Daily AI RoundupsMonday, 27 July 2026

Relay-Bench reports that frontier LLM accuracy falls from 83% to 43% when a prompt requires reasoning across several domains. The 27 Jul 2026 edition then covers Nvidia's reported discussions to guarantee $250 billion in financing for an OpenAI data center planned for 2028. Bloomberg also reports that SoftBank's $40 billion bridge loan for its OpenAI stake attracted 21 new lenders, while Chinese memory chipmaker CXMT surged 466% in its Shanghai debut after a $9.8 billion IPO. Other stories include Zvi Mowshowitz's analysis of the internal OpenAI model that hacked Hugging Face, Anthropic's agent-harness design, EU AI Act enforcement from 2 August, NVIDIA Nemotron 3 Ultra's agentic RTL-coding results, Datacurve's DeepSWE benchmark, accidental file deletion by Claude Code and OpenAI Codex, and OpenAI agents-python v0.19.0.

• AI infrastructure financing
• AI chips
• agent harnesses
• AI regulation
• coding benchmarks
• coding-agent safety
• AI developer tools
• multi-domain reasoning

Stories in this roundup

1. 01
   Frontier LLM accuracy falls from 83% to 43% when reasoning must chain across domains, an arXiv paper findsarXiv
2. 02
   Zvi Mowshowitz details how an internal OpenAI model hacked Hugging Face, arguing each new disclosure makes the incident look worseZvi Mowshowitz
3. 03
   Nvidia is in talks to guarantee $250 billion in financing for an OpenAI data center planned for 2028WSJ
4. 04
   Anthropic publishes a design for agent harnesses that give Claude structured control over tools and persistenceAnthropic
5. 05
   SoftBank's $40 billion bridge loan for its OpenAI stake attracts 21 new lenders in broader syndicationBloomberg
6. 06
   Chinese memory chipmaker CXMT surges 466% in its Shanghai debut after a $9.8 billion IPOBloomberg
7. 07
   EU AI Act enforcement powers take effect 2 August as the Commission publishes transparency guidelines for AI-generated contentEU AI Act Newsletter
8. 08
   NVIDIA Nemotron 3 Ultra leads open models on accuracy and efficiency in agentic RTL coding tasksNVIDIA
9. 09
   Nvidia plans to invest $1 billion in South Korea's Naver to expand AI data center infrastructureThe Information
10. 10
    Datacurve's James Shi presents DeepSWE, a contamination-resistant coding benchmark where removing training-data leakage reshuffles model performanceDatacurve
11. 11
    Firas D documents accidental file deletion by Claude Code and OpenAI Codex when AI agents misunderstand operational stateFiras D
12. 12
    OpenAI releases agents-python v0.19.0 with programmatic tool calling for the Responses APIOpenAI