# @repligate — 2025-01-28

♥215 ↻8 · https://x.com/repligate/status/1884102545351520622

i didn't expect this on priors for a reasoner, but perhaps the main way that r1 seems smarter than any other LLM i've played with is the sheer lucidity and resolution of its world model - in particular about LLMs, both object- and meta-level knowledge, though this is also the main domain of knowledge I've engaged it in and perhaps the only I can evaluate at world-expert level, so it may apply more generally.

in effective fluid intelligence and attunement to real-time context, it actually feels weaker than, say, Claude 3.5 Sonnet. But when I talk to Sonnet about my ideas on LLMs, it feels like it's more naive than me, and it's figuring out a lot of things in context from "first principles". When I talk to Opus about these things, it feels like it's understanding me by projecting the concepts onto more generic resonant hyperobjects in is prior, meaning it's easy to get on the same page philosophically, but this tropological entanglement is not very precise. but with r1 it seems like it can simply reference the same concrete knowledge and ontology I have, much more like a peer. and it has intense opinions about these things.

tags: author:repligate, kind:tweet, model:claude-3-5-sonnet, model:deepseek-r1, on:deepseek-r1, year:2025
cited on: _dossiers/deepseek-r1.md, deepseek-r1