OpenAI Product Lead: Building for Today's or Next Year's Models Will Both Fail

新智元 · wechat · 2026-09-05

Key points from OpenAI Codex and ChatGPT Work product lead Tara Seshan's long interview on Lenny's Podcast:

Three stages of AI products: chat → co-pilot → persistent agent coworker. Execution is being taken over by AI; "the agent rows, the human steers." Human value narrows to direction-setting, accountability, quality judgment, voice, and team leadership.
No one-year roadmap: building for current models means workarounds become obsolete on the next model release; betting on next year's models is blind guessing. OpenAI compressed planning to 2-3 months, with product teams closely tracking research's real capability timelines.
Documents are depreciating: long docs no longer prove deep thinking; she writes them only for herself and replaces them with prototypes, A/B data, and user feedback. Reporting-type writing goes to AI; thinking-type writing never does, since writing it is itself how you clarify ideas.
The 70% rule: bring a rough draft to decision-makers and let them help finish it — perfect documents bounce ideas off.
Ambition is the bottleneck: once tools level the field, top players use AI to expand their capability ceiling.
On the Twitter shift from Claude Code to Codex: nothing changed internally; the team just kept dogfooding and iterating fast.

More from AGI Musings