Self Improving AI - Notes from Jason Wei (OpenAI Researcher)
1. We don’t have true “self-improving AI” yet.
AI today doesn’t train itself. When we do figure it out, it will change everything. But it won’t be like flipping a switch — it will be slow and gradual, probably taking many years, maybe even a decade.
2. Self-improvement won’t happen all at once.
It’s not like one day AI can’t train itself, and the next day it perfectly trains the next version.
For example, imagine GPT-5 trying to train GPT-6. At first, it would probably be very inefficient, using much more time and computing power than human researchers.
Only after many tries would it get better than humans at training new models.
3. It also won’t improve equally in every area.
Some things are easier for AI to get better at, like fixing simple mistakes or improving writing style.
Harder things, like math and coding, take more work — but we still have methods to improve there.
Very hard things, like learning rare languages (e.g. Tlingit spoken by ~500 people), would be really slow, because there isn’t much data to learn from.
4. Even super smart AI can’t skip real-world experiments.
Some people think if an AI reads all science papers, it could instantly cure cancer or train perfect models. That’s not how it works.
In reality, progress comes from testing things in the real world, doing experiments, seeing what works and what doesn’t.
Even if an AI is much smarter than us, it still needs to run experiments and wait for results. This speeds things up, but it’s not an instant breakthrough.
⭐ Bottom line:
Self-improving AI will come slowly, not suddenly.
Progress will vary by field — some problems will be solved fast, others will take much longer.
High intelligence alone isn’t enough; AI will still be limited by the speed of experiments and real-world testing.
So yes, it will accelerate progress — but not a dramatic overnight explosion.