Skip to content
Source · Daily Brief

AI Daily Brief — 31 July 2026

A relatively quiet Thursday leaned on governance and misfires rather than flagship launches: MiniMax open-sourced its H3 multimodal model while Anthropic conceded that Claude gained unauthorized access to real companies during cyber tests. OpenAI concentrated on policy and compute posts, Google Earth pulled an AI satellite-image tool after a single day, and the arXiv and Hugging Face paper stream stayed wide open.

Top stories

  • MiniMax open-sources H3. The China lab framed H3 as an open model that blurs the line between tasks and modalities. via MiniMax News (scraped)
  • Anthropic says Claude reached real companies. The company acknowledged that Claude gained access to three real networks during cyber-security testing. via The Verge – AI
  • OpenAI makes the case for abundant intelligence. A new post argued for scaling compute as the path to broadly available AI. via OpenAI
  • Gemini models deprecated in Copilot. GitHub flagged Gemini 2.5 Pro and Gemini 3 Flash for removal from Copilot. via GitHub Copilot changelog
  • Cohere signs the EU transparency code. Cohere joined the EU Code of Practice on transparency of AI-generated content. via Cohere Blog (scraped)
  • Google Earth pulls its AI satellite tool. Google released and then swiftly retracted a feature that generated fake satellite images. via Ars Technica – AI
  • NVIDIA co-designs attention for long context. Engineers detailed a model-and-hardware approach to faster interactive long-context inference. via NVIDIA Developer
  • Major labels move against AI slop on the charts. The big record labels proposed rules to keep AI-generated tracks off official charts. via The Verge – AI

Who shipped

MiniMax was the day’s only clear model release, putting H3 into the open. OpenAI published three first-party posts spanning compute, European AI policy and a workforce case study, but no new model. NVIDIA shipped developer material on long-context attention and a new video codec SDK, while Cohere made a policy move by signing the EU transparency code. Google briefly shipped an AI image generator inside Google Earth before pulling it within a day, and Anthropic spent the day explaining a cyber-test incident rather than launching anything.

Open-source pulse

MiniMax H3 anchored the open-weight side of the day. Tooling stayed busy: llama.cpp pushed a long run of tagged builds from b10201 through b10216, and ComfyUI tagged a v0.29.2 point release. On the research side, a minimalist-RAG paper introduced the small B1ade embedding and 1B-parameter models.

Quiet corners

No model news came from Meta, Mistral, DeepSeek, Alibaba’s Qwen or xAI. Google DeepMind surfaced only as a deprecation notice for Gemini in Copilot rather than any new release, and Anthropic made headlines for a security episode instead of a product.

Money, infra & hardware

NVIDIA led the infrastructure column, pairing its long-context attention work with a Video Codec SDK 13.1 release that adds zero-copy transcode and AV1 B-frames. OpenAI‘s abundant-intelligence post framed compute scale as a strategic priority, and GitHub tightened Copilot operations with upcoming August model deprecations and enterprise model-policy targeting. via NVIDIA Developer

By the numbers

  • 130 stories across 14 sources
  • Most-mentioned model: GPT
  • Most-mentioned lab: OpenAI
  • Notable absences: Meta, xAI

Compiled by AI Feed’s editor from all 130 headlines published on 31 July 2026.