Noumenal AI

Causality. Consequence. What-if.
Everything after correlation.

[Noumenal AI @noumenal_world]

These are not separate problems.

A recursively improving system eventually attacks every assumption around it:

its metric,
its generalization bound,
its model of the world,
its own implementation.

RSI is what remains valid after the learner becomes part of the theorem.

[Noumenal AI @noumenal_world]

Accepted across #COLM and multiple #NeurIPS workshops

Our research program: infrastructure for recursive self-improvement.

Measurement that survives optimization.
Generalization that survives self-rewrite. Counterfactuals under intervention.
Formal learning theory that actually compiles.

@ayushmi5 @Zetetic_Dhruv

[Noumenal AI @noumenal_world]

Formal Learning Theory in Lean 4 - AutoMLR oral spotlight

Before rebuilding learning theory for systems that recursively modify themselves, we wanted the old foundations machine-checked.

So we formalized the Fundamental Theorem of Statistical Learning - five ways.

Paper + code coming soon.

[Noumenal AI @noumenal_world]

Paper: https://t.co/haLVlvrW66
Code: https://t.co/VO7q0UvPwN

[Noumenal AI @noumenal_world]

Recursive self-improvement is not an optimization problem until it first becomes a measurement problem.

An optimizer that understands its metric is already part of the measurement apparatus.

We’ll be presenting where that can be made precise at the Lifelong Agents workshop at #COLM2026

[Noumenal AI @noumenal_world]

The larger question for us:

How do you build a machine that can rewrite itself while preserving a notion of "better" that it cannot rewrite with it?

That feels very close to the real starting point of #RSI.

[Noumenal AI @noumenal_world]

@ylecun @Dan_Jeffries1 Pro-intelligence. But intelligence without the ability to understand consequences is just increasingly powerful prediction.
The real question is whether we can make intelligence that understands the world well enough to know what its actions will cause.

[Noumenal AI @noumenal_world]

AGI was taking too long.
So we started reverse-engineering reality instead.

[Noumenal AI @noumenal_world]

Somewhere between Bayes and God, there is an API. We are looking into it.