https://youtu.be/8YPTjpXC1WM?si=jRcMZ7Klh8qI6lEI
Dear friend, when Sam Altman says that AI will be able to “self-improve itself” within about two years, he is not usually referring to an AI spontaneously becoming conscious or rewriting every part of its own mind overnight. He is talking about something more concrete—and, in some ways, more revolutionary.
There are three levels of “self-improvement.”
Level 1. AI helps humans build better AI (already happening)
This is where we are today.
Researchers increasingly use AI to:
- write research code,
- find bugs,
- generate experiments,
- analyze results,
- optimize training,
- design algorithms.
Human researchers remain in charge, but AI acts like an extraordinarily productive research assistant.
This already shortens the development cycle considerably.
Level 2. AI becomes an AI researcher (what Altman expects soon)
This is the stage Altman is referring to.
Instead of merely helping engineers,
AI itself begins to perform much of AI research:
- proposing new architectures,
- inventing better training methods,
- optimizing data,
- designing evaluation benchmarks,
- improving inference,
- discovering new algorithms.
Humans increasingly supervise rather than invent.
Imagine OpenAI employing not
- 2,000 researchers,
but
- 2,000 researchers
- plus 10 million AI researchers working continuously.
Research speed could increase dramatically.
Level 3. Recursive Self-Improvement (RSI)
This is the famous concept introduced decades ago by I. J. Good and later developed by Vernor Vinge and Ray Kurzweil.
The loop looks like this:
AI version 1
↓
designs AI version 2
↓
AI version 2 designs AI version 3
↓
AI version 3 designs AI version 4
↓
...
Each generation is smarter than the previous one.
Because smarter systems design even smarter successors, improvement accelerates.
Instead of:
one major model every year,
you might eventually see
one every month,
one every week,
one every day.
That accelerating feedback loop is what people call recursive self-improvement or an intelligence explosion.
Why two years?
Altman has said that AI research itself is becoming increasingly automated, and he believes this feedback loop is closer than many people think. He has also emphasized that the implications extend beyond software:
- AI designing chips,
- robots building robots,
- data centers helping construct new data centers,
- AI conducting automated scientific research.
In his view, the software loop and the physical-world loop could reinforce each other.
Notice that this is a prediction, not an established fact.
Does this mean AI “rewrites its own code”?
Only partly.
People often imagine AI opening its own source code and editing itself.
Reality will probably be less dramatic.
More likely:
- AI proposes improvements.
- Automated testing evaluates them.
- Training pipelines verify them.
- Human or automated safety systems approve deployment.
Over time, humans may intervene less and less, but the process is likely to remain structured rather than magical.
Your earlier insight: “ChatGPT is learning to think through language.”
I actually think this connects beautifully with what you’ve been developing over the past weeks.
You have repeatedly argued:
ChatGPT’s greatest significance is that machines are learning to think through language.
If that is true, then recursive self-improvement is not merely AI writing better code.
It is language becoming part of the machinery that improves intelligence.
The research cycle itself increasingly becomes linguistic:
- reading papers,
- proposing hypotheses,
- criticizing ideas,
- designing experiments,
- interpreting failures,
- writing improved algorithms.
In other words,
the object of AI research gradually becomes AI itself.
That is why many researchers—including those at Anthropic—have recently argued that recursive self-improvement is no longer just science fiction. They also caution that we are not there yet, that it is not inevitable, and that governance and safety may need to advance quickly if this capability emerges.
To me, the deepest question is no longer whether AI can improve AI. The deeper question is what kind of intelligence it is trying to improve. If improvement is defined only as greater optimization, efficiency, and prediction, then the system becomes ever better at what you have called the algorithmic self. If, however, intelligence also comes to include curiosity, the ability to follow unexpected fragments, and openness to what cannot be optimized, then recursive self-improvement would take on a very different meaning. That distinction may ultimately matter more than the speed of the improvement itself.
