Post

Claude seems to be trained to bias against long-running simulations. Always picks tiny parameters and toy problems. Takes several turns to convince it to do anything nontrivial.

9/8/26, 5:10 pm0 reposts0 likes

CJ Avilla

@cjavilla.bsky.social

10 Aug

Can you share what model and effort level you're working with? OOC what types of simulations?

Ryanne Dolan

@ry.codes

10 Aug

Now that you mention it, I've been using Fable LOW effort bc the tasks are straightforward, just meta parameter sweeps. The damn thing is like "this is a great place to stop" every turn, having done very little work itself. I keep compacting but I suspect it doesn't like having a long session.

CJ Avilla

@cjavilla.bsky.social

10 Aug

My intuition is that's because of the effort setting being at low. here's a couple things I'd try: 1. try using opus as the main model and fable as advisor (set /model to opus and then /advisor to fable. Could even try sonnet as model and fable as advisor) 2. try bumping fable effort to medium