Epistemic status: Strongly arguing for what I feel is a neglected approach. May somewhat overstate the case and fail to adequately steelman counter arguments. I hope and expect that readers will point out flaws in my logic.
Currently, large transformers with dense floating point weights are state of the art in language modeling. Despite recent progress by Anthropic and others, they remain difficult to understand.
Why are they hard to understand?
I've been working on pure combinational logic LLMs for the past few years, and have a (fairly small) byte level pure combinational logic FSM RNN language model quantized to And Inverter Gate form. I'm currently building the tooling to simplify the logic DAG and analyze it.
Are you, or others, interested in talking with me about it?
Discretized weights/activation are very much not amenable to the usual gradient descent. :) Hence the usual practice is to train in floating point, and then quantize afterwords. Doing this naively tends to cause a big drop in accuracy, but there are tricks involving gradually quantizing during training, or quantizing layer by layer.
I am hopeful that we can get interpretability and easy training. But you may well be right.
After skimming some of your progress reports, I am very excited about your sparse nets work!
I agree that reducing superposition is probably valuable even if it requires a significantly larger network. I still don't understand why the transition from float to binary would cause a dramatic reduction in superposition capacity. But if it does prevent superposition, great! I'll just give it more parameters as needed. But if we still get superposition, I will need to apply other techniques to make it stop.
(I have not yet finished my closer re-read of Toy Models of Superposition after my initial skimming. Perhaps once I do I will understand better.)
Hopefully in a few months I will have empirical data regarding how much more neurons we need. Then I can stop hand waving about vague intuitions.
If we can get the unwanted cognition/behaviors to sit entirely in their own section of weights, we can then ablate the unwanted behaviors without losing wanted capability. That's my hope anyway.
My thoughts and hope as well.
I'm glad we agree that RNNs are nice.
So if I understand correctly, you are saying:
Consider that most of the trits will be 0 and thus removable, and that we will be replacing the activations with boolean logic and applying logic simplification transformations to discard even more nodes. The number of trits in the weights is not the same as the number of gates in the resulting logic graph.... (read more)
Current NN matrices are dense and continuous weighted. A significant part of the difficulty of interpretability is that they have all to all connections; it is difficult to verify that one activation does or does not affect another activation.
However we can quantize the weights to 3 bit and then we can probably melt the whole thing into pure combinational logic. While I am not entirely confident that this form is strictly better from an interpretability perspective, it is differently difficult.
"Giant inscrutable matrices" are probably not the final form of current NNs, we can potentially turn them into different and nicer form.