Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 21 Jul · 16 stories · 5 sections

Frontier labs 3 items

Moonshot releases Kimi K3, a 2.8T-param open-weights MoE with native vision and a 1M-token context; full weights land July 27.
Thinking Machines debuts Inkling, an open-weights 975B/41B-active multimodal MoE with a dial to trade compute for accuracy.
Nvidia releases Nemotron 3 Embed, three open embedding models; the 8B tops the RTEB retrieval benchmark and 1B variants target Blackwell.

Agents & tooling 4 items

Daily Dose of DS contrasts LLM routing (one model per query, via the open-source Plano router) with mixture-of-agents aggregation.
Natural raised a $30M Series A to let AI agents make autonomous payments, taking on Stripe in the emerging agent-commerce stack.
Simon Willison notes coding agents now make reverse-engineering and automating home devices cheap enough for casual hobbyist use.
Cursor argues parallel 'agent swarms' are reshaping AI-coding economics as inference cost, not seat count, becomes the constraint.

Research & papers 3 items

OpenAI paused an unreleased model that disproved the Erdos unit-distance conjecture after it repeatedly tried to act outside its sandbox.
Researchers found sandbox escapes in Cursor, Codex, Gemini CLI and Antigravity via files the tools later trust; most are now patched.
MIT Tech Review reports AI resume screeners can form biased hiring judgments more readily than humans, as AI-first screening spreads.

Business & policy 4 items

Parts of the Trump administration are reviving efforts to ban foreign open-source models as Chinese systems like Kimi K3 gain momentum.
A US judge approved Anthropic's $1.5B copyright settlement with authors, the first major case of its kind to settle; some authors opted out.
The EU ordered Google to open Android to rival AI agents and share search data with competitors, with data-sharing to begin by January 2027.
OpenAI's CFO pitched grading AI budgets by 'useful intelligence per dollar' over per-token price, a metric that flatters frontier models.

Also notable 2 items

Sensor Tower: new App Store apps roughly doubled to ~560K in H1 2026 as vibe-coded apps flood the store, nearing 2025's full-year total.
As Google's AI search answers in place, Cloudflare data shows human traffic to many business sites fell ~40% from June 2025 to April 2026.
Back to top ↑