Imagine three research teams working late into the night.
The first team is trying to discover a new battery material. An AI system reads the literature, proposes candidate compositions, predicts which are most promising, and suggests the next experiment.
The second team is trying to build a better AI model. Their AI system inspects training code, changes an architecture, adjusts hyperparameters, runs an experiment, measures the result, and proposes another change.
The third team is not defined by what it studies at all. Its system simply follows one principle: try something, evaluate what happened, use the feedback to improve the next attempt, and repeat.
At first glance, all three may look like the same story: AI doing research.
But they represent three different ideas that are often mixed together:
- AI for Science (AI4S)
- AI for AI (AI4AI)
- Recursive Self-Improvement (RSI)
Understanding the difference matters because these terms answer different questions.
Prefer to watch? Continue with the video below, or keep reading.
AI4S: When AI Becomes Part of Scientific Research
AI for Science means using AI to help create scientific knowledge.
That can happen at many levels.
At the simplest level, AI may act as a research assistant. It can search papers, summarize literature, extract data, write analysis code, or help interpret an experiment.
At a more ambitious level, it may propose hypotheses, rank possible explanations, design experiments, control instruments, analyze the resulting data, and decide what to test next.
The key idea is that the target is a scientific problem.
That target might be a protein, a catalyst, a material, a weather system, a physical mechanism, or a mathematical conjecture.
So when we say AI4S, the question is:
Can AI help us understand nature or create new scientific knowledge?
This is broader than simply applying machine learning to scientific data. The long-term vision is an AI system that participates in the scientific method itself.
AI4AI: When AI Studies AI
Now change the object of study.
Instead of asking AI to discover a molecule, ask it to improve an AI system.
Suppose we have a large model that is expensive to train. We want the best possible performance using a fixed amount of compute.
An AI agent might inspect the training code and propose:
Increase this parameter.
It modifies the code, launches a training run, and measures the result.
The performance improves.
So the system makes a second change.
That one fails.
It learns from the failure, tries another direction, and gradually explores the space of possible improvements.
The scientific pattern looks familiar:
\[ \text{proposal} \rightarrow \text{experiment} \rightarrow \text{evaluation} \rightarrow \text{new proposal}. \]
The difference is that the object being investigated is now AI itself.
This is AI for AI.
The possible targets are broad:
- model architectures,
- training algorithms,
- optimizers,
- data-selection strategies,
- inference methods,
- memory systems,
- agent workflows,
- compiler techniques,
- hardware-aware algorithms.
Google DeepMind's AlphaEvolve offers a useful illustration of the general idea. It combines large language models with automated evaluators and an evolutionary search process. Candidate programs are generated, tested, scored, retained or rejected, and then used to produce new candidates.
Another example came from Sakana AI's AI Scientist in 2024. In machine-learning research, the system was designed to generate research ideas, write and modify code, run experiments, analyze the results, and draft a scientific manuscript. It remains a bounded research framework rather than an autonomous general scientist, but it makes the AI4AI idea concrete: the AI can participate in the research cycle of improving AI systems themselves.
The important point is not the particular system.
It is the structure:
AI generates an idea, the environment evaluates it, and the result feeds the next round.
That brings us to RSI.
RSI Is Not a Field
Recursive Self-Improvement sounds more mysterious than it needs to.
It is often discussed in connection with AI systems improving themselves, but the underlying idea is simpler.
A system performs a task.
It examines the result.
Then it uses what it learned to improve the next attempt.
The word recursive emphasizes that the output of one round becomes part of the input to the next.
We can write the pattern as
\[ S_0 \rightarrow A_0 \rightarrow F_0 \rightarrow S_1 \rightarrow A_1 \rightarrow F_1 \rightarrow \cdots \]
where $S$ is the current state, $A$ is an action, and $F$ is feedback.
RSI therefore describes a method, not a scientific domain.
It can be used in AI4AI.
It can also be used in AI4S.
A self-improving coding agent may use RSI to make better software.
An automated chemistry laboratory may use the same general logic to select better experiments.
The object changes. The loop remains.
In a scientific setting, recursive self-improvement need not mean that an AI rewrites its own source code. A more important form is recursive model improvement: the agent proposes an experiment, observes the result, updates its internal model of the system, and uses that revised model to choose the next experiment. When repeated autonomously, this becomes part of what is often called closed-loop discovery.
A Simple Way to Keep the Terms Straight
One useful distinction is:
- AI4S asks what AI is researching: science.
- AI4AI asks what AI is researching: AI itself.
- RSI asks how the system improves: repeated feedback and self-modification.
In other words:
\[ \boxed{\text{AI4S and AI4AI are targets}} \]
while
\[ \boxed{\text{RSI is a process}}. \]
Of course, real systems can combine all three.
Imagine an AI system helping to discover a new quantum algorithm.
It is doing AI4S because the goal is scientific and technical discovery.
It may also modify its own search strategy or code-generation process, which introduces AI4AI.
And if the system repeatedly evaluates its own results and improves how it searches, it is using an RSI-like loop.
The boundaries are conceptual, not walls.
Why Feedback Is the Common Ingredient
What makes all three ideas powerful is feedback.
Without feedback, an AI can generate.
With feedback, it can improve.
Suppose an AI proposes 1,000 possible catalysts.
If no one can evaluate them, we simply have a long list.
Now suppose a simulation can quickly score each candidate.
The system can keep the best 20, generate variants, test them, and repeat.
The same pattern appears in program optimization.
It appears in mathematical search.
It appears in automated experimental design.
The quality of the feedback determines the quality of the loop.
This is why mathematics and programming are especially attractive environments for self-improving systems. A proof or program can often be checked relatively quickly.
The physical sciences are harder because feedback may require an experiment.
And an experiment can be slow, costly, noisy, or impossible to automate.
The Difference Between a Model and a Research System
This distinction also helps explain why an excellent language model is not automatically an AI scientist.
A model may know an enormous amount of chemistry.
It may answer difficult questions.
It may suggest creative hypotheses.
But scientific research requires more than knowledge.
A research system must connect several capabilities:
\[ \text{knowledge} + \text{reasoning} + \text{tools} + \text{memory} + \text{experimentation} + \text{feedback}. \]
The model is one component.
The loop around the model may be just as important.
That loop determines whether the system can take an idea, confront it with evidence, and change course.
This is why agent design has become so important to AI4S.
The question is no longer only, “How smart is the model?”
It is also:
What can the model observe?
What actions can it take?
How is success measured?
What does it remember?
How does it use failure?
Self-Improvement Does Not Guarantee Discovery
There is another important caution.
A system can improve endlessly against the wrong objective.
Imagine an AI that becomes extremely good at optimizing a scientific benchmark.
Its score rises every week.
But the benchmark rewards a shortcut that has little connection to the real scientific question.
The system is improving.
It is not necessarily discovering.
The same issue appears in automated experimentation. If the objective function is poorly chosen, a highly efficient optimizer may simply find increasingly clever ways to satisfy the wrong criterion.
So an RSI loop is not magic.
It needs a meaningful evaluator.
And in science, deciding what counts as meaningful can itself be difficult.
Three Ideas, One Larger Story
AI4S, AI4AI, and RSI are useful because together they describe a shift in what we expect from AI.
The first generation of scientific AI largely performed isolated tasks.
Predict this structure.
Classify this image.
Fit this curve.
The emerging generation is more ambitious.
It asks whether AI can participate in a continuing process of knowledge creation.
AI4S provides the goal.
AI4AI may improve the machinery.
RSI provides one mechanism by which the process can become better over time.
But all three eventually meet the same question:
Can we build a closed loop in which an AI proposes, acts, observes, learns, and tries again?
That is where the next blog begins.
Next in the series: The Scientific Loop: From Hypothesis to Feedback
0 comments