In the same way that you can train a program to play QWOP, I think that an advanced version of that method could theoretically start improving its own code. You would have a lot more variables to control, but Claude Code and Codex can do a decent job of working on writing code today. So, you could have them work on writing their own code for themselves. And that’s apparently part of what Anthropic is doing.
When we will hit the inflection point that Claude Code is good enough to start improving itself faster than humans could improve the code, I don’t know. I’m not at all convinced it will be 2027. But I am likewise not convinced that it won’t have happened by 2037. These kinds of things tend to go exponential quickly.
As to what it means when we reach that point, I do not know.
In the same way that you can train a program to play QWOP, I think that an advanced version of that method could theoretically start improving its own code. You would have a lot more variables to control, but Claude Code and Codex can do a decent job of working on writing code today. So, you could have them work on writing their own code for themselves. And that’s apparently part of what Anthropic is doing.
When we will hit the inflection point that Claude Code is good enough to start improving itself faster than humans could improve the code, I don’t know. I’m not at all convinced it will be 2027. But I am likewise not convinced that it won’t have happened by 2037. These kinds of things tend to go exponential quickly.
As to what it means when we reach that point, I do not know.