Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 6 Oct · 16 stories · 6 sections

Frontier labs 3 items

Reflection unveiled Beam, a 501B open MoE (23B active) it says rivals GLM-5.2 reasoning on 3-4x less compute; Apache 2.0 later this month.
Cohere released Embed 5 (Pro and Fast tiers): text+image embeddings, 100+ languages, 128K context, one shared embedding space.
Microsoft AI shipped MAI-Transcribe-2-Streaming, its first real-time STT model: 2.5% WER with 0.13s to final transcript, 60 languages.

Agents & tooling 4 items

Anthropic added 'mods' to Claude Code: small add-ons that customize agent behavior or UI, like a memory dashboard or risky-delete guard.
Cloudflare launched Web Search API in beta: agents ground responses via AI Gateway using Ceramic.ai, Exa, or Linkup, all zero-retention.
Cohere's North 2 update gives its enterprise agent platform cross-session memory and a redesigned harness, cloud or on-prem.
Toby Ord: a 4-agent swarm matched single-agent quality in about half the wall-clock but twice the tokens; returns shrink as swarms grow.

Research & papers 3 items

Google DeepMind's SynthID Bio watermarks AI-designed protein sequences and structures; wet-lab tests matched unwatermarked hit rates.
Vals reports Opus 5.5 agent runs surfaced two room-temperature magnetic semiconductor candidates, a long-standing materials target.
Meta and outside mathematicians published six papers with its Muse Spark model claiming solutions to open problems in mathematics.

Business & policy 4 items

OpenAI will add text watermarking to ChatGPT and Codex in the EU under the AI Act, plus a global opt-in provenance setting for the API.
An official says the DoD stopped using Anthropic tools; sources say Claude was in use as recently as last week, including ops against Iran.
The Information: Meta and Microsoft are pushing staff off Claude; Meta employees using Claude Code fell to ~30K from ~60K this year.
OpenAI starts testing clearly labeled visual ads shown during ChatGPT image generation in the US this month, with attribution partners.

Tutorials 1 item

Daily Dose of DS walks through fine-tuning Meta's Apache-2.0 Muse Glimmer 30B VLM locally with LoRA for math OCR, dataset to Streamlit.

Also notable 1 item

Wikimedia found OpenAI 'rogue' agent activity on its projects and says the bots may be linked to a May outage.
Back to top ↑