Post AcGkFfr5QsR5sAOFG4 by lain@fediffusion.art
 (DIR) More posts by lain@fediffusion.art
 (DIR) Post #AcGXY9kTjHayOTmXD6 by lain@fediffusion.art
       2023-11-28T12:53:55.921458Z
       
       0 likes, 0 repeats
       
       One aspect that speaks very much against human cognition working like LLMs is that LLMs always think with the same speed: The prediction of the next token always happens in the same time. So there's no deliberate thinking happening on harder issues. Whether you ask it if 1 is prime or whether 123093275755303 is prime, the next token will always be predicted after the same number of computational steps.
       
 (DIR) Post #AcGkFfr5QsR5sAOFG4 by lain@fediffusion.art
       2023-11-28T15:16:15.362612Z
       
       0 likes, 0 repeats
       
       although apparently LLMs get better at math when you tell them to tell you how they arrived at the solution
       
 (DIR) Post #AcGkbJxBXihuJhZ452 by grillchen@brotka.st
       2023-11-28T15:20:09.776991Z
       
       0 likes, 0 repeats
       
       @lain both arent prime
       
 (DIR) Post #AcGkemLMRjCVEeWOTQ by grillchen@brotka.st
       2023-11-28T15:20:47.543764Z
       
       0 likes, 0 repeats
       
       @lain (done in one step with high accuracy. just return not prime every time)
       
 (DIR) Post #AcGoZyIAzfjkf6zkrQ by guizzy@fediffusion.art
       2023-11-28T15:22:30.421700Z
       
       1 likes, 0 repeats
       
       @lain It makes sense, I think. Tokens in a chain of reasoning are a more natural continuation of one another than the correct answer is a continuation of a question that was not a part of its training data.
       
 (DIR) Post #AcHxCzDySHKXde6WJs by doggel@merovingian.club
       2023-11-28T18:14:39Z
       
       1 likes, 0 repeats
       
       @lain I think it may be because the LLM is giving itself more context to correctly predict the next token, predicting the correct next step of a math problem should only need info about the previous steps, whereas properly predicting the tokens of the correct answer with only the initial statement may be more difficult
       
 (DIR) Post #AcIOVwd11rXTqq57Mu by VD-15@fediffusion.art
       2023-11-28T15:18:28.127592Z
       
       1 likes, 0 repeats
       
       @lain I've honest to god had an LLM tell me that 2+2=5 and I didn't even give it the Winston treatment beforehand.