|
|
|
|
|
|
by vjk800
35 days ago
|
|
|
I feel this more and more as I age. Especially after having children. I used to be a "tech guy" (like most people here probably) and was excited about new technology. Now my main feeling when something disruptive (like AI currently) comes up is: "why the hell do people need to rock the boat". The thing is, I'm perfectly happy living my life as I have been living so far, concentrating on doing stuff with my children and having fun. When the world changes, stuff I need to worry about it: is this going to affect my job in the future? What is the long term effect of exposing my children to this? Is the stuff I teach my children going to be relevant in the future after this disruption has happened? I don't want to be forced to learn new stuff. I mean, I can learn new stuff occasionally for fun, but it's not fun if my life and salary depends on it. Fuck the tech bros trying to change everything up. |
|
|
Right now most of the weights are in the FFN neural nets, but I believe we can reformulate LLM's or distill them into a format where each token/character represents either a matrix or a geometric algebra multivector. If each weight becomes token/character associated, it means we could write mindless reaction speed games: have a 2D projection of token/word/concept multivector/matrix coordinates. Let the 2D projection smoothly evolve a grand tour through the space of projections, possibly only including new dimensions if user performance on the explored dimensions suffices. 95% of the tokens/words could be rendered at their correct positions, while say 5% is at their incorrect positions. So the user sees a subset of the character/word/tokens rendered, and since they are individually placed at a correct position 95% of the time, it allows a user to learn peripherally, the user just needs to click those tokens intentionally disturbed from their true positions. Consider the hypothetical scenario that eventually all coordinates are explored and hypothetically the user scores highly (perhaps 100% or 90%) then we could say that at least the high dimensional coordinates are properly uploaded. Now you could randomly render bigrams (the new matrix coordinates being the matrix product / the new multivector coordinates being the geometric product). Next we could hypothesize that with sufficient training the user can predict the correct position of bigrams (meaning the user brain is performing the same or similar computation). From this the likelihood could be computed, and so a user would start predicting the same next tokens the model would. It wouldn't be blind belief more like a user associating the claims with what the reaction speed game insinuates. Imagine having trained on random cyrillic characters and suddenly you understand the Greek ingredients on some food packaging. There is a lot of hypothesis here, but none of the steps seem impossible to me.
It would mean both kids and adults could sync up with what LLM's claim (without blindly believing in it, but when it helps predict a lot of things around you people may lean and trust more on it, so that danger is not entirely moot), and having an LLM acquaintance in your brain to consult.