Daily AI RoundupsSunday, 26 July 2026

The 26 Jul 2026 Calathea roundup leads with Sam Altman telling Ti Morse on Relentless that OpenAI redirected compute and teams from Sora and its browser toward coding agents. It also covers poolside’s synthetic-data pre-training pipeline, Kyle Mistele and Dan Luu on practical coding-agent workflows, and a bipartisan US bill that would let the government throttle or suspend risky AI models. Anduril is reportedly seeking funding at a $100 billion valuation. vLLM 0.26.0 adds Inkling support and DeepSeek-V4 performance work, while AMD introduces ROCm.AI for coding-agent optimization. Two papers examine in-memory attention state and LoRA memorization, and Microsoft open-sources MLVC, a learned video codec already being tested in Teams.

• AI agents
• coding agents
• AI policy
• AI funding
• model serving
• GPU software
• AI research
• video compression

Stories in this roundup

1. 01
   On Relentless with Ti Morse, Sam Altman says OpenAI redirected compute and teams from Sora and its browser toward coding agentsRelentless
2. 02
   poolside explains its synthetic-data pipeline for pre-training, including difficulty-tuned templates, multistage reshaping and automated quality filteringpoolside
3. 03
   Kyle Mistele argues coding-agent failures stem from poor loop design, then demonstrates control-theory-inspired iteration with small, verifiable diffsKyle Mistele
4. 04
   Dan Luu examines agentic coding workflows, benchmark reliability and why teams often report 2x rather than 10x gainsDan Luu
5. 05
   US lawmakers unveil a bipartisan AI ‘kill switch’ bill that would let the government throttle or suspend models deemed too dangerousPOLITICO
6. 06
   Anduril is reportedly in talks to raise funding at a $100 billion valuation, more than three times last year’s markTechCrunch
7. 07
   vLLM v0.26.0 adds Inkling support and DeepSeek-V4 performance work in 411 commits from 212 contributorsvLLM
8. 08
   AMD introduces ROCm.AI to help coding agents deploy, debug and optimize models for Instinct hardwareThe Register
9. 09
   An independent paper proposes keeping LLM attention state in memory to reduce KV-cache data movementZenodo
10. 10
    Kaizhen Tan, Heqing Du and Yang Feng measure how much data LoRA adapters can memorize, finding placement matters more than parameter countarXiv
11. 11
    Microsoft open-sources MLVC, a learned video codec already being tested in TeamsMicrosoft