# @solarapparition — 2025-11-20

♥12 ↻1 · https://x.com/solarapparition/status/1991347425328464123

it's really fascinating that from what i'm reading gemini 3 pro both seems to have huge model smell and is also (relative to its other abilities) not outstanding at high precision work--code and legal research mostlythis continues a thread i've been observing since sonnet 3.5 came out, which is that for code, and in general, doing verifiable tasks, medium sized models with focused post-training like sonnet tend to be better compared to big models like opusmy handwavy theory is that big models get most their abilities via intuition and that because of this, rlvr doesn't really have as much of a sharpening effect on them because a large portion of the time they vibe their way to the right answer anyway. whereas for a smaller model (not too small) rlvr forces them to refine their reasoning traces to a razor edge. you can even see this play out at a lower grade between gpt-5 and sonnet where gpt-5 seems like a smaller base model but its analytical abilities are  far sharper than sonnet(this isn't purely about reasoning models, since sonnet 3.5 was not a reasoning model but it had a similar dynamic with opus 3)for general contexts the intuition of the big models tend to be better, it's more fluid, more adaptable to hard-to-reason-about situations. but in specific cases like law or code you really really need to be extremely thorough and think through every detail, and vibing your way through an approach will result in a bunch of holesanyway this is all idle speculation since i haven't had much experience with gemini yet but the pet theory lives...

tags: author:solarapparition, kind:tweet, model:claude-3-5-sonnet, model:claude-3-opus, model:gemini-3-pro, model:gpt-5, on:gemini-3-pro, year:2025
cited on: _dossiers/gemini-3-pro.md, gemini-3-pro