AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Sam Altman on OpenAI's Core AI Subscription Vision and Road to AGI

· Training Data

In this live conversation at Sequoia's AI Ascent, Sam Altman walks through OpenAI's origin story and strategic vision. He recounted how the 14-person team in 2016 was a pure research lab with no product plan, eventually releasing GPT-2 as model weights, launching the GPT-3 API in mid-2020 (which Silicon Valley noticed but the world largely ignored), and shipping ChatGPT on November 30, 2022, after observing that users loved chatting with the model in the playground despite its poor chat capabilities. ChatGPT now serves over 500 million weekly users. Altman outlined OpenAI's ambition to become users' "core AI subscription," comprising the model itself, ChatGPT-style surfaces, future devices, and eventually a platform layer. He framed the platonic ideal as a tiny reasoning model with a trillion tokens of context holding a user's entire life, without any weight customization. He acknowledged the $40B round at a $340B valuation and emphasized OpenAI's no-master-plan approach, working forward from immediate tasks rather than backward from a grand vision. On research, he said the biggest leverage remains algorithmic breakthroughs, with perhaps a few 10x or 100x improvements still available. Looking ahead, Altman predicted 2025 will be the year of agents (with coding as a dominant category) and AI-assisted scientific discoveries, while 2027 will see robotics transition from curiosity to serious economic value creator. He described coding as central (not vertical), voice as a major unsolved surface that could enable entirely new device categories, and incumbents as doomed to a familiar pattern of two years of denial followed by a too-late capitulation. He closed with advice on founder resilience, arguing the post-crisis rebuilding period (day 60) is psychologically harder than the acute crisis itself.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.