DeepSeek using intuition vs logic

Posted by DeviantPlayeer@reddit | LocalLLaMA | View on Reddit | 5 comments

Reply to Post

5 Comments

[-]

UnreasonableEconomy@reddit

Is it actually evidence of models being able do to embedding arithmetic "at will"? Would be super interesting if that's the case.

[-]

I think so. Subtracting Japan from Germany doesn't make any sense, not logical, nor intuitive and I didn't give it any room to think and shuffle them around, so it's pure intuition, yet it gave the expected answer. I also gave it more complex cases like a-b+c= and then told which letter is what. In that case it just combined 3 words together without spaces every time.

[-]

DeviantPlayeer@reddit (OP)

https://preview.redd.it/186hgopb3jee1.png?width=639&format=png&auto=webp&s=628a9f2bb6ed2d7d09001eaabfd01c3898e99995 R1 did a better job. Bratwurst was one of the plausible answers but it rejected it because it was too straighforward.

[-]

Utoko@reddit

now we know!

[-]

DeviantPlayeer@reddit (OP)

When using intuition it either says Schnitzel or Bratwurst, while using logic it outputs multiple pages of text depending on how confusing the expresion is. The more weird the order of words the more it thinks. With intuition the order of operation seems to be irrelevant, only affecting Bratwurst/Schnitzel ratio. However, it can't handle expressions where the first word is negative, it treats it as positive instead.