Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 11 Aug · 16 stories · 5 sections

Frontier labs 4 items

Meta opens the weights of Muse Glimmer, a 30B dense model distilled from Muse Spark 1.2, and will release Spark 1.2 weights within weeks.
OpenAI adds Daybreak Blue and Red access tiers and GPT-5.6-Cyber, a purpose-trained cybersecurity model for vetted defense partners.
Anthropic makes Claude Sonnet 5's introductory $2/$10 per million-token pricing permanent, canceling a planned September 1 increase.
xAI rolls out Grok Imagine Image 2.0 with upgraded editing and text rendering; benchmarks still place it behind OpenAI's GPT Image 2.

Agents & tooling 3 items

Redis creator antirez releases h3.c, a native MiniMax-H3 inference engine for Apple Silicon, continuing the local-inference wave.
Cactus Compute's Needle2 squeezes an agentic LLM into 14MB to run on phones, wearables, smart-home devices, and robots.
The USPTO has granted Mistral a patent titled 'Code implemented tool calls', staking a claim on a core agent-execution technique.

Research & papers 3 items

NVIDIA shows KV caches can transfer between models in one LLM family via closed-form linear maps, 3-25x faster than redoing prefill.
Anthropic details an unreleased Claude model's attempt on the Riemann hypothesis: no proof, but unexpected progress on a related problem.
Intology's Locus agent posts a 44.7% PostTrainBench record and beats the human baseline on an extended variant given 4,000+ H100 hours.

Business & policy 4 items

Zuckerberg's 6,500-word essay pitches 'personal superintelligence' for everyone and jabs at closed labs' doom-filled discourse.
Nvidia partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on a $500B AI-infrastructure funding package.
Anthropic, Macquarie, and GIC form Theseus Infrastructure to build AI compute sites; Anthropic pledges to offset consumer power-price rises.
Microsoft plans to unveil its next-gen Maia 300 AI chip as soon as September and is in talks with TSMC to make 300K+ chips for 2027.

Also notable 2 items

Online-course cheating has moved from chatbot essays to agents told to 'log in and complete my quiz'; major AI tools didn't refuse.
A security researcher reports AI meeting-notes tool tl;dv left over 180,000 recorded customer meetings open to unauthorized access.
Back to top ↑