| I believe uploading neural weights in the form factor of a reaction speed game will become the norm. Imagine the token weights etc to live an an N dimensional embedding. Consider a random 2D isometric projection, now imagine plotting (a subset of) the tokens on screen at their respective locations, with the background having patterns conveying the orientation of the projection. Imagine tokens appearing FIFO, and 95% or so of the tokens at their correct position, and 5% at incorrect positions. The user is expected to identify misplaced tokens in the projection. Each the frame the perspective slowly changes. Since we guarantee ~95% of the tokens on screen are correctly placed this enables the human to absorb their correct high dimensional location subconsciously. Imagine this game is made a bit addictive. Imagine eventually the user gets high scores and thus knows most of the coordinates of the token embeddings in a coordinate free way (no axes had to be drawn) At that point all the information the ANN required is undisputedly present in the humans brain, absent how those coefficients are used. This is where the fun starts. The alcorithm can calculate the likelihood of observing 2 tokens in a certain order. It can generate more probable pairs, this will correlate with the information in the brain. For example imagine the user knows some spanish, and that the language model was multilingual. Suddenly the brain starts picking up the correlation between close juxtapositions of certain tokens, and their positions learnt from the computer game. While playing and getting better scores in the game, the user starts noticing its own grammar and vocabulary improve, because the brain has been helped in better estimating next tokens... People will be able to learn math, languages, programmnig languages, tables of chemistry, etc... with substantially less effort. |