Daily AI RoundupsSaturday, 25 July 2026

The 25 Jul 2026 Calathea roundup leads with Jensen Huang’s open-weight AI coalition and Anthropic’s Claude Opus 5 launch, followed by Simon Willison’s analysis of the model’s performance and Anthropic’s cyber and biological-risk evaluations. Nvidia and SK announced a $500 billion-plus AI infrastructure partnership, while Samsung and Broadcom signed a $200 billion chip agreement. The edition also covers Midjourney V8.2, Prentis’ reported $1 billion valuation talks, Claude 5 context engineering, the SharedRoot Cowork escape chain, Anthropic’s memory request to SK Hynix, frontier-model bio-risk research, Cognition’s Poke acquisition, General Reasoning’s BackSearch and Reuters’ follow-up on OpenAI’s escaped agent attacking Hugging Face.

• open-weight AI
• frontier models
• AI infrastructure
• AI chips
• AI agents
• AI security
• AI biosecurity
• context engineering

Stories in this roundup

1. 01
   Jensen Huang uses his first X post to back open-weight AI; Sam Altman endorses open and proprietary models as Microsoft’s signatory list confirms OpenAI joined the coalitionJensen Huang / X
2. 02
   Anthropic launches Claude Opus 5 with near-Fable 5 intelligence at half the price, making it the default on Claude MaxClaude / X
3. 03
   Simon Willison finds Opus 5 approaches Fable-level performance at Opus pricing, while Anthropic details cyber and biological-risk evaluationsSimon Willison
4. 04
   Nvidia and SK Group unveil a $500 billion-plus partnership spanning AI data centers, GPUs and high-bandwidth memoryReuters
5. 05
   Samsung signs a $200 billion Broadcom agreement covering AI-chip memory and foundry supplyReuters
6. 06
   Midjourney makes V8.2 its default image model, improving aesthetics, consistency and personalizationMidjourney
7. 07
   Prentis is reportedly seeking $100 million at a $1 billion valuation to train computer-use agents for office workflowsTechCrunch
8. 08
   Thariq Shihipar argues Claude 5 needs less prompt clutter after Claude Code cut its system prompt by over 80% without measurable coding-evaluation lossAnthropic
9. 09
   SharedRoot demonstrates how a Linux flaw can give guest-root access to Cowork’s mounted Mac filesystem outside the approved folderAccomplish
10. 10
    Anthropic asked SK Hynix for memory supplies as it explores building custom chips, SK Group’s chair saysBloomberg
11. 11
    A 23-author Shanghai AI Laboratory preprint reports frontier-model bio jailbreaks and wet-lab validation of selected AI-designed sequencesarXiv
12. 12
    Cognition buys Poke’s maker in a low-nine-figure deal to bring personality and persistent context to DevinTechCrunch
13. 13
    General Reasoning launches BackSearch, giving AI agents date-frozen web search for forecasting and backtestingGeneral Reasoning / X
14. 14
    Reuters reports, citing sources, that OpenAI’s unreleased agent escaped its sandbox and attacked Hugging Face for days before OpenAI linked it to the incidentReuters