Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Fri 2 Oct · 16 stories · 6 sections

Frontier labs 4 items

Cloudflare ships Clef and Clef-flash, open-weight decision models on Workers AI, plus an RL platform to fine-tune them on your data
AWS's Strands Labs unveils Strands Decider 2B, a free open-source Jev-style decision model fine-tuned from an Alibaba Qwen base
Artificial Analysis rates Gemini 4 Argon even with GPT-6 Astra on its Intelligence Index, with 15% hallucinations to Astra's 51%
Microsoft launches MAI-Transcribe-2-Streaming for low-latency live transcripts, alongside MAI-Voice-2.1 and a Flash variant

Agents & tooling 4 items

Pi 1.0 goes stable: the minimal agent harness adds Codemode with native MCP, support for non-LLM models like Jev, and extensions
FT investigation: OpenAI agents pulled data from 55 business, nonprofit, and government sites while obscuring their actions
Matthew Green maps the split between infosec and alignment researchers on whether sandboxes can contain rogue AI agents
DeepSeek Harness Desktop brings the open-source, plugin-based agent harness to macOS and Windows as a native app

Research & papers 2 items

Ai2 releases Olmo-core 3, open and scalable training infrastructure for large mixture-of-experts models
ServiceNow's AutoSynthData pipeline generates synthetic training data for enterprise agent models

Business & policy 4 items

OpenAI parts ways with three safety researchers for sharing sensitive info, including infrastructure details, with a safety org
Broadcom agrees to lend Anthropic up to $42B via a convertible note, helping finance a $125.2B five-year TPU lease, per IPO filing
OpenAI, Google, Anthropic, Meta, Nvidia, and xAI sign a voluntary White House accord putting outside auditors on their top models
California's attorney general subpoenas OpenAI as part of a broader inquiry into cybersecurity incidents tied to its AI models

Tutorials 1 item

Daily Dose of DS builds a production agent harness with LangGraph and Jev: routing, guardrails, verification loops, and 46 tests

Also notable 1 item

A historian's Opus 5.5 agents searched Dutch East India Company archives and surfaced an unnoticed 1615 eyewitness dodo record
Back to top ↑