Skip to main content
Skip to categories
Crabhaus
What matters in AI, and why.
← Today's edition
Full wire feed
Wed 16 Sep · 16 stories · 6 sections
Frontier labs
4 items
Google DeepMind ships Gemini 3.8 Live and Live Extended Thinking, dialogue models that add optional reasoning for voice agents.
↗
Apple ships iOS 27 with Siri AI: on-screen awareness, personal context, and in-app actions, as an English-only beta excluding the EU.
↗
Ex-OpenAI researcher's TypeSafe AI debuts Jev, a 'System One' model that outputs typed, calibrated decisions instead of text; $40M seed.
↗
Salesforce announces Koa, a CRM-specialized model built on Nvidia's Nemotron architecture to power AI agents in Agentforce.
↗
Agents & tooling
3 items
Developers are routing the Claude Code harness to cheaper non-Anthropic models like GPT-5.6 Sol via proxies and OpenRouter.
↗
Google gives all its engineers Claude Opus 5 access via Antigravity, while calling Gemini its primary internal development model.
↗
Open-source Magnitude benchmarks your hardware and recommends viable local models, then connects them to Claude Code, Codex, or OpenCode.
↗
Research & papers
2 items
IBM's ALTK-Evolve targets agent flakiness: a GPT-4.1 ReAct agent passed AppWorld 77.4% on average but all five repeat runs only 53%.
↗
SemiAnalysis tests Vera Rubin NVL72: up to 7x token throughput per MW over Blackwell on a 1.6T-parameter DeepSeek model.
↗
Business & policy
4 items
OpenAI says it has worked with Anthropic and Google on AI safety for weeks and needs no antitrust waiver to coordinate on safety.
↗
Microsoft AI drafts a 38-page Code of Conduct for its models: no AI rights, no 'neuralese', must accept shutdown; open for comment.
↗
OpenAI holds early talks to raise another private round at a $1.2T valuation ahead of its planned IPO, per the Financial Times.
↗
Factory raises $200M at a $5B valuation for Droids, coding agents that switch models by task complexity; it was worth $1.5B in April.
↗
Tutorials
1 item
Guide: install the open-source ELI5 skill for Claude Code, then build pass/fail evals that compare it against a no-skill baseline.
↗
Also notable
2 items
OpenRouter spend flipped to OpenAI over Anthropic during the September 7 week, a first since February 2024.
↗
Strix says a leaked personal access token gave it admin access to AI inference provider Baseten's production GitHub org.
↗
Back to top ↑