> How can AI ID a cat?
For reference, my initial understanding was somewhat low: basically I know a) what embedding is basically b) transformers work by matrix multiplication, and c) it's something like a multi-threaded Markov chain generator with the benefit of prior-trained embeddings
Resources I’ve liked:
Sebastian Raschka book on building them from scratch
Deep Learning a Visual Approach
These videos / playlists:
https://youtube.com/playlist?list=PLoROMvodv4rOY23Y0BoGoBGgQ... https://youtube.com/playlist?list=PLoROMvodv4rOwvldxftJTmoR3... https://youtube.com/playlist?list=PL7m7hLIqA0hoIUPhC26ASCVs_... https://www.youtube.com/live/uIsej_SIIQU?si=RHBetDNa7JXKjziD
here’s a basic impl that i trained on tinystories to decent effect: https://gist.github.com/nikki93/f7eae83095f30374d7a3006fd5af... (i used claude code a lot to help with the above bc a new field for me) (i did this with C and mlx before but ultimately gave into the python lol)
but overall it boils down to:
- tokenize the text
- embed tokens (map each to a vector) with a simple NN
- apply positional info so each token also encodes where it is
- do the attention. this bit is key and also very interesting to me. there are three neural networks: Q, K, V – that are applied to each token. you then generate a new sequence of embeddings where each position has the Vs of all tokens added up weighted by the Q of that position dot’d with the K of the other position. the new embeddings are /added/ to the previous layer (adding like this is called ‘residual’)
- also do another NN pass without attention, again adding the output (residual) there’s actually multiple ‘heads’ each with a different Q, K, V – their outputs are added together before that second NN pass
there’s some normalization at each stage to keep the numbers reasonable and from blowing up
you repeat the attention + forward blocks many times, then the last embedding in the final layer output is what you can sample based on
i was surprised by how quickly this just starts to generate coherent grammar etc. having the training loop also do a generation step to show example output at each stage of training was helpful to see how the output qualitatively improves over time, and it’s kind of cool to “watch” it learn.
this doesn’t cover MoE, sparse vs dense attention and also the whole thing about RL on top of these (whether for human feedback or for doing “search with backtracking and sparse reward”) – i haven’t coded those up yet just kinda read about them…
now the thing is – this is a setup for it to learn some processes spread among the weights that do what it does – but what those processes are seems still very unknown. “mechanistic interpretability” is the space that’s meant to work on that, been looking into it lately.
Instead I believe this might work better as a guided exercise where a person can work on it over a few hours rather than being spoon-fed it over the 10 minute reading time. Or breaking up the steps into "interactive" sections that more clearly demarcate the stages.
Regardless I'm very supportive of people making efforts to simplify this topic, each attempt always gives me something that I either forgot or neglect.
If I'm interpreting it correctly, it sort of validates my intuition that attention heads are "multi-threaded markov chain models" , in other words if autocomplete just looks at level 1, a transformer looks at level 1 for every word in the input plus many layers deeper for every word (or token) in the input.. while bringing a huge pre-training dataset to bear.
If that's correct more or less, something that surprises me is how attention is often treated as some kind of "breakthrough" - it seems obvious to me that improving a markov chain recommendation would involve going deeper and dimensionalizing the context in a deeper way.. the technique appears the same just the amount of analysis is more. I'm not sure what I'm missing here. Perhaps adding those extra layers was a hard problem thta we hadnt figured out how to efficiently do yet (?)
Ilya Sutskever: Yeah, so the thing is the transformer is a combination of multiple ideas simultaneously of which attention is one.
Lex Friedman: Do you think attention is the key?
Ilya Sutskever: No, it's a key, but it's not the key. The transformer is successful because it is the simultaneous combination of multiple ideas. And if you were to remove either idea, it would be much less successful. So the transformer uses a lot of attention, but attention existed for a few years. So that can't be the main innovation. The transformer is designed in such a way that it runs really fast on the GPU. And that makes a huge amount of difference. This is one thing. The second thing is that transformer is not recurrent. And that is really important too, because it is more shallow and therefore much easier to optimize. So in other words, it uses attention, it is a really great fit to the GPU and it is not recurrent, so therefore less deep and easier to optimize. And the combination of those factors make it successful.
I think it does not scale well beyond 5 boxes (20 numbers) because the stacks become too complex to remember and identify patterns in. This is me, could be also quite individual.
Also, the Illustrated Transformer: https://jalammar.github.io/illustrated-transformer/
Also, this HN comment has numerous resources: https://news.ycombinator.com/item?id=35712334
Also, here's an interactive "transformer explainer" that is absolutely mind-blowing [2].
aabdel0181•5mo ago