Caveman Thinking

[phene]

I've noticed that though the final output from my prompts is caveman-ized, if the model does any thinking, that thinking output is not using caveman. Is there something that can be done to improve thinking so that it does use caveman?

[RivasMario]

like the models rational internally is using caveman speak to save on tokens?

[phene]

Some of the "thinking" comes out as output, and it's not subject to caveman's rules about language.

[marcebdev]

I've encountered the same thing, handing off to subagents, results from subagents, general thinking is the same. I thought that was the main utilization of this, to help really reduce context window growth.

I'm honestly not sure how useful it pans out if it's just the response being extremely terse especially since that's usually the part that's helpful to be more human-readable and not in another language if you're using Wenyan mode

[JuliusBrussee]

Unfortunately no AI providers allow us to truly edit how thinking works so hopefully in a future update but for now thinking remains as normal

[marcebdev]

I'm curious if you could share any potential examples where this really benefits workflows?

I think many probably just prefer or know use workflows they're comfortable with, but it seems like this would be most helpful with keeping plans more terse or reducing output for long running chats, is that accurate?

I ask just to better understand how to best use this tool.

[yamiteru]

AI models internally treat thinking very differently than the normal output.

There have been attempts to change the style in which AI models think but basically all of them agreed that it's basically impossible and if somehow you do achieve some form of control then the quality degrades drastically.

I think caveman thinking will require AI models to be trained differently in regards to thinking.

[RivasMario]

yeah most of the time when someone does the caveman speak to you face to face you kind of just fill in the blanks yourself subconsciously. Maybe think more concise thoughts mode. Or like use the proper word to describe the thing instead of a sentence, but then that would require a bigger vocabulary.

[yamiteru]

I recommend checking out "symbolic thinking".