> We might not understand particular "emergent" capabilities
The emergent capabilities are the only capabilities we care about
> We might not understand particular "emergent" capabilities
The emergent capabilities are the only capabilities we care about
For allignment maybe.
For the core functionality and the optimizations we don't really need to know how the emergent capabilities decide on particular answers.
Which is why we could build LLMs before those features ...emerged for us to see, and why we can just code LLMs with the numerical NN algorithms we use, and do now have to go in and change individual weights.
Paraphrasing
We don’t? I sure would like to know!We do understand "thinking" though. That's literally the whole point of Attention is All You Need, the attention mechanism is what separates the transformer architecture from other neural networks. It's well worth a read if you haven't gone over it yet.
Features like chain-of-thought, long-horizon contexts and RoPE/YaRN all extend this thinking capability very transparently. The only remaining thing to study is the data and weights, which probably isn't going to contain some sort of miraculous revelation.
Attention is at about the same level of abstraction as spike trains or action potential, IMO. It’s a mechanism, it doesn’t tell you much at all about how actual concepts get represented. (It merely defines the substrate with which they can be represented.)
The entire field of Mechanistic Interpretability exists because just understanding Attention does not in any way help you to understand why a certain NN responds with a certain hallucination about a certain Chinese boat in this specific context.
> The only remaining thing to study is the data and weights
To me this is like saying “the only thing left to study in the brain is the connectome; probably going to be boring, we understand it already”. It’s almost all of the hard/meaningful stuff! It’s where intelligence and consciousness lives!
The entire field of mechanistic interpretability has gotten nowhere. The formalized, "causal" understanding of latent space conceptualization isn't any better than an LLM cargo cult.
It's wholly possible that you could study one set of weights for decades, and find nothing. There's no guarantee that any patterns outside of human language exist in that data. In this specific context, it's satisfying enough to state that [Chinese] and [boat] were both tokens in the tokenizer, activated by a feedforward pass through weights that favor [boat] after [Chinese]. There's not any guaranteed solution to this. There's not even any guaranteed problem; that hallucination is an expected behavior.