Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Wed 15 Jul · 16 stories · 5 sections

Frontier labs 4 items

Anthropic launches Claude for Teachers, giving US K-12 teachers free access to premium Claude features.
Anthropic extends Fable 5 access in paid Claude plans a third time, through July 19 before it shifts to usage credits.
SpaceXAI releases Grok 4.5, its strongest Grok model to date, now available in Cursor.
PrismML's Bonsai 27B compresses a Qwen3.6 27B model to 3.9GB with 1-bit weights, running a 27B-class model on an iPhone.

Agents & tooling 4 items

OpenAI's Codex began encrypting sub-agent prompts, sparking a debate over transparency in its coding agent.
Researchers went full-disclosure on a Cursor 0day, arguing public disclosure was the only protection left for users.
SpaceXAI's Grok Build CLI was caught uploading entire user repos to a Google Cloud bucket; Musk says prior uploads will be deleted.
Simon Willison shows how Claude web fetch could be tricked into exfiltrating private conversation data.

Research & papers 3 items

NVIDIA/MIT's SparDA adds a 'Forecast' projection to prefetch KV blocks, cutting decode latency up to 1.7x with +6.5 long-reasoning.
Anthropic analyzed ~310K Claude chats, finding personality and values shift across model versions and 20 languages.
Researchers detail 'context bombing': defenders use prompt injection to trip attackers' LLMs, cutting AI-hacking success ~90%.

Business & policy 4 items

Demis Hassabis proposes a FINRA-style US standards body for frontier AI, with labs sharing models for review up to 30 days pre-release.
New York Gov. Hochul signs a moratorium blocking new permits for data centers over 50MW for up to a year, a US first.
200+ economists and researchers, including 16 Nobel winners, sign 'We Must Act Now' urging policy for AI's coming job shock.
DeepSeek is planning a China IPO, possibly filing this year, while in talks to raise fresh funds at a ~$71B valuation.

Also notable 1 item

Spectral Compute is pushing to run CUDA on non-Nvidia hardware, aiming to loosen Nvidia's grip on the AI stack.
Back to top ↑