Supermartingales

The Doomsday clock now sits at 89 seconds to midnight, and the recent wave of AI developments has shaved off at least one more. I promised myself this would be my last post on the subject, and I intend to keep that promise — but the occasion demands one more entry.

I will start with some technical grounding. For all times $s \leq t$, a supermartingale is a stochastic process satisfying

$$\mathbb{E}[X_t \mid \{X_{\tau}\}_{\tau \leq s}] \leq X_s.$$

Intuitively, the expected future state of the process is no greater than its present state — things tend to get worse over time, or at best stay the same. A martingale is the fair version: expected future equals present. A supermartingale is the unfair version: the drift runs against you. Keep this in mind.

In late 2026, a preprint claiming a solution to the Navier-Stokes existence and smoothness problem — one of the seven Millennium Prize Problems — began circulating, with the authors crediting a significant portion of the proof search to an AI system. The mathematical community’s reaction was, to put it charitably, uneven. Some celebrated it as the dawn of a new era. Others reacted with what I can only describe as existential fury: if a machine can crack Navier-Stokes, what exactly are we doing here?

I want to sit with that fury for a moment, because I think it is revealing. The anger is not really about the proof — nobody is upset that Navier-Stokes might be solved, that would be wonderful. The anger is about credit, and credit is only precious if you have decided that mathematics is fundamentally a competition for it. Follow that logic and you arrive somewhere uncomfortable: if your identity is tied to being the one who produces the result, then your expected future self-worth — conditioned on AI getting strictly better — is itself a supermartingale. It can only go down. Every new capability is another tick of the clock, another second shaved off the Doomsday clock of your relevance.

This framing contains a hidden assumption I argued against in my last post: that a mathematician’s value is their output. Reject that premise, and the entire Navier-Stokes discourse looks completely different. The question is no longer “did a human prove it?” but “do we now understand it?” Understanding — the slow, personal, occasionally maddening process of internalizing why something is true — is not something a model can do for you. It can hand you the proof. It cannot hand you the comprehension.

The mathematicians who seem most at peace right now are the ones who fell in love with the process before they ever cared about the product. The ones who spent a Saturday afternoon genuinely delighted by a clever change of variables, not because it would appear in a paper, but because it was beautiful. For them, AI is just another tool — occasionally useful, occasionally impressive, never a threat to the thing they actually care about.

And here is the honest conclusion I keep arriving at: life, in the most accurate accounting, is a supermartingale. Things tend not to improve on their own; entropy wins, credit gets misattributed, Millennium problems get solved by machines, and the Doomsday clock never runs backwards. The process is unfair by construction, and no amount of productivity changes that. But this is precisely why the premise — that our worth is our output — is so dangerous. If the game is rigged, optimizing your score is a losing strategy from the start.

The optional stopping theorem tells us that a supermartingale stopped at the right time has a perfectly controlled expected value. I think that is the only honest response to all of this: not to beat the unfairness, but to choose deliberately when to stop letting it be the thing you measure yourself by. Once you accept that the game is rigged, you are finally free to play it for the right reasons.