When I asked Tim Shi, co-founder of Recursive, how AI might develop scientific taste, he pointed to the history of deep learning. “If you think about this whole field of machine learning, back in the day there were many people who thought deep learning was a terrible idea and you shouldn’t pursue it, but there was this small group who kept working on deep learning anyways.” In hindsight, we know that persistence was worthwhile.
As we build AI systems that can choose their own research directions, how do we recognize what’s really promising, especially when the near-term value isn’t clear? And how can AI learn from experts while still testing ideas those experts think won’t work?
Tim is focused on trying to answer these questions. At Recursive, he’s building AI systems that can move beyond what they learn from humans. Tim previously co-founded Cresta in 2017, where the idea was to learn from the best customer support agents and make their expertise available to everyone else. Before that, he worked on reinforcement learning at Stanford and OpenAI.
“There are directions that humans haven’t explored yet because humans have biases,” he told me. “So AI has to be much better than the best experts.”
Recursive’s goal is to automate scientific discovery, starting with AI research itself. The vision is for a system to conduct research that helps produce better models, which can then help it become a better researcher. In our conversation, Tim and I discuss what it means to make this work and what’s needed at each key step: choosing what ideas to investigate, deciding what counts as progress, and knowing when something deserves more attention.
What “self-improving” means
“Self-improvement can mean many different things,” Tim told me. It can mean that a model improves through training against an objective and a dataset that it was given by humans.
It can also mean that it learns iteratively through experience, trying different things and then getting feedback about what did and didn’t work. “You’re learning from experience. You’re not learning from predefined data,” he said.
A more ambitious form of self-improvement would mean a system could actually decide which objectives to pursue. Scientific research often changes direction when there’s new evidence and a result’s significance isn’t always clear right away. Therefore, a system that only rewards near-term progress toward a fixed goal risks overlooking those possibilities.
This is where Recursive is focused. Their goal is to automate the scientific process through which these seemingly unobvious discoveries happen. “The original idea was to build an automated AI researcher that can automate the scientific method,” Tim said. To do this, the system needs to propose hypotheses, implement experiments, evaluate the results, and then use what happened to decide what to try next.
If a research system helps produce more capable models, those models could, in turn, help the system do better research. That is the recursive process the company wants to develop. Getting there will be a gradual progression, “starting with assisting the expert human researchers,” Tim said, and eventually “automating a lot of the scientific discovery.”
The challenge of scientific taste
As I wrote about here in my conversation with Mihai Codreanu, choosing what to test is getting harder as AI makes it easier to generate new ideas. For example, feasibility vs expected value. One hypothesis might be easy to test but scientifically incremental whereas another might be much more uncertain (and expensive), but potentially a lot more valuable. Choosing between this tradeoff means using judgment about what the likely result will be and its impact.
Today, Tim sees human researchers playing an important role in directing agents, drawing a parallel to the role that a principal investigator plays in directing a group of graduate students. “A lot of the research tasks can be completed autonomously, but you still need taste and direction,” he said.
Over time, he hopes systems will become more capable of choosing their own directions. Part of this means exploring ideas that people have prematurely ruled out. For example, during the development of Flint, Recursive’s model for coding on local hardware, the automated research system tried quantizing the model’s output layer, reducing the numerical precision used to represent its weights. Human experts have typically been skeptical of this approach because of the expected loss in capability. But in this case, the system tried it and validated the result. “Human experts would think that idea would not work, and then AI proved it actually works,” Tim said.
Here the system pursued a novel idea and produced a result that could be evaluated against clear goals. But what happens when there’s no immediate improvement from an experiment, but it could become a stepping stone to a bigger discovery? How does a system decide whether to keep pursuing that direction?
“That’s taste and it’s a really hard question,” Tim said.
Different researchers make different judgments about what is worth pursuing. Of course some of those will be wrong, but it’s only by having many different directions that you can actually discover something unexpected. “That’s why you want to simulate a community of scientists,” he said.
Over time, Tim thinks research experience could help models develop better taste. “Human scientists train in the PhD program, in research positions, for many, many years to develop a taste for what is a good problem to work on,” he said. “I think the model will over time develop that.”
The goal is to encourage “curiosity in the model” and have it “explore a wider space of possibilities.”
Defining what “good” looks like
By giving agents a wider space of possibilities and more freedom to experiment, evaluation becomes harder. What if it’s not initially obvious that a result could lead to something bigger? Or what if the system found a solution, but also a way to exploit the measurement?
Tim described experiments in which reported accuracy rose above 100%. He suggested that agents might be changing the evaluation or finding ways around the experimental constraints. Aka: reward hacking.
Recursive built monitoring systems to help detect this. “It’s AI systems that monitor other systems, but they have to embody the human value,” Tim said, which can be hard. Any metric captures only some of what people care about. “When you define a metric, sometimes you actually miss some of these important components where a human will naturally think what is good or bad.”
At Tim’s previous company, Cresta, defining success often meant having many conversations with customers. Cresta first needed to understand what an organization wanted to improve and how it would recognize a good outcome. “That rubric takes time to define because the customer themselves may not know. The evaluation can be very nuanced depending on the kind of problem you’re solving for.”
This is likely to continue to play a very significant role in what it means to deploy recursively self-improving systems. People need first and foremost to really understand their priorities. Then they need ways to express those priorities and understand whether an evaluation fully encapsulates them.
“Defining that rubric is going to be the key component of RSI,” he told me.
Some of that knowledge is hard to articulate in advance. This is another area where human judgment has to be part of the research system. “You can’t just read a book about playing chess,” Tim said. “You have to learn from the masters. Expertise is very situational. You know what to do in what scenario.”
The infrastructure to test new ideas
As I wrote about here and here, the bottleneck is moving downstream. There’s a big difference between having more hypotheses and having more discoveries. As research systems become more capable, the capacity to test what they propose becomes incredibly important.
“Validation is actually sometimes easier than the actual generation,” he said. “Today, a lot of math can be validated autonomously. If you build a plane that wants to fly at a certain speed and height, you can validate it in simulation or in actual tests. And we can test if a drug works by doing clinical trials.”
But validation still takes time and resources. And depending on the field, that can be very costly. “Validation is expensive and you have to pick and choose which idea to validate,” Tim said.
This is one reason why Recursive is starting with AI research. Many of the experiments can be run entirely on computers. “You can buy more compute and you can run more experiments. Whereas when you get to biology and physics, it takes longer to validate and you have to build physical infrastructure,” Tim said. The scientific method applies across domains, but the practical work of validation changes. Collecting evidence about safety and effectiveness takes time. As Tim put it, “AI will increasingly get good at generating theories and validating theories, but it’s limited by the infrastructure to run experiments fast enough.”
I asked Tim whether labs need to be fully autonomous for recursive self-improvement to work. He said not necessarily because AI systems can propose experiments that people can help carry out. In fact, over time, he expects humans to play a bigger role in building the capacity to test ideas that AI generates. “I don’t think human scientists will be out of a job. I think we’ll shift towards building the infrastructure,” he told me.
That said, this doesn’t mean the supply of useful ideas from AI is sufficient. “Ideation itself can be a bottleneck,” he told me. “There’s not enough good ideas being generated.”
And even the notion of a “good” idea is often unclear in the beginning. The history of deep learning is a reminder of why it’s so important to sometimes pursue unpopular directions. AI researchers will need room to pursue unconventional ideas, as well as the experimental capacity to find out where they lead. This is Recursive’s goal, and it depends on enabling different researchers to follow different intuitions, test them, and build on what they discover. As Tim put it, “You never know what will actually drive the fundamental progress in science.”
Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.


