Advisory boards are now infrastructure.
That sounds like a strange thing for someone who spends his days wiring up agents and debugging tool calls to say. But look at what OpenAI has been assembling. In 2026 the company announced that one of its AI systems resolved key parts of the Navier–Stokes equations, a problem that has been open for about 90 years. It also formed an advisory council on wellbeing and AI. Separately, it created a safety advisory group for generative AI, alongside giving its board veto power over releases.
Those three things get covered as unrelated news items. From where I sit, building bots for a living, they are the same story told twice: as capability, and as governance. And both halves have something concrete to teach anyone shipping an agent this quarter.
A math result is a verification story
Navier–Stokes describes how fluids move. It has resisted a full proof since the 1930s, which is why “resolved key parts” is the careful phrasing everyone is using. Not solved. Parts.
That distinction matters more than the headline. Mathematics is the one domain where we have a mechanical way to check whether a machine is lying to us. A proof either holds or it does not. There is no “seems plausible, ship it.” Compare that to the output of the average bot you and I build, where correctness is a vibe check against a handful of test prompts.
Not everyone is convinced, either. Critics have already published pieces arguing the math announcements amount to research misconduct. I am not in a position to adjudicate that, and I would not try from a blog post. But the fact that the argument is happening at all is itself instructive. When an AI produces a claim, the claim and its verification are two separate artifacts, and the second one is the expensive part.
What that means for your pipeline
If you are building anything that produces answers people act on, steal the mathematician’s habit:
- Separate generation from checking. Different model call, different prompt, ideally a different mechanism entirely. A deterministic validator beats a second opinion from the same model.
- Prefer domains with checkable output. Code that compiles and passes tests. SQL that returns rows. JSON that matches a schema. If your bot’s output cannot be mechanically checked, you are the checker, forever.
- Log the reasoning trace, not just the answer. When someone disputes a result six weeks later, the trace is the only thing that lets you figure out whether the system was right for the wrong reasons.
- Say “parts” when you mean parts. Precision in your own release notes buys you credibility you will need later.
Two advisory groups and a veto
The governance half is less glamorous and probably more transferable. A wellbeing council. A generative AI safety group. A board that can block a release. Strip away the corporate scale and what you have is a review layer that sits outside the team doing the building.
Every bot builder I know eventually reinvents some version of this, usually after an incident. The support bot that told a customer something expensive and wrong. The agent that had write access to a production table because it was convenient during development. The chat assistant that handled a user in genuine distress about as gracefully as a vending machine.
You do not need a council for a side project. You do need the function that a council performs: somebody whose job is to ask what happens when this goes badly, and who is not the same person racing to ship it. On a solo project that can be a checklist you actually fill out. On a team of five it can be one person per release wearing the skeptic hat.
The wellbeing part is a product requirement
Of the two groups, the wellbeing council is the one I would pay closest attention to as a builder. Skeptics have fairly asked whether the company will listen to the experts it recruited. Fine question. But the underlying admission is the useful bit: how people feel while using your bot is now treated as a design surface, not a side effect.
If you ship conversational interfaces, you have already made wellbeing decisions whether you meant to or not. How your bot responds to frustration. Whether it encourages someone to keep talking to it instead of a human. Whether it knows how to end a conversation. Those are architecture choices, and they belong in your design docs next to your rate limits.
What I am actually taking away
The math result tells me the ceiling is moving, and that the interesting work is in verification rather than generation. The advisory groups tell me the organizations closest to the frontier have decided review layers are not optional overhead.
Both point the same direction for those of us building smaller things. Make your bot’s output checkable. Give someone the job of doubting it. Neither requires a research budget, and both will save you a bad week.
đź•’ Published: