Andy Halliday

Let me just do a quick aside here that in a regular LLM that doesn't have this conditional memory, what happens is every single time that the inference run is done, many of the layers of the deep neural network are being used to

Figure out what these static entities are.

Let's think about these as entity names, like the names of individuals or places of things, things that aren't going to change.

They're rooted in the basics of physics and or the world.

So this new conditional memory from DeepSeq is taking those static sort of engram is the technical term for it, embeddings, and using them to do a lookup primitive.

Think of it as a very fast short-term memory.

And it puts those things into that, which is outside the computational process of the deep neural network.

And so that preserves many more layers in the inference process to be used for reasoning rather than this computation of the static content.

Anyway, that's probably way more than I'll need to say about that.

Let's just suffice to say that as a result of doing that, this new process boosts the performance on knowledge and reasoning benchmarks

by efficiently freeing up the neural network's resources for more complex reasoning by offloading them into this, you know, really fast and at time of inference process.

Carl's joined me.

Welcome, Carl.

I've been holding down the fort here until you joined me.

I was just, you know, talking probably too long about DeepSeq's Ngram conditional memory system.

Breakthrough, I think.

I think this is a breakthrough.

And once again, the Chinese are doing something that's creating a much more efficient model.

And I think that's important.

I'll toss it to you, Carl.