[moomoo11]

And that program would be written different each time and sometimes fail.

[sshine]

...and the LLM, given an agentic loop, would ingest its own error message and correct itself...

...and eventually it'd persist some knowledge in a context window to not make that mistake for a while...

...and then it'd forget and make the same mistake again...

[moomoo11]

exactly