Andy Halliday

But there's an argument that's made that in order to get to superintelligence, it has to have the foundation of world model causality and understanding of real world physics, not just comprehension of text.

Yes.

Thank you.

Okay.

I have a weave off of this actually news that I just came across today, which is really interesting.

A new research paper published in archive by a group of researchers at the university of Chicago and what they've discovered by analyzing actual reasoning models is

using their ability to look into the vague transparencies to observe what's actually happening inside the model.

So there are ways to inspect the model and what it's doing when it's performing reasoning.

And this is a really interesting development here.

Reasoning models generate societies of thought.

So what they found is that reasoning models that they're examining

exhibit much greater perspective diversity than instruction-tuned models, the ones that have been fine-tuned to be able to follow a sequence of steps, like step-by-step reasoning, if you will.

But reasoning models that go beyond that and have the ability to apply logic, in effect, to novel situations, they show perspective diversity.

Meaning in the internal structures, when they examine the concept spaces that are developing inside the deep neural network, there's a kind of separation of different perspectives.

like point and counterpoint.

So what this does is it creates a set of conflicts in the reasoning model across different personalities, let's call them, inside the model.

So then I'm going to read from this abstract because it's really, it doesn't take very long and I think it's really interesting.

So here we go.

Quoting from the abstract.

Multi-agent structure manifests in conversational behaviors.