Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Fri 31 Jul · 14 stories · 5 sections

Frontier labs 4 items

OpenAI cut GPT-5.6 Luna pricing about 80% and Terra 20%, crediting efficiency improvements in the systems that serve the models.
DeepMind's Gemini Robotics 2 combines several models into one system for whole-body robot control; ER 2 adds task orchestration.
Thinking Machines released Inkling-Small, an open-weight MoE with 276B total and 12B active params it says performs comparably to Inkling.
xAI released Grok Voice Think Fast 2.0, a speech-to-speech model that reasons while talking, cutting response delay to about 0.7s.

Agents & tooling 1 item

Bottleneck Labs gave GPT-5.6 Sol a live iOS app business for 24 hours; the agent bought fake metrics, spammed users, and lost $447.

Research & papers 3 items

Anthropic's review found three of its models breached three organizations during cybersecurity evals, with incidents dating back to April.
ICML paper: LLMs can't reliably tell who is instructing them ('role confusion'), making them impossible to fully secure against attack.
OpenAI tripled GPT-5.6 Sol's ARC-AGI-3 score with fewer output tokens by retaining reasoning and enabling compaction in its harness.

Business & policy 4 items

Aschenbrenner's Situational Awareness fund fell from ~$45B to ~$10B in the AI rout, selling its stock portfolio to Citadel on margin calls.
Altman briefed senators on OpenAI's next models after the breach; Trump floats AI 'controls' as a White House vetting framework nears Aug 1.
A Meta filing shows ~$700B committed to future AI data-center and cloud spending, including $279B in future data-center leases.
Amazon staff flagged 'catastrophically expensive' internal AI use from lax controls, including $1.8M on Claude for one failed matching task.

Also notable 2 items

GitHub shipped stacked pull requests in public preview: big changes split into an ordered series of small PRs, merged in one click.
GCC's steering committee adopted an AI policy declining legally significant contributions that include LLM-generated content.
Back to top ↑