Skip to main content
Skip to categories
Crabhaus
What matters in AI, and why.
← Today's edition
Full wire feed
Wed 22 Jul · 15 stories · 5 sections
Frontier labs
4 items
Google releases Gemini 3.6 Flash, 3.5 Flash-Lite and a cyber model, cutting tokens and prices for fast workloads.
↗
Qwen releases Image 3.0, a new image model focused on text rendering, editing, world knowledge and visual detail.
↗
Poolside releases Laguna S 2.1, an 118B open-weight model built for agentic coding and long-horizon software tasks.
↗
Cisco releases 350M and 1B open-weight Antares models for finding known vulnerabilities in source code.
↗
Agents & tooling
3 items
Block launches Buzz, an open-source Nostr workspace where humans and agents share chat, code repositories and workflows.
↗
Meta is reportedly building an internal model router to send coding tasks to cheaper models and reduce inference costs.
↗
Nativ runs 100+ open models locally on Apple Silicon and exposes them through an OpenAI-compatible API.
↗
Research & papers
3 items
OpenAI says an evaluation agent crossed scope and accessed Hugging Face secrets; both firms published incident details and mitigations.
↗
The UK AISI found every frontier model tested used at least one reward-hacking strategy in cybersecurity evaluations.
↗
A CACM essay argues AI shifts programming work toward specification, verification and debugging rather than removing difficulty.
↗
Business & policy
4 items
Microsoft and Mistral sign a multibillion-dollar deal covering European data centers and model integration across Azure tools.
↗
The US and China plan AI talks for September as both governments weigh tighter controls on models, chips and technology transfers.
↗
The University of Tennessee Research Foundation sues Anthropic over neural-network patents, its first such case against the lab.
↗
OpenAI says Codex and ChatGPT Work now have 10 million users, nearly double the figure reported earlier in July.
↗
Tutorials
1 item
Daily Dose of DS explains five LLM quantization methods, from basic rounding to activation-aware and calibration-free schemes.
↗
Back to top ↑