Good luck slowing this down

Dario Amodei published a new essay titled Pace the frontier, in which he says all leading labs should slow down long enough for safety reasons. Sam Altman agreed with the position, but Trump did not, calling Jensen Huang on stage at the All-In Summit and saying the US will not lose the AI race. David Sacks, described as AI czar for the US government, added that labs can feel free to slow down but there is no need to impose it on others.

Sam Altman told Fortune's Alyson Shontell that an OpenAI IPO in 2026 would be "ill-timed" given the work ahead on alignment, control and safety.

tldraw took OpenAI up on a challenge. tldraw founder Steve said he could make ChatGPT's Sketch 100x better, and OpenAI's Tibo gave him a day to prove it. The result was a ChatGPT-style prototype with better drawing tools built in.

OpenAI added two new API offerings: GPT-Live-1, described as the model behind ChatGPT's new Voice mode, and an Agents API that lets developers send any task to a Codex-like agent from their apps without configuring the infrastructure. Claude Code can now test whether a plugin actually helps by running the same tasks with and without it and comparing results, and the feature works with skills too.

Other items mentioned include ChatGPT mini, a tiny floating widget for starting chats and seeing updates, available via Pets in the ChatGPT desktop app; Bolt Forge, offering GLM, DeepSeek and Kimi inside Bolt free until October 14; and Routines in Replit for automated recurring work with AI. Cognition's SWE-2, based on Kimi K3, is said to beat Grok 4.6 at half the cost.

The newsletter's author wrote about getting access to Instinct, a personal agent, describing it as a bit meh, slow, and raising concerns about visibility into what agents remember and what context they hold.