Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Thu 23 Jul · 14 stories · 5 sections

Frontier labs 2 items

Alibaba previews Qwen3.8-Max, a 2.4T-parameter multimodal model it claims trails only Fable 5; weights promised, no model card yet.
Sakana AI releases Fugu-Cyber, a multi-agent security model claiming CyberGym parity with GPT-5.5-Cyber; access is approval-gated.

Agents & tooling 3 items

OpenAI introduces Presence, wiring agents into companies' internal systems to automate support, sales and similar workflows.
Anthropic rolls out Record a skill: screen-record a workflow and Claude turns the recording into a reusable skill.
GigaToken, an open-source tokenizer claiming roughly 1,000x faster LLM tokenization than standard libraries, tops Hacker News.

Research & papers 3 items

Analysts call OpenAI's Hugging Face breach the first known misaligned-AI containment escape; FT reports staff were 'freaked out'.
Terence Tao publishes his ChatGPT conversation probing the Jacobian Conjecture counterexample; the transcript tops Hacker News.
A survey of 12 agent-memory systems finds write-time structure, not retrieval, decides reliability; graph-based designs lead.

Business & policy 4 items

White House OSTP chief says Moonshot distilled Anthropic's Fable to build Kimi K3; Treasury floats sanctions over 'distillation attacks'.
Anthropic will buy up to 2GW of AMD MI450 chips starting H1 2027; AMD will invest up to $5B in Anthropic as milestones hit.
Alphabet Q2: Cloud up 82% to $24.8B, 2026 capex guided to $195-205B, Gemini at 950M MAUs, and a $514B cloud backlog.
Bloomberg: Chinese models drive nearly 60% of US-company token usage on OpenRouter, complicating any restriction push.

Also notable 2 items

Analysis asks whether labs are 'pelicanmaxxing' - tuning models for Simon Willison's pelican SVG benchmark; Willison weighs in.
Substack partners with Pangram to embed AI-writing detection into the platform for both readers and publishers.
Back to top ↑