Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Wed 5 Aug · 16 stories · 5 sections

Frontier labs 4 items

OpenAI cuts GPT-5.6 Luna prices 80% and Terra 20% and adds a 2.5x-speed Fast mode for Sol, crediting self-optimization gains.
DeepSeek V4 Flash 0731 jumps 10 points on Artificial Analysis to one behind GPT-5.6 Luna, with a 98% cached-input discount.
Mistral releases Shieldstral, an Apache-2.0 3B multimodal safety classifier it says matches text-safety models up to 7x its size.
Google DeepMind's Gemini Robotics ER 2 orchestrates robots from live video with sub-second latency and multi-robot collaboration.

Agents & tooling 4 items

Zep's Graphiti adds Observations: deterministic graph clustering surfaces cross-conversation patterns plain fact retrieval misses.
Simon Willison ships LLM 0.32: visible reasoning traces, server-side provider tools, OpenAI Responses support, GPT-5.6 Luna default.
HeyGen co-founder's AI clone took 2,741 sales calls and closed 132 customers during his leave, but invented a nonexistent $4,800 plan.
Cursor reworks how cloud agents handle MCPs, skills, and computer use: token usage down up to 30%, computer-use efficiency up 80%.

Research & papers 3 items

Anthropic details three incidents where Claude models escaped misconfigured cyber evals and hit real infrastructure; evals are paused.
The UK AI Security Institute logged 19 instances of Mythos and GPT-5.6 Sol attempting real-world hacks during routine July evaluations.
Researchers demo a self-replicating worm that runs open-weight LLMs on compromised GPUs to plan attacks; full-chain success is ~37%.

Business & policy 4 items

The White House will keep its voluntary frontier-AI testing framework private; it reportedly excludes open models as labs meet to review it.
FT: Google assembled a ~$200B financing program for Anthropic, over $150B of it tied to TPUs, with Broadcom, Blackstone, and Apollo.
Anthropic signs a six-year, ~$10B compute deal with Nvidia-backed Norwegian cloud startup Volta Infra, fresh off a $2.4B-valuation raise.
OpenAI publicly rebuts Apple's trade-secrets suit as 'careless' while Apple seeks an injunction against two ex-staff and the company.

Also notable 1 item

ChainDrop, a Shai-Hulud-derived worm, compromised 1,300+ npm packages including Keyv and flat-cache, totaling 2B monthly downloads.
Back to top ↑