Post #2412747

@worik@mastodon.social @jadedtwin@corteximplant.com @WhiteCatTamer@mastodon.online @Eatsbluecrayon@rollenspiel.social Dude, recently I tried to have deepseek solve a problem and it couldn't, it went in circles for a long time and then the answer was wrong. If it was truly intelligent and being capable of actual thought, I could tell it where it went wrong and it would be able to learn from it. But do you know what happens when I send another message? The LLM sees the previous messages, but it doesn't see the previous thinking blocks, to save context tokens. Try to talk to it about things it mentions in its thinking blocks. You can't, because it literally doesn't have recollection of ever having any of those thoughts. You can change the frontend to include all the thinking blocks, but in most cases it performs worse because the context is filled with fluff (and because all of the training data for fine tuning only ever has a single thinking block). Before deepseek R1, and even before chatgpt o1 (the first version with "thinking"), I made a local LLM "think" by having a multi user conversation (basically several NPCs) and then turning one of the characters into the actual "assistant" and the other one into "assistant's thoughts". It's just lengthening the conversation with multiple argumentation lines so it can figure out the best one. Is a chatroom "thoughts"? Is the person that talks the most what constitutes as "thinking"?