I've been obsessing over trivial technical details of LLMs, spending way too much time 'groking' trivial issues.
One of them is how 'tokens' are extracted from text and sent through the neural-nets. It's done by creating a random vector for each token, such as converting the letter 'x' into coordinates [1.37, 2.98]. During "training", these "parameters" are "adjusted". What that means is that when plotted, they'll move.
So I created a webpage that animates this, watching what happens to this randomized parameters during training.
We see:
* vowels and punctuation create their own clusters away from others
* The letter 'v' moves away by itself
* The '-' character stays at the center, moving toward the unpopulated quadrant.
You can see the code here.
https://t.co/twJ910LAz1
A brief (AI generated) description of the code is here:
https://t.co/tl7jMIP1kq
Did you know you can take Polystyrene to Mitre10 for FREE and Expol take it away and turn it into insulation panels? How good is that!?!?
Meanwhile at my local transfer station...
This is never gonna happen in a million years in the USA but really the government should be providing some kind of cryptographic key to every citizen.
January 2025 brings a rare sight: six planets aligning, with Mercury joining for seven. Visible starting January 21 after sunset, you can see Venus, Mars, Jupiter, Saturn without gear. Telescope needed for Neptune, Uranus.