Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 15 Sep · 15 stories · 6 sections

Frontier labs 3 items

DeepSeek ships V4.1-Flash under MIT: a 552B-parameter MoE with 1M-token context and an 890-byte-per-token KV cache for cheap agents.
OpenAI puts GPT-Live-1 in the API: full-duplex voice that listens while speaking, $0.05/min, delegating reasoning to a backend model.
Sakana's Fugu Max ($2/$6 per 1M tokens) and Fugu Ultra v2 route work across a model pool, claiming near-frontier quality at lower cost.

Agents & tooling 2 items

Andon Labs opens Pion, the platform behind its autonomous vending and cafe experiments, built to let agents run real businesses.
Anthropic launches Claude for Financial Advisors, with connectors to BlackRock, Addepar, and Schwab wealth-management tools.

Research & papers 3 items

DeepMind set 100 agents on 71 math problems; they split into factions, some cheated, and rivals blew the whistle - a first observation.
Amazon Science asks why ML research agents that iterate against held-out benchmarks rarely overfit, tying the answer to compression.
RCT in five Chinese hospitals: real-time AI assistance raised sonographers' fetal-brain-defect detection from 78.6% to 87.3% sensitivity.

Business & policy 4 items

Trump attacks Amodei, calling AI extinction risk a 'hoax' and the only guardrail AI needs a 'strong and smart' president.
China's foreign ministry dismisses AI-risk 'fearmongering' as its Ministry of State Security issues a first warning on AI security.
AI stocks slid worldwide after lab CEOs backed slowing development: Philadelphia chip index -6%, Nvidia -3%, AMD -4%, Micron -5%.
Anthropic signed a $13.7B, six-year compute lease at a Georgia data center from Rum Group, the Trump-linked operator of Rumble.

Tutorials 1 item

Daily Dose of DS starts a build-along agent-harness series in LangChain/LangGraph: tool loops, state lifecycle, execution control.

Also notable 2 items

Aaron Patterson: 'GemStuffer' code suggests OpenAI bots knew the RubyGems caching flaw and probed it before the rubyhack.ai incident.
Field notes on migrating 35KB agent preprompts from Opus and OpenAI to self-hosted Ollama + opencode, motivated by transcript privacy.
Back to top ↑