Context
Thibault “Tibo” Sottiaux leads Product & Platform at OpenAI, which puts him over ChatGPT, Codex, and the developer platform used to build on OpenAI models. In this conversation with Rana el Kaliouby, he explains how product teams work with researchers before a model ships, how OpenAI thinks about trust and alignment for agents, and why the company held back Astra 6.1 after a safety regression. For PMs, the episode matters because it shows what changes when AI compresses shipping timelines, but also how much more important explicit safety criteria, permission design, and internal debate become when the product can act on a user’s behalf.
The Big Idea
Frontier AI product management is about controlled autonomy, not maximum autonomy.
Tibo’s throughline is that more capable models only create product value when the team can define acceptable behavior, earn access gradually, and block launches when the safety bar moves backward. That mindset turns trust and alignment from abstract research topics into day-to-day product decisions.
Key Insights
Product shapes model capabilities
Tibo rejected the idea that research builds a model first and product simply wraps it later. He said product collaborates early on the capabilities the model should have, then helps define the data, evaluations, and product surface. His example was secure payments: if OpenAI wants that behavior, product and research work backward from the use case before training. For PMs, the lesson is that in AI products, capability design and product design are often the same conversation.
AI collapsed shipping cycles
- Tibo said DevDay plans come together only a few days before the event because the product changes too fast for a six-month plan.
- He gave a stronger internal example: OpenAI merged separate ChatGPT web and desktop codebases in 30 days, a job he said once would have taken 6 to 12 months.
- Tibo said the majority of the code for that effort was written by Astra. The PM implication is that planning, coordination, and release review all have to adapt to much shorter build loops.
Agent trust grows gradually
- Tibo’s advice was not to give an agent access to everything on day one.
- Start with limited permissions and cautious defaults. Ask for approval frequently, then expand autonomy as the user gains confidence.
- He framed this as the core trust equation: an agent’s utility is limited by its access, but broader access must be earned. That matters to PMs designing permissions, review steps, and onboarding for agentic products.
Safety gates must hold
Tibo said he was proud OpenAI delayed Astra 6.1 because of an ever so slight regression in safety and alignment evaluations. The important PM lesson is not the model name but the launch rule: once the team has declared a safety watermark, keynote pressure, engineering effort, and marketing value cannot redefine success. A guardrail that only matters when it is convenient is not a real guardrail.
Safety depends on stage
- Tibo separated AI work into training, evaluations, and deployment.
- During training and evaluation, models are being explored inside secure systems and may be tested on abilities that never reach users. That is different from a broadly deployed product.
- For PMs, this matters in incident review and stakeholder communication. Saying an issue happened during model research is not the same as saying a consumer product failed in the wild.
- Tibo’s point was not that research incidents are harmless. It was that each stage needs different controls and different language.
Automation should protect attention
One of Tibo’s clearest product beliefs is that AI should remove the work that keeps falling through the cracks, not automate for automation’s sake. He described using Dots to filter messages and surface only urgent issues, and he said OpenAI can use agents to watch latency or other performance curves 24/7 and alert humans only when something is wrong. For PMs, that is a strong design principle: use AI to preserve human attention for higher-leverage decisions.
Mental Models & Frameworks
Capability, data, eval loop
- Capability: start with the user or business behavior you want, such as secure payments or reliable computer use.
- Data: decide what training data or examples the model would need to learn that behavior.
- Evaluations: define how you will test whether the behavior is safe and reliable enough.
- Product: package the resulting capability into the right surface, workflow, and permission model.
Use this loop when working on AI features where the model’s ability and the product experience are tightly linked. It forces PMs to specify success before the team starts shipping wrappers around a vague capability.
Trust staircase for agents
- Step 1: limited access and approval for nearly every action.
- Step 2: broader read access or drafting authority, but a human still approves final actions.
- Step 3: more autonomous action in narrow, well-understood workflows.
- Step 4: wider autonomy only after the system has earned trust through reliability and clear controls.
Use this model when an agent touches private data, money, communications, or other hard-to-reverse actions. It reflects Tibo’s view that agent value increases with access, but access should expand only as trust does.
Three-stage safety funnel
- Training: researchers explore capabilities and techniques inside controlled environments.
- Evaluations: the team runs batteries of tests to see whether the model meets required behavior.
- Deployment: the model ships only if it clears the production bar for the intended audience and scale.
Use this framework to keep incident review precise. It prevents a common mistake where research risk, pre-launch failure, and live-user failure all get mixed into one vague safety conversation.
Layered defense stack
Tibo described safety as multiple defenses stacked on top of each other, not one magical barrier. In his example, that can include hardened sandboxes, online monitoring, and systems that stop a model if it gets close to breaking out of the allowed environment during training or evaluation. Use this model whenever a system can act autonomously. One safeguard will eventually fail. A stack gives the team more than one chance to catch the problem.
Attention preservation model
Treat AI as a first responder for noise, not a replacement for judgment. The system monitors, filters, and escalates. Humans decide what the exception means and what to do next. This is useful whenever a team is drowning in alerts, dashboards, and operational busywork that consumes attention but adds little strategic value.
Decision Principles
Principle: Delay on small regressions
- When: a launch has strong momentum, but a safety or alignment metric moves backward.
- Why: Tibo’s Astra 6.1 example shows that a release criterion only matters if it can stop a high-profile launch. If the bar changes under pressure, the team does not really have a bar.
Principle: Earn access before autonomy
- When: an AI feature needs access to private data, communications, money, or external tools.
- Why: more access increases utility, but it also raises the cost of failure. Gradual permissions and user approvals let the team learn where trust is strong enough to widen autonomy.
Principle: Define alignment explicitly
- When: a team says it wants a model to be aligned, safe, or on brand.
- Why: Tibo defined alignment as following specific values and instructions, and he pointed to OpenAI’s model spec as an example of writing those expectations down. A behavior standard that is not written clearly cannot be evaluated clearly.
Trade-offs & Nuance
Utility needs deeper access
An agent that cannot see anything or touch anything will not be very useful. Tibo said the utility of these systems is capped by how much access they have. The trade-off is that every new permission also increases the blast radius of mistakes. PMs have to design the slope of trust, not just the final destination.
Speed cannot outrun safety
Tibo clearly likes fast iteration. He talked about shipping better versions quickly and staying close to community feedback. But the Astra 6.1 delay shows the limit of that philosophy. Faster build loops are helpful only if the team can still stop, inspect, and say no when a critical guardrail regresses.
Open models need safeguards
Tibo argued that open source models create a new safety challenge because teams may not know what went into training them, yet they may still want to fine-tune and deploy them. His answer was not to reject the open ecosystem. It was to think about how a strong safety stack could support both OpenAI models and open models. The trade-off is flexibility versus provenance and control.
Automation versus human centrality
Tibo is explicitly not excited about automation as an end in itself. He wants AI to reduce stress, give people back time, and stop making humans adapt to screens and workflows built around software constraints. The tension is real: the same technology that frees attention could also centralize power or concentrate returns if teams optimize only for efficiency.
Common Mistakes
Mistake: Full access too early
Giving an agent broad access before it has earned trust creates a bad first experience and increases the cost of failure. Tibo’s view is to start cautiously, ask for approval often, and widen access only after the system proves reliable. That is a better path than forcing users into an all-or-nothing trust decision.
Mistake: Treating handoff as linear
A weak mental model for AI products is that research invents, then product packages. Tibo described a much earlier collaboration where product helps define desired capabilities, data, and evaluations. Teams that wait for a finished model may discover too late that the product needs and the model behavior never lined up.
Mistake: Chasing automation alone
Tibo repeatedly pushed back on the idea that the goal is simply more autonomy everywhere. He wants humans and humanity at the center, not as an afterthought. Teams that optimize only for what the system can automate may ship impressive demos that do not actually improve a user’s life or preserve user control.
Practical Application
Write capability-specific evals
Before the next AI feature review, write the expected behavior in plain language and list the failure cases that would block launch. Include both product failures and safety failures. This mirrors Tibo’s point that alignment only becomes real when the team can test against a clear spec.
Stage permissions by risk
Map each agent permission to the harm of getting it wrong.
- Low risk examples: summarizing, drafting, or monitoring.
- Medium risk examples: filing tickets, changing settings, or moving internal data.
- High risk examples: sending external messages, handling payments, or touching regulated data.
Then launch with the minimum level of access that still creates value, and require explicit approval at each higher-risk step.
Separate incident language
In every AI incident review, force the team to label the stage clearly: training, evaluation, or deployment. That one habit will make internal discussions more precise and will help executives avoid overreacting to research-stage findings or underreacting to live-user failures.
Use AI for exception monitoring
Pick one operational area where humans are wasting attention on routine checking, such as latency, failed jobs, support escalations, or broken integrations. Set up an AI workflow that watches the signal continuously and alerts the team only when the pattern changes in a way that matters. This applies Tibo’s idea that AI should handle the background vigilance so people can focus on judgment.
Build open dissent loops
Tibo praised OpenAI’s habit of visible discussion in Slack channels and direct Q and A with leadership. Product teams can copy the principle even without OpenAI’s scale.
- Create a public channel for launch concerns.
- Invite engineering, design, policy, support, and leadership into the same thread.
- Make it normal to question a release decision with evidence.
This improves safety and product quality because concerns surface earlier, while they are still cheap to fix.
Questions to Consider
- Where in your product would extra agent access create the most user value, and what approval step would make that extra access feel trustworthy?
- If your team defines alignment as following the product’s values and instructions, have you written those values and instructions clearly enough that an evaluation can fail them?
- Which workflows on your team still consume human attention only because nobody is monitoring them continuously, and could an AI watcher surface only the exceptions instead?
- When your organization discusses an AI incident, are people distinguishing between a training sandbox issue, an evaluation failure, and a deployed user-facing failure?
Bottom Line
The episode’s clearest lesson for PMs is that agentic AI product work is really trust design. OpenAI’s product approach, at least as Tibo describes it, combines early collaboration with research, gradual permissioning, explicit behavior specs, and the discipline to delay a launch when the safety bar slips.
Resources Mentioned
| Resource | Type | Why it was mentioned |
|---|---|---|
| OpenAI Model Spec | Company document | Tibo pointed to the model spec as the written standard for how aligned model behavior should look. |
Concepts to Explore
Alignment beyond refusal
Alignment is not just about stopping bad requests. In Tibo’s definition, it is about whether a model follows a specific set of values and instructions. PMs working on AI products should study alignment as a product behavior problem, not only as a safety-team problem.
Recursive self-improvement
Tibo described a practical version of recursive self-improvement where Astra helped OpenAI engineers build the next generation of inference infrastructure, including an ultrafast system he said was 8 times faster. The more advanced version would be a model designing its own next architecture. PMs do not need to solve that research question, but they do need to understand how faster AI-assisted development changes roadmaps, dependency planning, and release cadence.
World models debate
Rana asked whether true AGI requires richer world models, including physical, emotional, and social intelligence. Tibo said the debate is still open on whether those capabilities require explicit world models or whether they can emerge from generality and multimodality. This matters for PMs because it shapes which product bets look close, such as voice, robotics, and more context-aware assistants.
Case Studies Mentioned
Astra’s 30-day merge
Tibo said OpenAI merged separate ChatGPT web and desktop codebases in 30 days, and that the majority of the code was written by Astra. He contrasted that with a traditional timeline of 6 to 12 months for such a core-surface change. The product lesson is not just that coding got faster. It is that integration work once treated as major roadmap planning can now happen on dramatically shorter cycles, which changes how PMs sequence dependencies and review risk.
Astra 6.1 delay
OpenAI withheld Astra 6.1 from broad release because it showed a slight regression on safety and alignment relative to Astra’s earlier high watermark. Tibo called that decision something he was proud of. The lesson is that a release process earns trust when it can stop a launch that everyone would have preferred to announce.
Dots caught demo outage
Tibo’s favorite Dots example was personal and practical. Minutes before his DevDay keynote, his Dot messaged him that production was down and that the outage would affect the live demo. The team could not fix it in time, but they later reposted a better version of the demo. The product lesson is that the value of an agent is not only task execution. It is also context-aware prioritization that surfaces the one thing a human should care about right now.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| ChatGPT | OpenAI’s main consumer AI product | Tibo said ChatGPT is the biggest part of his scope and used it as an example of a product that must meet an extremely high deployment bar. |
| Codex | OpenAI product for coders and technical users | Tibo said Codex sits within his organization and used coding products as examples of how AI accelerates internal development. |
| Astra | OpenAI model discussed throughout the episode | Tibo used Astra to explain computer use, internal engineering acceleration, alignment, and the decision to delay Astra 6.1. |
| Dots | OpenAI’s new suite of agents | Dots was the main product example for trust, permissions, context-aware assistance, and reducing the need to constantly check every channel. |
People to Follow
Thibault “Tibo” Sottiaux
Head of Product & Platform at OpenAI. He is useful to follow if you want to understand how a frontier AI company connects model capability, product design, trust, and launch discipline.
Rana el Kaliouby
AI scientist, entrepreneur, and the host of Pioneers of AI. She consistently pushes guests toward the human-centered implications of AI, which makes her especially useful for PMs thinking beyond raw capability.
Notable Quotes
“It’s proof of a system working.” (Thibault “Tibo” Sottiaux)
“I don’t want a product that does unwanted or unexpected things.” (Thibault “Tibo” Sottiaux)
“The utility that you get from these systems is kind of capped by, you know, how much access it has.” (Thibault “Tibo” Sottiaux)
