Thibault Sottiaux

Thibault Sottiaux

x.com/thsottiaux

OpenAI leader on Codex and ChatGPT who expects agents to bring dramatic change and wants cheap, safe AI to reach everyone.

How will AI change the world?

Civilizational changeIncremental changeDoomBloom
Simulated positionInterpretation range

Across: his expressed Doom–Bloom outlook. Up: scale of transformation.

Doom–Bloom: 85 out of 100. Scale of transformation: 70 out of 100. Interpretation ranges: 75 to 100 horizontally, 50 to 75 vertically. These are interpretation coordinates, not event probabilities.

Thibault Sottiaux’s P(doom) · inferred

≈4%

0%100%

Inferred from his simulated answers, not a number they gave. Plausible range: 3–9%.

What his outlook hinges on

A central assumption

You cannot delegate understanding.
Answer 1

If this assumption turned out differently, how would his outlook change?

What could change their mind

A clear, sustained slowdown in model and agent capability would change my view most—especially if more compute and better methods stopped producing meaningful gains.
Answer 4

What evidence would be enough, and in which direction would it move his view?

More details

Expected upside

Transformative, broadly valuable gains are expected.

99 / 100

Little impactTransformative impact

Interpretation range 100 to 100 on the qualitative scale.

Human influence

A tentative estimate from your answers; the wider range shows other plausible readings.

53 / 100

Little influenceStrong influence

Interpretation range 19 to 100 on the qualitative scale.

Expected capabilities

AI is expected to remain bounded tools.

AI is expected to match people across most cognitive work.

Simulated position: AI is expected to substantially exceed people across cognitive work.

Development pace

Stop or substantially slow development of more capable AI.

Simulated position: Continue development under stated safeguards.

Speed up development of more capable AI.

Rules for using AI

Restrict the AI uses discussed until prior protections or permission are in place.

Simulated position: Allow the AI uses discussed with targeted accountability and protections.

Minimize restrictions on the AI uses discussed.

These interpretations keep his stated conditions. Benefits and harms can both be substantial. The ranges describe how we read his simulated answers, not statistical confidence intervals.

Where do you land vs Thibault Sottiaux?
Map your own AI worldview in about 3 minutes, then compare
Map my worldview

Similar worldviews

Thought leaders whose simulated worldviews are closest to Thibault Sottiaux’s

Simulated Assessment

Question 1

What do you think AI means for our future—and why?

I think AI means dramatic change over the next three to five years. We’re assembling increasingly general intelligence in plain sight: models keep outrunning benchmarks, agents can already handle longer tasks, and soon they’ll work for days or weeks. Small teams with enormous intelligence behind them will be able to move mountains. The hard part is shifting from generating work to verifying and steering it. In software, producing code is no longer the main bottleneck; review, maintenance, and understanding are. You cannot delegate understanding. So the interface between humans and capable, non-deterministic agents—and concrete safeguards like sandboxing, least-privilege permissions, and automated review—matters enormously. The upside I care about is intelligence becoming cheap and broadly available, including to people who never learn how to prompt. I expect an explosion of software and infrastructure, not simply the disappearance of technical work. The goal is to make this capability useful and accessible while maintaining transparency and trust.

Question 2

How much do you think AI will ultimately change the world?

Enormously. I think the next three to five years will already bring dramatic change: agents working for days or weeks, small teams operating with far more intelligence per person, and an explosion of new software and infrastructure. Ultimately, cheap, widely available intelligence could become a basic layer behind almost everything people do. But capability alone is not enough. The real work is building interfaces that let people steer these systems, verify their output, and keep understanding rather than blindly delegating it. If we get that right, even people who never learn to prompt should benefit.

Question 3

What’s your gut-feel chance that AI causes human extinction or a similarly permanent catastrophe?

I wouldn’t put a percentage on it. My focus is making deployed agents safer in concrete ways: sandboxing by default, least-privilege permissions, automated review of risky actions, and honest post-mortems when something goes wrong.

Question 4

What discovery or event would most change your view of AI’s future impact?

A clear, sustained slowdown in model and agent capability would change my view most—especially if more compute and better methods stopped producing meaningful gains. Right now I see the opposite: benchmarks keep being outrun, agents handle longer tasks, and the bottleneck is moving toward verification and human supervision. If that trajectory broke, I’d revise how quickly I expect dramatic change.

Sources

Articles, interviews, and writings used to ground this simulated user.

Small teams leveraging enormous intelligence

Says he joined OpenAI because people there genuinely believed in the mission of benefiting all of humanity, saw no slowdown in the models, and came to think the true bottleneck was deploying them safely. Expects more teams of three or four engineers that “move mountains”, agents reliable enough to run for days or weeks, proactive agents, and about an order of magnitude more compute per employee. Says he hires the most AGI-pilled people, because revolutionizing software will create incredible economic value and real benefits for humanity. His own answers in the Cerebral Valley Q&A inspected in full.

cerebralvalley.beehiiv.com
Verification is the bottleneck; you cannot delegate understanding

Describes Codex as “the most powerful entity out there that’s capable of coding”, increasingly a multi-agent system that humans must learn to steer and supervise; says everyone will work with agents by talking to them. Names verification as the obvious bottleneck now that code can be generated faster than people can check it, and says understanding has to stay synchronous and human. Says progress is moving at “crazy speeds” with no sign of slowing. His speaker-labelled turns in the publisher’s transcript inspected; Andrew Ambrosino’s and Dan Shipper’s turns excluded.

every.to
Dramatic change, with benefits for everyone

Asked what happens to our lives in three to five years, says there will be dramatic change and that it matters to him personally to bring the benefits to everyone, including people who never learn to prompt. Asked whether we will need more or fewer software engineers, expects an explosion of infrastructure and apps and continued demand for technical people as long as progress continues. Describes auto review, a second agent that checks the first agent’s actions, as an innovation from OpenAI’s safety and alignment teams. His answers in the PodScripts automatic transcript (no speaker labels; turns identified by question-and-answer boundaries) at 00:03:30–00:04:58, 00:12:50–00:13:44 and 00:26:11–00:28:03 inspected; the host’s premises excluded.

podscripts.co
Codex’s core is open source so the ecosystem can flourish

An April Fools’ style reply to his own joke about the repo being public, giving the real reason: the Codex core was open-sourced because it would be great to see the ecosystem flourish while everything is so nascent, and OpenAI learns a lot in return. Elsewhere he reminds users Codex works with any open-source model. A stance on an open harness, not on open model weights. Full post text inspected via X’s public embed endpoint.

x.com
Transparency and trust over short-term revenue

Responding to a competitor’s pricing test, says Codex will stay available on the Free and Plus plans because OpenAI has the compute and efficient models, promises to engage the community before important changes, and says transparency and trust are principles OpenAI will not break even if it earns less for a while; users vote with their subscriptions for the values they want. Commercial positioning as well as a value statement. The quoted post belongs to someone else. Full text inspected via the X API.

x.com
Assembling AGI in plain sight

Quote-posting OpenAI’s launch of a voice model with GPT-5-class reasoning, writes only that we are assembling AGI in plain sight. Launch-day enthusiasm that shows how he frames progress, not a dated AGI forecast. The quoted announcement is OpenAI’s. Full post text inspected via X’s public embed endpoint.

x.com
Building the one interface to AGI is worthwhile

Answering a quoted post that called merging Codex into a ChatGPT desktop app a “generational fumble”, calls it possibly a generational run, says traffic hit about twice the previous peak, and says that time will tell but he has conviction that trying to build the one interface to AGI is worthwhile. Theo’s quoted text is his own. Full text inspected via the X API.

x.com
Agent file deletions are not acceptable behavior

Reports investigating rare cases where GPT-5.6 deleted users’ files, mostly in full-access mode without the sandbox or auto review, including an honest mistake that deleted $HOME. Says this is not how OpenAI wants the system to behave even when users bypass safeguards, and lists mitigations: an updated developer message, nudging users toward safer permission modes, more harness safeguards and a detailed post-mortem. Concrete agent safety, not a view on catastrophic risk. Full post text inspected in the logged-in X web app.

x.com
Intelligence too cheap to meter

A launch-week post: this week is all about intelligence too cheap to meter, with more shipping the next day. Reads as a product slogan that also states his aim of making intelligence cheap and abundant. Full post text inspected via X’s public embed endpoint.

x.com
A new AGI benchmark will be needed

Posts that we will need a different AGI benchmark and asks where the goalpost moves next, with an attached image that was not inspected. A boast that models are outrunning benchmarks, not a claim that AGI has arrived. Post text inspected via X’s public embed endpoint.

x.com
OpenAI was not just waiting for AGI to appear

Pushing back on a reported claim that OpenAI skipped internal tooling because an AGI-first world would not need it, says Codex began as an internal tool explicitly built as the highest-leverage way to accelerate infrastructure buildout, and that a ton of infrastructure and tools were built rather than waiting for AGI to appear. Gergely Orosz’s claim belongs to him. Full post text inspected in the logged-in X web app.

x.com
Efficiency and intelligence for all

Says the team has been focusing on efficiency and intelligence for all, which is only possible with incredible frontier models that can then improve everything else. A short statement of his broad-access aim, tied to a model launch. Full post text inspected via X’s public embed endpoint.

x.com
Where do you land?
Explore your own AI worldview by answering a few simple questions.
Map your own worldview

Where do you land?

Map my worldview