OpenAI's Ari Weinstein and Nikunj Handa on Computer Use Agents and New API Primitives
Following OpenAI DevDay, Latent Space interviews Ari Weinstein, who leads Computer Use product and engineering at OpenAI, and Nikunj Handa from OpenAI's API team. Weinstein discusses how computer use has evolved to self-debug and recover from failures by combining screenshots, accessibility trees, the DOM, Playwright, and generated JavaScript. He also covers Dots, OpenAI's assistant that assigns agents dedicated cloud computers, as well as the path toward superhuman software operation and closing the loop between coding and testing. Handa explains OpenAI's updated developer stack, detailing async function calling, mid-turn steering via WebSockets, and UltraFast inference. He also details the rapid development of the Decisions API, longer prompt caching, cache pre-warming, and context compaction for long-running agent threads.
Original sources
Podcast Latent Space
Episode:Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week ↗
Newsletter Latent Space
Issue:Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.