Circuit tracing LLMs reveal some bizarre reasoning processes

Comments

• antonim@lemmy.dbzer0.com ⁨1⁩ ⁨day⁩ ago

  The way it does math is mostly as people have already assumed - approximating instead of doing it “the normal way”. It’s 2025 and at this point absolutely nobody should be surprised that AI “confidently describe[s] the standard grade-school method, concealing its actual, bizarre reasoning process”.

  As for poetry,

  Here, the model settled on the word “rabbit” as the word to rhyme with while it was processing “grab it.” Then, it appeared to construct the next line with that ending already decided, eventually spitting out the line “His hunger was like a starving rabbit.”

  this is exactly how many poets write rhymed poetry too, it’s not even remotely bizarre.

  Still, it is interesting and good to see some concrete advancement in the study of AI reasoning. Hopefully it will contribute towards reducing the mystification of the whole thing.

  source

  • Acamon@lemmy.world ⁨1⁩ ⁨day⁩ ago

    Yeah, I didn’t find it particularly bizarre. Both are very natural ways to process verbal information. Anyone who’s ever tried to do arithmetic in a new language knows that we don’t just abstractly do math, a big part is that we know that seven plus eight is fifteen. That’s why they used to teach multiplication tables by rote. It would be lot more bizarre if an llm had independently devised a reliable mathematical algorithm.

    source
• jasoman@lemmy.world ⁨1⁩ ⁨day⁩ ago

  So it is human like, I seen bizarre things from humans every day.

  source