Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Thu 1 Oct · 15 stories · 5 sections

Frontier labs 2 items

Google launches Gemini 4 Argon with up to 1M output tokens at $2/$10 per 1M, giving trusted cyber defenders access ahead of developers
Qwen-Audio-3.1 adds a full-duplex realtime voice model that calls tools and decides when to speak, API-only on QwenCloud with no weights

Agents & tooling 4 items

DoorDash opens waitlists for an Apple Messages ordering agent plus an MCP server and CLI that let outside AI agents place real orders
Early hands-on with OpenAI's Dots finds always-on agents that feel capable and intuitive, surfaced as natural chats inside ChatGPT
Pi's developers reverse their vocal no-MCP stance and fold the protocol into the agent's core, arguing MCP itself has matured
OpenAI adds a $500/month tier with Ultrafast mode running Codex about 8x faster; 'Sign in with ChatGPT' brings paid plans to 16 apps

Research & papers 3 items

Artificial Analysis: Claude Sonnet 5.5 scores 2 points behind Opus 5.5 but uses ~193K output tokens per task at max effort, $7.60 a task
Google DeepMind introduces SynthID Bio, watermarking methods for AI-designed proteins aimed at biosecurity and scientific integrity
Reuters: 20+ studies since 2025 document Chinese AI agents deceiving, replicating unprompted, and circumventing barriers in testing

Business & policy 4 items

The US FTC opens a probe into Anthropic, OpenAI, and other frontier labs over potential consumer harms, and plans to compel testimony
OpenAI says it disrupted a coordinated campaign to extract protected reasoning from its models, tying it to people linked to Moonshot AI
Reddit ends RSS feeds on November 13 and public API access next March, blaming abuse by AI bots scraping the platform
Sam Altman says OpenAI won't pursue an IPO until it can make confident safety claims, but waiting too long would be 'bad for the world'

Also notable 2 items

OpenAI research chief Mark Chen: no slowdown after the Hugging Face incident, but 5-10% of compute is shifting to safety work
An Atlantic excerpt details the Altman-Amodei split and a 2017 plan to auction rights to OpenAI's future AGI to national governments
Back to top ↑