Before we know what matters.

AI gives me new ways to write code, try ideas and run experiments. With Mutome, I want to use those abilities to explore questions whose importance we do not yet know. I also want each experiment to leave a better starting point for the next.

An unexpected pattern can be worth examining before we know what it means. The first useful result may be a question precise enough to test. Finding that question, comparing possible explanations and choosing what to measure next is the work I want to do.

Choosing where to look.

I first looked at famous open problems. We already know why their answers matter. But their prestige also draws attention and resources. I do not want winning that race to be the condition on which Mutome depends.

Recent announcements sharpened that choice for me: OpenAI’s Navier–Stokes announcement and mathematics report, alongside Anthropic’s Fermat formalization project. These are different kinds of work. I am citing the reports that prompted my reflection, not independently validating their results.

Large labs can explore unfamiliar questions too. An obscure subject is no protection from competition, and obscurity alone is no reason to study it. My freedom is to follow a question before there is agreement about its value.

Finding questions is not enough.

Tao’s discussion of AI and mathematical research, and his closing post on what surrounds a solution, helped me put this into words. My takeaway is that deciding what to study deserves as much attention as finding an answer.

We can invent a strange equation in seconds. That is not necessarily a good reason to spend a year on it. I am more interested in an unexplained pattern, a connection between ideas or an observation that existing models struggle with.

Does an explanation make a different prediction? Can an accessible experiment check it? What would its outcome change about our next step? These questions help choose where to spend effort without pretending to score a field’s future importance.

Not knowing the importance in advance.

If every experiment must promise a major discovery before it begins, we tend to follow subjects already recognized as important. Some observations become useful only when another piece of work gives us a way to understand or apply them. Others lead nowhere.

We may not know what a result will be useful for. We should still know how it was obtained, the conditions in which it holds and what could overturn our interpretation. Uncertain importance does not have to mean uncertain evidence.

An experiment can leave a counterexample, a reusable measurement tool or a reproducible behavior even when it does not answer the original question. I want Mutome to preserve those results with their limits, so a later question can make them useful again.

That does not mean treating every saved idea as knowledge. A note needs a reason to survive: an effect reproduced, an explanation ruled out under stated conditions, or a tool another experiment can use. Keeping a conversation is easier than establishing any of those things.

Putting ideas to the test.

I want Mutome to propose explanations, compare their predictions and use observations to shape the next question. Memory supports that work by carrying forward what has been established. It cannot replace the work of testing a new explanation.

If two models explain the same data, we need to find where their predictions differ. A closer fit to old measurements does not prove we have the right explanation. Sometimes the useful conclusion is that the available experiment cannot distinguish them.

OpenAI’s account of research acceleration also informed this reflection. I want to count the human work, failed attempts and verification behind a result, rather than equate more activity with more progress.

Simulation gives us an inexpensive starting point. It lets us test a method against a world whose rules we know. Success there does not establish a law of nature. That requires confronting explanations with real observations and their uncertainties.

What counts as progress.

A plausible formula is a candidate. To claim an improvement, it must beat a relevant comparison under the stated conditions and survive independent tests. A well-established negative result can also advance research by showing what fails, where and why.

Our first oscillator campaign produced candidate formulas but did not reach a conclusive evaluation. That exposed a gap in the process: exploration could consume the resources needed for verification. It taught us something about running the lab, not whether the proposed formulas were right.

The next demonstration I want is concrete: establish a result on one problem, retrieve it when it is relevant to a separate problem, and test whether it helps. A lower search cost or a justified decision to stop could matter as much as a better final score. We must choose the criterion before the test.

A better starting point.

This is the ambition I want to build Mutome around: explore questions whose importance we do not yet know, while making each experiment a better starting point for the next. That cumulative benefit remains something to demonstrate, not a capability I can already claim.

Many attempts will lead nowhere. Some will narrow the possibilities. A few may reveal a question we would not otherwise have thought to ask. I want to be able to follow those questions, and to explain what we learned along the way.