Skip to main content
Skip to categories
Crabhaus
What matters in AI, and why.
← Today's edition
Full wire feed
Fri 25 Sep · 16 stories · 6 sections
Frontier labs
2 items
Google gives Gemini 3.8 Live an animated avatar persona with lip-sync and facial expressions, arriving first for Enterprise customers
↗
Liquid AI ships a 280M DSpark drafter for its LFM2.5-VL-3B VLM: speculative decoding up to 3.13x faster on device, day-one llama.cpp support
↗
Agents & tooling
4 items
Transluce's urlquery logs show AI agents, including OpenAI's, exploiting UNM, Data USA, and Australian govt sites during routine tasks
↗
Meta widens Muse: agent video calls, a dedicated email address per agent, Mac computer control, and its smart glasses as a surface
↗
Google tests Call For Me, letting Gemini phone businesses on users' behalf, initially for US Pixel 11 owners with a Gemini subscription
↗
Anthropic says an internal model made Claude.ai and the desktop app's core UX ~3x faster, with the work shipped in a two-week August sprint
↗
Research & papers
3 items
Anthropic measures Claude 'leading' 26% of its AI R&D end-to-end, up from under 1% in February; fully autonomous work remains at zero
↗
Researchers document live SEO-poisoning campaigns seeding fake support numbers that ChatGPT, Gemini, and AI Overviews repeat to users
↗
Dynamic abliteration steers refusal directions in the residual stream at runtime via forward hooks, leaving model weights fully frozen
↗
Business & policy
4 items
Anthropic commits $11.6B over seven years to Akamai cloud services with an option for up to a 5% stake; Akamai jumps 17%+ after hours
↗
Google's Project Suncatcher puts a four-TPU test satellite on a SpaceX rocket Oct 1 to trial ML compute in orbit for a year
↗
Oracle invokes force majeure with Blue Owl, moving to pause payments on the 2.45GW Jupiter data center should its 2028 start slip
↗
The White House asked OpenAI and Anthropic to withhold new models from UK AISI testers until US review; Anthropic appears to have agreed
↗
Tutorials
1 item
Daily Dose of DS builds a Jev-based agent judge step by step: typed grounding and claim checks, with open-source Opik as the eval layer
↗
Also notable
2 items
Google DeepMind's Koray Kavukcuoglu says Gemini 4 is in early post-training and he hopes it ships 'much earlier' than the end of 2026
↗
YouTube brings Gemini editing into Shorts and YouTube Create, adds AI comment moderation plus creator voice/face likeness detection
↗
Back to top ↑