if you have a candidate that lies 5% of the time, but u cant tell what 5% because they lie convincingly, would that help or hurt your business to hire them?
So, I think it's a bit misleading, or perhaps unintuitive, rather, that this technique was labelled "MoE". It's more like S-Lora, where the model actively swaps out relevant LoRAs at inference time. It's not strictly speaking anything "new" as such, but a series of existing techniques tied together into a simple package. I'm not sure how useful it really is to the broader community, particularly given that it's not open source, and that there are existing techniques, like mechanistic interpretability, that should essentially do something really quite similar at the end of the day, to say nothing of advancements in reinforcement learning which will not eliminate an LLM's ability to lack confidence (raw LLMs actually have a pretty good internal estimate before instruction tuning of how accurate the facts they're saying are, we just destroy it in fine tuning atm, but forcing them to answer confidently).
Hmm. Lots of PR stunts on their blog. So still... skeptical. I really don't get the main trickery, and 200 API calls per month is not enough to get a proper test-through. "Internal memorization. Tuning the weights, not RAG. You can layer them." /via X.
This is interesting and somewhat aligns with how the brain seems to work. We have general capabilities that we use all the time, but we are also able to retrieve memories even after years of not accessing them. So it implies that we have weights that change, and memories that are more static / MoE-like where we can pull them up at will.
Nice! That does add a lot more comfort in correct answers. The "mixture of agents" model architecture is coming in with some good stuff too (not as good as this though - this is big). We're not far from some really smart agents...
It’s special because it swaps in those experts within a larger architecture. Related research on polysemanticity also suggests that sparsity will enhance explainability and steer ability