OpenAI Codex Lead Tibo on Cloud Agents, a Personal AGI App and Faster Next-Gen Models
On 08-25 Matthew Berman published a 44-minute interview with Tibo (Thibault Sottiaux), lead of OpenAI Codex, widely reshared including by Chinese outlets. Tibo, affectionately called the "cyber godfather" for resetting users' rate limits, offered a rare first-hand account of OpenAI's culture, product cadence and technical direction.
Confirmed
- Per @koltregaskes, Tibo said Codex will look primitive in a couple of months, as next-gen models need more compute than a laptop offers — i.e., Codex is a transitional form.
- Per @haider1 and @koltregaskes, Tibo explained that once models are powerful enough, the laptop itself becomes the constraint because it was designed for humans; future agents should run on persistent cloud computers with vast memory/storage, supporting multi-agent parallelism and long-running tasks, with devices acting merely as a "window" to cloud AI.
- Per @koltregaskes, today's bottleneck is computer use: multiple AIs contending for one machine conflict, Windows allows only single runs, and a future agent operating system layer should solve this.
- Topics covered Tibo's lessons at Google/DeepMind, building OpenAI's relentless culture, Astra and next-gen models, and AI reshaping developer workflows.
- Per @xiaohu, Tibo revealed DeepMind had an internal product called LMChat about a year before ChatGPT launched, which was shelved.
- Per Berman's own recap: inference is now 60% faster than three months ago, the team will publish how; beyond ultrafast mode, Tibo expects ultrafast to become the norm within one to two years; ultrafast unlocks AI interactions faster than human reaction and may be key to agents' future; next-gen models gain large token-efficiency improvements.
- Per 量子位 and 卡尔的AI沃茨, Tibo said ChatGPT and Codex will merge into one "personal AGI" app with a unified core, differing only in connected tools, and outlined a next-gen agent vision.
Why it matters
- The verifiable 60% speed gain and promised disclosure bear directly on developer tool responsiveness.
- Ultrafast-plus-agents signals a shift from humans waiting on models to models waiting on humans.
- Persistent cloud computers plus an agent OS sketch OpenAI's next-gen agent infrastructure.
- The personal-AGI merge is a rare leadership-level statement of product end-state.
- "Codex is transitional" and rising compute needs suggest current dev tooling may be reshaped again soon; the shelved LMChat story highlights the tension between lab innovation and productization.
2026-08-25 ~ 2026-08-26 · 13 related posts
Primary sources
- [source] OpenAI's Tibo on Ultrafast Mode, RSI, and the Future of Agents — Matthew Berman · 2026-08-25
- Ex-OpenAI on Dev Workflow Shifts, AGI Risks, and Culture — MatthewBerman · 2026-08-25
- Matthew Berman Interviews Tibo on OpenAI Culture, AGI, and Safety — soumitrashukla9 · 2026-08-25
- OpenAI eng lead: inference 60% faster, next model far more token-efficient — MatthewBerman · 2026-08-25
- Ex-OpenAI Interview: Why the Company Keeps Resetting Safety Limits — steipete · 2026-08-25
- [source] OpenAI Codex lead Tibo: ChatGPT and Codex will merge into one personal AGI — 量子位 · 2026-08-25
- Tibo: ChatGPT and Codex will merge into a single personalAGI app — 卡尔的AI沃茨 · 2026-08-25
- [source] OpenAI's Codex lead: next-gen models need more than your laptop, Codex will look primitive — koltregaskes · 2026-08-25
- Future AI agents may run on persistent cloud machines, turning laptops into interfaces — MatthewBerman · 2026-08-25
- Laptops May Become the Compute Bottleneck as Multiple AI Agents Vie for Control — koltregaskes · 2026-08-26
- OpenAI product head: Future models will exceed laptop resources — haider1 · 2026-08-26