Post #1223849
2026-03-18 04:02 UTC
@dhobern Like most things "it depends" because they tend to wrap a bunch of other things around the LLM to adjust its behavior. At its base a user prompt is literally just tacked onto a system prompt and fed to the model, and yes, it's weighted more heavily in the resulting computation.
Some of the models will do a thing where they run that prompt through another system first (or runs it against an LLM in a loop) and then feeds _that to an LLM_.
But it's all just adjusting statistics.
Replies (1)
-
@cthos@mastodon.cthos.dev 2026-03-18 04:04
@dhobern There are also "Mixture of Experts" where there's a router that processes the tokens of the input as they run through the system and routes those to models that are "specialized" on a given thing, and then recomposes it out the other end. But again, statistics.