Tim Davis

But is that actually the future of superintelligence?

I would argue that when you think about what is intelligence and how do you define intelligence, you're sort of saying there's this agent or an entity that's taking actions, receiving feedback, trying to understand the environment that it's in against some goal.

Right.

And, you know, and this is where I think it's hard to say that the future of intelligence is going to exist in the data center medium trained on large scale corpuses of human text that have existed in some ways in the past.

Right.

The future of super intelligence, in my opinion, you know, is this sort of.

reinforcement learning approach, you know, if, and there's, you know, Richard Sutton actually just did a wonderful, you know, blog post and interview on this, where, you know, you really look at reinforcement learning as an approach that needs to be scaled out into the world.

Well, if it's out into the world, then what is the infrastructure, one, that's going to enable us to get out into the world, this process of continual training and continual learning that

software needs to power that hardware needs to power that and then you get into like the physics of of heat thresholds and thermal layers and you know all of these you know just even battery capacity out in the real world to actually have a constant training and learning process that occurs you know in reality not in a data center and so you know i sort of feel like that

Today, we are working towards systems that have a very strong economic reward.

There's no question that what has been built so far is incredibly valuable.

We see this in code generation and many of the use cases that have become commercially valuable.

But I don't necessarily know that training models and continuing just to build our gigawatt facilities

to continue to have this perception of intelligence in regurgitating you know a lot of you know history of human text particularly and even I would argue even in video generation and image generation I don't believe the models understand what they're actually doing and and hilariously you know we actually don't understand what these models are doing and I think you know

Dario Amodi has written some wonderful essays on mechanistic interoperability and understanding how do these models work.

But if you step back, as engineers, what other part of engineering as a society would we be comfortable with structuring and using products when we don't understand how they work?

If you go to a structural engineer and you say, is that bridge going to hold up?

There's a probability it will.

There's a probability that it will.

And we certainly hope, I certainly believe that it should, but I don't know if it will.