The Self Learning Process: An RSI Focus

What Is RSI? How AI Could Learn to Improve Itself

First, How Ordinary AI Learns

The Baseline

RSI: The Loop That Rewrites the Loop

The Big Idea

RSI stands for Recursive Self-Improvement. It is when an AI improves the very process that makes it smarter. Not just better answers. Better at getting better.

Recursive means the output feeds back in as the input. Each upgrade builds the next upgrade.

The Chef Who Trains Better Chefs

Analogy

Ordinary learning: a chef cooks, tastes, and adjusts the recipe. The food gets better.

RSI: the chef redesigns how chefs are trained, then trains a better chef.

That new chef redesigns the training again, and trains an even better one.

The improvement is not in the dish. It is in the chef making machine.

Each Generation Engineers the Next

How RSI Works

Learning vs. RSI

The Difference

Early Signs of RSI Today

Real Examples

Self-play: AlphaZero mastered chess by playing millions of games against itself, with no human games to learn from.

AI improving AI: Google DeepMind's AlphaEvolve found code speedups that cut time off training the Gemini models it runs on.

AI writing AI code: at leading labs, coding agents now write a large share of the software used to build the next models.

Self-made training data: models generate practice problems, check their own answers, and train on what worked.

None of these is full RSI. Humans still steer each loop. But the pieces are arriving.

"The first ultraintelligent machine is the last invention that man need ever make."

Origin, 1965

I. J. Good, mathematician and WWII codebreaker. He called the runaway version of RSI an "intelligence explosion." His full sentence adds a condition: provided the machine is docile enough to tell us how to keep it under control.

What Could Slow RSI Down

Reality Check

Compute: every upgrade still needs chips, power, and time to train.

Testing: some improvements can only be proven in the real world, which moves at human speed.

Diminishing returns: the easy wins come first. Later gains may get harder.

Human checkpoints: labs can require review before any new version is deployed.

So RSI could be an explosion, a steady climb, or a loop that stalls. Experts disagree.

Why Everyone in AI Watches RSI

Why It Matters

Speed: if each generation builds the next faster, timelines could compress from years to months.

Oversight: the faster the loop, the harder it is for humans to review each step.

Alignment: an AI's goals and values need to survive every upgrade, not just the first one.

Opportunity: the same loop could accelerate science, medicine, and software for everyone.

Better at Getting Better

Key Takeaway

Ordinary AI learning is a feedback loop on answers. RSI is a feedback loop on the learning itself. Partial versions already exist. Whether full RSI arrives, and how fast, is one of the biggest open questions in AI.