Interview

How to Understand the Next Wave of AI Before Everyone Else | Thibault Sottiaux and Matthew Burman

OpenAI’s Tibo (Thibault Sottiaux) joins Matthew Burman to break down what’s next for Codex, ultra-fast AI, recursive self-improvement, and the future of personal AI agents. We also get into OpenAI vs. Anthropic, why Codex usage is exploding, and how AI could soon work faster than humans can keep up with.

Source attribution: Original source

Interview details

Publication date: Sep 3, 2026
Duration: 44m

Speakers

Matthew BermanInfluencer, Forward Future Thibault SottiauxEngineer, OpenAI

Viewpoints in this interview

Fast experimentation should coexist with coherent product quality

Sottiaux says OpenAI tries to preserve a bottom-up culture where people can quickly test and ship ideas while balancing that freedom with simplicity, performance, and overall product coherence.

Thibault Sottiaux Product culture Source at 3:12

Companies must remain willing to disrupt themselves

Sottiaux argues that organizations need conviction, rapid user feedback loops, and a willingness to redirect resources toward promising new ideas even when doing so disrupts established priorities.

Thibault Sottiaux Product culture Source at 4:37

Personal agents should deeply understand and adapt to users

Sottiaux describes the desired personal agent as one that understands a user's goals, daily activities, and surrounding context, reacts when needed, acts proactively, and minimizes the need for users to manage underlying agent mechanics.

Thibault Sottiaux Personal AI agents Source at 7:53

Future agents will need resources beyond personal laptops

Sottiaux argues that computers designed around human workloads will become a constraint for increasingly capable models, which can potentially operate across far more concurrent applications and therefore require greater computational resources.

Thibault Sottiaux AI compute Source at 8:58

AI should adapt to human attention rather than demand multitasking

Sottiaux says AI products should manage information and interruptions around human attention, with sufficiently fast models enabling users to remain in a continuous creative flow rather than manually coordinating many parallel agents.

Thibault Sottiaux Personal AI agents Source at 11:17

Some complex processes can become mostly autonomous

Sottiaux distinguishes personal agents from full automation, describing systems that could handle tasks such as performance optimization, regression repair, and vulnerability patching with humans involved mainly for high-risk approvals.

Thibault Sottiaux AI automation Source at 12:59

Coding and general-purpose AI are converging into one agent

Sottiaux says merging Codex and ChatGPT reflects a future in which the same underlying multimodal agent handles both technical and nontechnical tasks while its interface adapts to each individual's needs.

Thibault Sottiaux Personal AI agents Source at 14:45

AI interfaces should become ambient and natural

Sottiaux expects future AI interactions to extend beyond text toward ambient experiences that can understand surrounding context and support natural voice-based interaction.

Thibault Sottiaux AI interfaces Source at 19:19

OpenAI focuses more on its strengths than competitors

When asked about Anthropic, Sottiaux says his focus is on building capable and efficient models, distributing powerful technology broadly, and accelerating toward OpenAI's own values rather than closely tracking competitors.

Thibault Sottiaux AI competition Source at 21:04

Product failures should be compensated rather than merely acknowledged

Sottiaux says Codex usage resets began as a way to compensate users when the product broke or performed below expectations, reflecting a broader principle of materially making up for poor experiences.

Thibault Sottiaux Customer trust Source at 24:03

Frontier models can help optimize their own supporting infrastructure

Sottiaux says more capable models are being used to re-engineer inference and serving infrastructure, producing substantial cost and speed improvements that OpenAI aims to pass on to users.

Thibault Sottiaux AI efficiency Source at 27:52

Recursive improvement extends beyond models training models

Sottiaux says recursive self-improvement can include models improving the inference stack, kernels, cloud-agent systems, products, and other infrastructure that determines how effectively the models themselves can be used.

Thibault Sottiaux Recursive self-improvement Source at 30:54

Greater model capability requires greater safety investment

Sottiaux says alignment and safety become increasingly important as model capabilities grow, and describes a training pause as necessary for teams to understand and harden the system before restarting.

Thibault Sottiaux AI Safety Source at 32:25

Ultra-fast AI is especially valuable when every second matters

Sottiaux says OpenAI uses ultra-fast inference for high-stakes situations such as incident response, where reducing latency has immediate operational value.

Thibault Sottiaux Ultra-fast AI Source at 34:29

Low latency enables more interactive creative workflows

Sottiaux says very fast AI can support shared canvases, rapid image and prototype generation, and real-time steering, allowing users to evaluate ideas by creating them rather than waiting on lengthy generation cycles.

Thibault Sottiaux Ultra-fast AI Source at 38:51

Today's ultra-fast speeds may become broadly standard

Sottiaux predicts that continuing gains in token efficiency, inference hardware, and system performance could make today's ultra-fast speeds close to the default within one or two years, while premium higher-compute tiers remain available.

Thibault Sottiaux Ultra-fast AI Source at 40:09

Efficiency gains can make advanced intelligence broadly accessible

Sottiaux argues that continued improvements in model efficiency will make capabilities that are currently expensive much cheaper, supporting OpenAI's stated goal of broad access and practical utility.

Thibault Sottiaux AI access Source at 42:12