NederlandsSelf-improving AI
Self-improving AI is AI that helps make itself or the software around it better, for example by designing experiments and improving code. The idea is that progress could speed up as a result.
Self-improving AI is AI that takes part in improving AI. It can start small: a model devises an experiment to improve another model, or it writes code that makes a system faster. The idea behind it is a feedback loop. If AI gets better at making AI, the next generation can improve faster still. This is also known as recursive self-improvement.
How does it work?
AI research consists of a number of recurring steps: working out what you are trying, carrying it out, reading the outcome and learning from it. Each step can be handed partly to AI. A model can come up with hypotheses, set up experiments, write and test code, and use the results for the next round. The young company PhAI Labs built such a team of AI agents, which comes up with research ideas, tests them and keeps track of what worked. The system learns from that record: after two rounds its hypotheses were better (link).
This concerns pieces of the process, not yet the whole. Humans still decide a lot: which question is worth asking, and what happens with the outcome.
Why it matters
Self-improvement is one of the ideas behind the road to the singularity: AI that improves itself could speed up progress. That is why it is important to measure how far AI has got with this.
On the TasteVal test from P-Zero Research, Claude Opus 5.5 designed better AI experiments than the best of 24 experienced AI researchers, using less than half the computing time (link). The test measures only part of the research: the question itself is already fixed. There are also examples from practice. The Chinese company Unitree is working on Rehearsal: AI reads research itself, writes robot code, tests it in a simulated world and then puts it on real robots (link). Steps like these show where development stands now and what the next steps are.
Articles about Self-improving AI
-
Research group METR: Anthropic is probably improving its AI models twice as fast as before
METR spent eleven weeks inside Anthropic and estimates that the company is now improving its models about twice as fast, probably mainly because AI helps with the research. The uncertainty is large and much of the evidence stays secret.
-
Tencent lets AI build its own tools for protein research and mathematics
Tencent has presented the second version of its research AI, Hyra. It delivers not only solutions but also tools that can be used again later. Models for proteins and molecules are already public, so other researchers can now test them themselves.
-
Newest AI model can read the inside of other AI and write it itself
Researchers at MIT had OpenAI's GPT-6 Astra read and adjust the numbers that small AI networks are made of, without any training. It even wrote networks that play a computer game well. The researchers call these very early results.
-
Unitree lets AI improve the software of its robots by itself
Robot maker Unitree is working on Rehearsal: AI reads research by itself, writes robot code, tests it in a simulated world and then puts it on real robots. Founder Wang Xingxing now expects robots to handle most tasks on command in 2027 or 2028.
-
AI model designs experiments better than 24 experienced AI researchers
On a new test by P-Zero Research, Claude Opus 5.5 designs better AI experiments than the best of 24 human experts. The model needs less than half the computing time to do so and costs about one thirtieth as much.
-
Free coding model for the laptop was largely developed by AI itself
The AI company Recursive has released Flint, a free AI model that helps with programming and runs on an ordinary laptop. According to the company, its own research AI did a large part of the work: a first step towards AI that improves itself.
-
AI researcher learns from its own work and comes up with better and better hypotheses
The young company PhAI Labs built a team of AI agents that comes up with research ideas, tests them and keeps track of what worked. The system learns from that record, and its hypotheses improved after two rounds. The code and model are publicly available online.
