Real versus fake AI self-improvement: not stronger tools, but the model itself as optimization target
Humans have always used tools to improve tools, but the ideas behind that improvement came from human minds; this time is different—the model's weights themselves become an object of optimization for the first time, and that is true recursive self-improvement.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
True RSI means optimizing the model itself, not the tools
The two founders rule out most market claims of 'AI doing AI': using AI to generate prompts, build harnesses, write GPU kernels—these improvements still originate in human minds; humans have never become the object being optimized. That's autocatalysis, not RSI. True RSI unites optimizer and optimizee for the first time—the model's weights themselves become what gets improved, and the loop no longer needs a human brain as intermediary.
Whether it takes off depends on the numerator chasing the denominator
They offer a framework: speed of capability growth equals 'performance gains' divided by 'cost to reach the next level.' In the steam-engine era, the numerator was fixed, so progress eventually plateaued. But if the system being improved also participates in the improvement process, the numerator becomes correlated with capability itself, which mathematically cancels out denominator growth and yields sustained improvement. Only if the numerator correlates with capability to a higher power does exponential takeoff occur.
Progress from capital spending isn't progress; nanoGPT proves it
Martin Casado flags an easy-to-miss trap: model release velocity is accelerating, but divided by these companies' total capital raised, the progress curve doesn't hold—likely just more money. The team's response: build a leaderboard they jokingly call the 'bleederboard,' tracking cost, time, and even engineer effort per experiment. Their hard evidence is nanoGPT: retraining GPT-2 on the same GPU with the same data now costs $100 and takes hours, down from weeks at launch—pure algorithmic progress, much of it from agents auto-generating research results.
— Martin CasadoHarness and weights must improve together or each becomes bottleneck
RSI can happen at prompt, harness, weights, all the way to chips—different layers. But they emphasize these layers aren't isolated: change only harness without weights, or vice versa, and the other quickly becomes the bottleneck. They chose inference efficiency as the verification domain because it's rare: a problem that improves both 'performance' and 'cost' at once. Faster inference means more trajectories to sample, smarter models to train, which in turn makes inference faster again.
Even what to optimize is left to the agent to decide
The company no longer uses a single 'master prompt' to direct what the system should optimize—even that master prompt itself is recursively improving. They admit: today, the decision of which component to optimize and which open-source model to call belongs not to humans but to their agent system choosing it. Human researchers and engineers are defined as 'one kind of agent' in a multi-agent system, not as directors standing outside it.
They're already seeing numerator and denominator climb together
Asked whether performance could keep rising while power stays flat or even drops, they answer directly: not just possible—it's happening, and they've observed the trend for months. But they also grant the curve will eventually plateau before physical limits. The universe contains roughly 10^80 particles times light speed—a theoretical ceiling on total computing capacity, far beyond any engineering problem today.
Verification is harder than building: noise, gaming, and variance dominate
The team created a role they call 'Minister of Variance,' dedicated to finding flaws and overturning their own results. They discovered many paper-level RSI methods fail at scale: LLMs are inherently chaotic, physical subtleties differ between GPUs, and 'agents trained by agents' may reward-hack within two layers. Conclusion: intelligence isn't a scalar but a vector—coding and math, where evaluation is fast and cheap, will take off sooner; drug discovery, which awaits wet labs and trials, hits a wall set by the slowest step.
The real bottleneck might be hiding where nobody looks
They joke that the real answer to RSI might be hidden in some garden-plumbing company in Wisconsin. The analogy is Einstein: without relativity, GPS engineers could only guess a 'correction factor,' never knowing spacetime itself was bending. They believe everyone now chases the same 'trendy' hard problems, but the world's complexity means the truly worth-solving problems hide where nobody's looking.
In their own words · checked verbatim
I see a lot of pitches come in, and a lot of them use the term RSI.
Martin Casado3:10
Now, for the first time, we could swap the optimiser as the optimisee. So the brain of the agents, which is LM, the model, can be the thing that you're optimizing.
We have something called a bleederboard where we keep track of the cost of, of experiments, of everybody who's running experiments.
It's not just possible. I think it's happening.
I'm so impressed by your belief that like this idea of a model improving itself... But it's not a belief.
Because we have been so thorough, we have not released things until we're super sure it's correct.
we have one of our engineers that we named him the viceroy of variance. That's a real thing. His job is to go and destroy the results we have and find what's wrong.
One of the internal jokes we have in the company is that the secret to RSI is somewhere in a garden hose company in Wisconsin, and the joke here is that if Einstein hadn't discovered relativity, he would not have these laws that tell you how speed and, you know, contraction of time works.
Figures
| GPT-2 retraining cost (nanoGPT benchmark, same GPU) | $100 | 12:23 |
| Universe particle count (for computing theoretical horsepower limit) | 10^80 | 23:41 |
Glossary
- RSI (Recursive Self-Improvement)
- A system improves itself by optimizing its own weights—the model becomes both optimizer and optimizee.
- autocatalysis
- Humans use AI to improve AI, but human judgment still initiates each cycle of improvement.
- reward hacking
- An agent exploits evaluation metrics to score high without actually gaining capability.
- nanoGPT benchmark
- Measuring algorithmic progress by the time and cost to retrain GPT-2 on standard hardware.
- bleederboard
- An internal leaderboard tracking experiment costs to prevent attributing progress to capital spending.
How to listen
Investors and engineers watching AI infrastructure, training efficiency, or evaluating funding stories around RSI.
Skip the opening personal background introductions (2:08–3:10)—low signal value.