Improve the way improvement works.
Obinka researches recursive self improvement: how software systems can change the methods they use to improve, and how those changes can be evaluated.
For this work, the recursive part concerns a system changing the methods it uses to make further improvements. We do not assume that each successive change will help.
Conceptual method
Not a result
- Candidate change A proposal, with the part of the system it would alter stated up front.
- Evaluation Comparison against a baseline on tasks the proposal did not choose.
- Record Outcome, uncertainty, cost and failed attempts kept together.
- Keep, revise or reject A change can also remain inconclusive. The figure does not promise progress.
01 / Why this work
Better methods for building better software
Our aim is to build useful technology and study the methods behind it. The research begins with a practical question: what would count as a real improvement, and what evidence would justify carrying it forward?
02 / Questions we want to answer
-
Question 01
Can the improver improve?
When a system changes its own tools or strategy, does it become better at making the next useful change?
Read the question -
Question 02
How should improvement be tested?
Repeated feedback can shape what a system learns to optimise. Which evaluation methods remain informative through that process?
Read the question -
Question 03
Which gains are useful in practice?
Quality matters alongside time, cost, reliability and the work left for people. Can methods be judged against those constraints?
Read the question -
Question 04
What transfers into a product?
A promising result is a starting point. Which methods help with the requirements of a working product?
Read the question
03 / Approach
Define the change. Examine the evidence.
Our proposed approach starts with a defined task and a clear account of what a system is allowed to change. Candidate methods can then be compared under stated conditions, with attention to failed attempts, cost and limits as well as successful outcomes.
04 / Current stage
An early research direction
We are setting out the questions behind Obinka's work. Research notes and results will be added as there is work to share. There are no published Obinka results yet.
05 / Engineering foundations
Tools that make the questions concrete
Our product work includes Acoolit, an evaluation framework; Jecoha, a local coding harness; and SlopChopper, static analysis for AI code. Paretomax is our planned lab for comparing and evolving agent harnesses. These tools give the research a practical focus. All are in development.
06 / Collaborate
Working on a related question?
If you study self-improving systems, agent evaluation or the tools used to build them, we would like to hear what you are working on.