Our mission

Language models are starting to help with research, including research on AI. They already read papers, write code and draft arguments, and the Serre platform uses them to read a field and judge a research direction.

Research on AI is research too. As models get better at research, they get better at the research that improves the next models, and those help again. We call that loop the self-improving machine, and hold it to one condition: every step checkable by people.

Science that moves faster. Most of a research project is not the idea. It is reading what is already known, finding out whether the idea is new, designing the experiment that would decide it, and writing up what happened. Each of those can take weeks of a skilled person's time, and a model that reads and writes quickly can help with each.

When that time shrinks, more questions get tested in a year, and a wrong idea is dropped sooner. We are optimistic about that, and we work on it directly.

What goes wrong in a loop is worse than what goes wrong in a single answer, because each turn builds on the last.

Confident errors

A model can give a wrong answer in the same fluent voice as a right one. In a single answer that is a mistake; in a loop it becomes the premise of the next turn.

Claims nobody can check

A result that arrives without its evidence can only be trusted or ignored. When results come faster than anyone can trace them, trust becomes the default.

Lost oversight

If the loop turns faster than people can follow, decisions get made that no person saw. Then nobody can say where a mistake entered, or steer away from it.

Checkable at every step. A loop nobody can check is a loop nobody can steer. We think the loop is worth running only if each turn leaves evidence a person can check before the next begins.

That takes two things. Every claim carries what it rests on: the passage it quotes, the evidence behind each decision, and who made it. And checking has to stay cheaper than producing, or people will stop doing it as the loop speeds up. The first is how we build the Serre platform. The second is an open research question, and we work on it.

Verification research

When can an answer be checked more cheaply than it can be produced? The answer decides how fast a checked loop can turn.

Evidence in the Serre platform

A claim taken from a paper carries the exact passage it rests on, and is rejected if that passage is not in the paper's text. Each decision is logged with the evidence it rested on and who made it, in a record that is only ever added to.

A person decides to commit

The platform judges a research direction on the evidence, but committing to it takes a person's decision, logged with the evidence that person saw. The platform cannot make it alone.

Open specifications

Open specifications anyone can read, implement and check, each with its own public repository, licence and governance. The first is T32: a compact universal timestamp.

This page says what Serre AI does today. What follows is not done, and we say so.

The full research cycle

The Serre platform reads a field, judges a direction and plans its submission. Running experiments, writing papers and handling reviews are planned, not built.

Results

Whether checking can stay cheaper than producing is the question our verification research asks, not something we have shown. Our first publications are in preparation.

Verification

When an answer can be checked more cheaply than it can be produced