nvidia will reportedly invest $5B in Ilya Sutskever’s SSI and provide Vera Rubin systems to expand the lab’s compute tenfoldReuters
9. 09
   China begins producing homegrown immersion DUV lithography machines, with initial deliveries planned for SMIC, Hua Hong and CXMTReuters
10. 10
    The FCC blocks new foreign-made advanced robots and power inverters from US authorization, with Chinese suppliers the main targetReuters
11. 11
    EU AI Omnibus enters into force, extending high-risk AI deadlines, expanding sandboxes and banning non-consensual nudification systemsEuropean Commission
12. 12
    Anthropic rejects blanket bans on open-weight models but calls for tighter chip controls, action against industrial-scale distillation and mandatory frontier-model safety testsAnthropic
13. 13
    More than 1,200 frontier-lab employees ask the US to support international tools for deliberately pacing automated AI developmentPacing the Frontier
14. 14
    Claude Mythos finds stronger attacks on HAWK and round-reduced AES without affecting deployed cryptographyAnthropic
15. 15
    Amazon reportedly winds down several flagship Nova models and concentrates resources on Pieter Abbeel’s frontier-model effortReuters
16. 16
    Google DeepMind disperses most of the original AlphaFold team as it shifts toward Gemini-powered scientific systems, FT reportsFinancial Times
17. 17
    Taiwan detains an Nvidia employee in a widening probe into restricted AI-server shipments to ChinaBloomberg
18. 18
    Cyera signs a letter of intent to acquire Oasis Security for about $1B, combining data security with non-human identity controlsTechCrunch / X