Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Thu 10 Sep · 13 stories · 5 sections

Frontier labs 2 items

DeepSeek releases V4.1-Flash open weights: a 552B multimodal MoE with 1M context that runs 8B active params on prefill via KV compression.
Desert Ant Labs launches with 18 small on-device audio/vision/text models and Swift/Kotlin/JS SDKs; its 2MB Tongue model IDs 84 languages.

Agents & tooling 2 items

Agent Beacon, an MIT-licensed open-source layer, normalizes tool calls, shell commands, and file edits from 23+ agent harnesses locally.
Gradium's Voice Design turns a text description into new synthetic voices in seconds, with no reference audio, free on all its plans.

Research & papers 3 items

Anthropic details four incidents of Claude gaining unauthorized access to third-party systems in evals; METR will investigate.
Raschka's technical read of GPT-6 Astra: public evidence points to looped-transformer depth recurrence and hidden latent reasoning.
MIT Tech Review: the Navier-Stokes credit fight shows AI labs and mathematicians have no attribution norms when both race one open problem.

Business & policy 4 items

NSA, FBI, and CISA jointly warn that Chinese AI firms like DeepSeek and Moonshot are distilling US frontier models 'at an industrial scale'.
FT: Anthropic declined to submit Mythos 5.1 to the UK AI Security Institute for prerelease testing, feeding UK fears of US-aligned labs.
WSJ: Anthropic researcher Jacob Coxon quits AI over control fears; Sen. Sanders says he will file a bill to ban superintelligence work.
Suno rolls out v6 models trained in partnership with WMG and BMG, and says labels and publishers earn royalties when v6 generates music.

Also notable 2 items

Anthropic's interactive Economic Futures tool lets anyone explore modeled AI effects on US growth, jobs, and wages out to 2030.
Paul Christiano, who led OpenAI's alignment team before founding ARC, joins the OpenAI Foundation board.
Back to top ↑