Post B5quFb1zJ0n2FeISlU by u0421793@toot.pikopublish.ing
 (DIR) More posts by u0421793@toot.pikopublish.ing
 (DIR) Post #B5qszLqDug04YOU4RM by futurebird@sauropods.win
       0 likes, 0 repeats
       
       People get mad when you call LLMs "spicy autocomplete" but my investigations into recreating and implementing small versions of this tech make me think that nick name is very accurate. Basically, it's a method to predict the next content in a text file. The whole conversation between you and the LLM is one file, and the LLM tries to find the most likely next text based on the training data. There is something significant here: LLMs were trained on internet forums and social media.
       
 (DIR) Post #B5qszM7wqmB7RMmE4G by liiwi@mastodon.social
       0 likes, 0 repeats
       
       @futurebird Someone recently used term "Augmenting Intelligence" and I thought it describes much better.
       
 (DIR) Post #B5qszMJe9HXI1eFZIm by futurebird@sauropods.win
       0 likes, 0 repeats
       
       @liiwi It kind of implies something intelligent rather than probabilistic is going on though. If I have a hat filled with quotations of wisdom and I pull one out and read it now and then some of the time it will align with what is going on and seem very perceptive. If I have three hats with such quotes and they are labeled "good" "bad" and "cryptic" and I pick one based on the mood people might think I'm a genius.
       
 (DIR) Post #B5qszMK07xos2kPqr2 by futurebird@sauropods.win
       0 likes, 0 repeats
       
       Thus the training data didn't just contain text, but rather text where each passage is tagged and attributed to a particular user. This aspect of the training data was critical in creating the illusion of talking to another person. An LLM doesn't just predict the next text. It predicts the next text that might come from another user. You need to hard code this in to make it work well. Leave it out and there is no conversation.
       
 (DIR) Post #B5qszMYBHFA6kj3AxM by djsumdog@djsumdog.com
       0 likes, 0 repeats
       
       I call them "Weighted Random Word [or Code] Machines." I have a friend who said he wasn't going to continue the conversation if I was using "slurs." I called him a Cogger Lover.
       
 (DIR) Post #B5qszMwzmz0lzgezdQ by futurebird@sauropods.win
       0 likes, 0 repeats
       
       For example if I give an LLM without user seperation this text:"It's a lovely day." It might continue with "The sun was shining."But with user separation it focuses on responses to "it was a lovely day" from other users and the training data might suggest "I agree, it's wonderful weather."So interaction with an LLM is like posting on a forum, it gives you and average of typical responses with one small change: most LLMs have a strong positivity bias programmed in.
       
 (DIR) Post #B5qszNKONzj7AFbg6S by futurebird@sauropods.win
       0 likes, 0 repeats
       
       Because let's be real, if you posted "It's a lovely day." on an internet forum you might get a response like "No it's not, noob."LLMs are heavily weighted to give supportive, and constructive responses. I wonder what they might be like without these limitations? Without the limitation to make the response from another user they might be much less deceptive. That they are popular shows that many people just want a nice moderated online community where people treat each other with respect.
       
 (DIR) Post #B5quFb1zJ0n2FeISlU by u0421793@toot.pikopublish.ing
       0 likes, 0 repeats
       
       @futurebird@sauropods.win this is a simplistic view – that it’s all about token prediction of similar vectors using gradient descent to arrive at the more likely next token to place in line with what is already there. There’s also RLHF – reinforcement learning through human feedback. This involves the human dressing up as a wizard and sitting behind the curtain ensuring that all the responses are the sort of responses that the wizard I mean human would actually prefer and approve of. Technically, this is achieved using a lot of smoke and some carefully placed bidirectional mirrors which act as beam splitters to construct a hologram which fools people into thinking that the machine did it all, instead of poorly paid workers.
       
 (DIR) Post #B5quFbHwLhYB37lCd6 by djsumdog@djsumdog.com
       0 likes, 0 repeats
       
       The early models did involve hundreds of people who were given multiple generations and clicked on the one that was the least idiotic for eight hours a day, but I think a lot of the newer ones just violate OpenAI/Anthropic's user agreements and use the existing models for reinforced feedback learning. (DeepSeek likely did this). They're likely also using the feedback users give GPT/Claude during use.So yes, it's going to always have the bias of whatever the rules were printed out for those original employees as well as their own personal biases.But it's not too over simplistic an explanation. These are next word (token) prediction machines. Each thing that's generated requires the entire context of text to be passed through the machine for each new word. The models themselves are also deterministic, it's just that the generator doesn't always pick the most likely next token. It might randomly select the token at 96% instead of 98% to introduce some variability.