First of all anyone making an AI can and should and even probably does know how their code works, they should know how their Backprop, vectors, etc purposes and results are. This is so obvious but it needs to be said.
Next, they all also should and probably do know why the backprop/ word vectors/ negative sampling works and is useful, and I do agree maybe here a bit they don't know why even when they invented it. Deeper, why and how does backprop or word2vec or positional encoding etc predict the next letter or word sub part? How does it deal with or does it deal with long range dependancies? Why I mean, what I mean, does it help? BPE helps ignore rare context memory matching, but DO they SAY this? Or they just say use BPE, it, helped? Looks like to me the latter! This they all fail at pretty bad, by the looks of their way they talk and can't explain GPT etc at ease, but me and some others are on top this real good. Next, to the bloody details, no one really can look at all the things that helped predict the next letter or word or sentence, you can and can ask the AI to explain, but this isn't always needed much of the time, give or take, basically. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T73cb0deded02df8c-Md91cc7ac0f60be0cf9433afd Delivery options: https://agi.topicbox.com/groups/agi/subscription
