AIs get 'dumb' (coding) after about 200k tokens? How is everyone handling this
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
A developer community discussion highlights a concern that AI models, particularly for coding, become less effective or "dumb" after processing approximately 200,000 tokens.…
Started watching a youtube the other day on AI coding by Matt Pocock (on the aidotengineer channel). He said after 200k tokens the ais start to get dumb, better to try start fresh soon after that, or at least realize you are working in the 'dumb zone'.
So I have been trying to start a new session soon after I hit the 200-250k. Problem is it feels way too short. By the time I'm around 200k there's just not much actual coding done yet.
Anyone else following this protocol? Keeping sessions smaller and restarting early? Curious how you get enough real work done before you hit it, if you noticed benefit after following it, or if you think it's BS etc.. Please add if you had a real engineering role pre the ais, or you are a post ai coder, for context.
Not a token thing for me, I'm on max plans for both. (Running Opus 4.8 and Codex 5.6-sol)