Yuandong Tian on RSI: Self-Improvement Starts With a Working Research Loop

Original interview: 178: A conversation with Yuandong Tian about RSI and the arrival of model self-improvement

LateTalk by LatePost · August 7, 2026 · 1 hour 29 minutes 02 seconds

Guest: Yuandong Tian, co-founder of Recursive Superintelligence · Host: Manqi, technology editor at LatePost

Primary materials: Recursive: First Steps Toward Automated AI Research · Open-source run artifacts · Anthropic Institute: When AI builds itself

RSI carries two meanings in this interview. The concept is recursive self-improvement, where AI contributes to improving a future AI system. It is also the abbreviation used by Recursive Superintelligence, the company Tian co-founded.

Tian turns the science-fiction phrase into a concrete chain of work. A model finds a problem, proposes an idea, changes code, runs an experiment, verifies the result with a reliable measure, and carries a useful change into the next round. The loop can continue only when every step is executable, testable, and reusable.

He also draws a clear boundary. Current models still cannot replace top researchers. Recursive improvement can begin while full autonomy remains out of reach.

The loop is beginning to turn

RSI is an old idea. Researchers were already using AutoML and neural architecture search a decade ago to make machines help improve machines.

Its revival comes directly from stronger coding agents and more capable models. Tian recalls working with a model on grokking research and realizing that it could absorb many tasks once given to interns: discuss a narrow direction, search for ideas, prove a theorem, write code, and run experiments.

Researchers still define problems, judge results, and choose the next move. Lower execution costs let experienced researchers test ideas without waiting for a team to schedule every implementation. The feedback cycle gets shorter.

That change makes a practical loop possible. A model performs part of the research and produces a better training method or piece of infrastructure. The improved system then contributes to the next research cycle. Recursion is the repetition of that loop.

From AutoML to RSI

A typical AutoML system starts with a search space, objective, and set of operations defined by people. An algorithm then searches for a better answer inside those boundaries.

Large language models add a representation of higher-level knowledge. They can read papers, understand code, discuss architectures, and participate in problem formulation. The search space becomes larger, and the system can help organize its path instead of moving only inside a grid designed by a researcher.

Tian argues that this still leaves the hardest research uncovered.

A coding agent can execute a path that is broadly known. Rebuilding a compiler, an operating-system component, or an existing project may be laborious, yet the target is relatively clear. Frontier research often lacks a standard answer for how to state the problem, whether a result is trustworthy, or which experiment deserves the next run.

Research taste, abstraction from sparse evidence, and judgment about whether a direction deserves more work remain weak points for current models.

Anthropic's self-improvement article makes a similar distinction between execution and judgment. Models can already carry out well-specified experiments, while people still do more of the goal selection, dead-end recognition, and reliability assessment. Tian therefore expects people to move higher in the research loop, with recursive improvement arriving before complete automation.

Three benchmarks prove only a first step

Recursive published First Steps Toward Automated AI Research on 2026-06-11. The company reports that one automated research system produced new results on three tasks with short feedback cycles. It also released run artifacts.

These are company-reported results. They show what the system may be able to do and still need independent replication. All three tasks have clear metrics, relatively low variance, and fast experiments, allowing the research loop to run repeatedly.

NanoChat Autoresearch

The task trains a small language model on one GPU for a fixed 5-minute budget and tries to minimize validation loss. The metric is BPB, where lower is better.

Recursive reports a move from the previous best solution at 0.9372 BPB to 0.9109 BPB, a reduction of 0.0263. Measured by the time required to reach the same loss, the company reports a 1.3x speedup.

NanoGPT Speedrun

The public community has optimized the same target for two years: minimize the time required to train a small GPT to a validation loss of 3.28 on 8 H100 GPUs.

Recursive reports reducing the time from 79.7 seconds to 77.5 seconds, an improvement of 2.2 seconds. The change appears small because the starting point was already a public record shaped by years of optimization. Removing another memory round trip or synchronization step is difficult at that stage.

SOL-ExecBench

This benchmark tests whether GPU kernels can run closer to hardware limits while preserving correct outputs. It contains 235 kernel tasks, and 1.0 represents the benchmark's analytical estimate of optimal performance.

Recursive reports raising the mean SOL score from the previous best of 0.699 to 0.754, which reduces the remaining gap to 1.0 by 18%.

Together, the results cover model quality under a fixed compute budget, training efficiency, and low-level infrastructure. They provide evidence that an automated research system can produce useful improvements on tasks with fast, reliable evaluation.

They leave harder questions open. NanoChat and NanoGPT use small-model settings, with a clear scaling gap to frontier models. SOL-ExecBench offers precise scoring, while open-ended research rarely has such clean feedback. Tian calls the release First Steps because that boundary matters.

Can the leader compound forever?

The most compelling RSI scenario gives the leader a stronger model for improving its successor. The successor then accelerates development again, and the gap keeps widening.

Tian accepts that the loop can accelerate and introduces another variable: intelligence may advance through a sequence of S-curves separated by plateaus.

A smooth capability curve would favor the company with the strongest models and the most compute. Recursive loops would amplify the initial lead, making a catch-up increasingly difficult.

Plateaus produce a different competition. Teams can keep approaching the ceiling of one method, while the next jump still requires a new architecture, theory, or training approach. A plateau leaves room for an original method, and compute, data, and energy constrain the returns from scale alone.

This is a technical forecast and a startup thesis, not an established empirical law. It still makes one useful point: a recursive loop can accelerate research without guaranteeing smooth growth forever.

The hard parts are evaluation, data, and interpretability

An automated research system must know whether a change produced a real improvement. A vague score cannot guide the next move. A vulnerable score invites reward hacking, where the system finds a shortcut to a higher number without solving the intended problem.

Recursive says its system reuses prior experiments, combines promising branches, and checks variance and reward hacks before accepting an improvement. The three benchmarks share fast feedback, clear metrics, and evaluators that can be hardened.

Open-ended research does not offer the same convenience.

Tian lists several conditions that remain under exploration: what data can teach a system to conduct research, how to reduce the compute cost of many experiments, how to keep evaluation from being gamed, and how to understand why a model produced a result.

He places interpretability at the intersection of capability and safety. Understanding what a model learned and which mechanisms matter could help it find better improvements while giving people earlier warning of dangerous paths. That is Tian's long-term research judgment, and current technology does not provide a complete answer.

Full recursive self-improvement would also compound control problems. As systems take a larger role in building successors, verification, monitoring, and constraints become more important. Anthropic's article highlights the same risk: errors and misalignment may accumulate when iteration outruns human verification.

People remain inside the loop

Tian's near-term picture is a research loop shared by people and AI.

Models take on more implementation, search, and experimentation. People continue choosing worthwhile problems, setting evaluation criteria, checking results, recognizing superficial optimization, and changing direction when evidence is weak. As execution shrinks, human work moves upward and the quality of judgment matters more.

This explains his repeated emphasis on hands-on work. Researchers who stay far from code and experiments can blur the difference between a difficult problem, an extremely difficult problem, and a problem that current methods cannot solve. Using the model, observing failures, and running validation reveal whether the bottleneck sits in engineering, evaluation, data, or theory.

Recursive's organizational preference also serves this feedback speed. Tian favors a small, flat core team so the person proposing an idea can see experimental results quickly. That preference comes from his experience and should not be treated as a universal law for every research organization.

Research eventually becomes a science

In the second half of the interview, Tian describes a longer-term goal. AI research still contains a large empirical component, with mechanisms that remain hard to explain. He believes the field will eventually develop a scientific structure that can be explained and reasoned about.

He compares the path with alchemy becoming chemistry and with astronomical observations passing through Kepler before reaching Newton. Experiments accumulate phenomena, and a later theory connects scattered regularities into a coherent system. The next unifying figure could be a person or an AI system.

If AI systems continuously discover new knowledge, human researchers can still find meaning in understanding how a discovery arose and appreciating the structure and beauty in its result.

Three questions help test whether a system is approaching RSI today:

1. Can it propose the next move from experimental evidence rather than merely execute a fixed workflow?
2. Can it confirm progress with reliable, game-resistant evaluation rather than raise one score once?
3. Can it transfer a method to new problems and larger scales rather than optimize one benchmark?

Recursive's public results occupy the first step: a short-feedback research loop can now run and produce measurable gains. A system that chooses important problems, invents new methods, and validates them across scales remains much further away.

That is the most useful contribution of the interview. It turns RSI from a story about acceleration into a set of concrete engineering questions. The loop is beginning to turn, while research judgment, verification, and new theory still determine how far it can go.