Benchmarking for world models is a very interesting task. You might see a very broad pattern with my blogs. It's about a lot of stuff around world models. That is because I've recently been reading about it a lot, and a lot of my research efforts recently have gone into sort of really understanding how world models work, how they're going to be good for us, and how we are going to test world models.
Like how do we know that a world is truly a good depiction of the real world and is going to be able to, with the most efficiency, simulate stuff in real life.
By the way, most of what I call research efforts are just me sitting for long hours with all the information I have and thinking about everything from the very first principles, as if what I already know about society and the world in general doesn't exist.
Today, in this blog, I'm going to be talking about the third part, which is- how do you benchmark world models? I think a very common mistake that humans, and there's a reason why I say humans, would make is that they would want to verify the output of a world model against what a human would do. Now, this is right and wrong, depending on how you want the future of the world to look like:
If you want AI to just be another human being, which I don't see a point of, then that approach is right.
But if you want AI to be something truly productive and to be able to physically help you. With stuff that you don't want to do in your day-to-day life and be able to do that and also be able to help humanity in general with other more life-threatening situations like war, then I don't think that training their models like what a human would do is truly helpful in situations like these.
Because inherently human decisions are rational, but not always right. If everyone in the world started thinking rationally, then there would be no conflicts, because everything would be justified, right? but it doesn't work like that, and we're not here to discuss philosophy, although these blogs might seem like that, but in a more technical sense.
Anyways, if you want AI to be more like a super intelligence layer, then you've got to start training on the expected outcomes and train it to always pick the most efficient path to achieve that outcome. Now, what I mean by that is, there are going to be infinite ways to get to a certain outcome, right, what we humans may call "the journey". Sometimes for some people, the journey may take X time, but for some others, it may take 2 X or 3 X, even 10 X the time. Right? So the journey is different, but the outcome is the same.
Your efficiency will be maximum when you take the least time- in this case, it will be X. So the training has got to be, for optimising for the most efficient paths. Now, these situations are not a mathematical equation or a mathematical algorithm which you solve for, at least not yet, as we improve the capability of our current world models. This maybe gets more towards a mathematical problem, rather than a statistical problem although both of those may lie under the same umbrella for some.
So in order to benchmark a world model, what do you do? you need to train another world model to always pick the most efficient path. For achieving a given outcome from a certain situation. So essentially, this becomes like a chicken and egg problem. Where in order to get one, you have to get the other, but to get the other, you have to get the one, right?
For example, if we are training, let's say robots for war, there is no first-person data available for war footages. Other than FPV shooter game videos, which trust me, there are a lot of them on YouTube, and a model can essentially get trained on those because not only is there an affirmative side of killing an opponent, but there is also a conflicting side, which is getting killed. So you can appreciate the skills and moves in the algorithm that can lead to a kill and penalize heavily the skills that remove that can lead to your death, which can essentially be then transferred to these war robots, right?
The guns are the same, the shooting mechanism is the same. The only thing that is different is the terrain, and that is something you're gonna have to manually gather the data for, right, in the terrains that these robots are gonna fight, like a hilly terrain or a flat terrain, or a marshy area. The movements differ with that. But the shooting and aiming and targeting mechanism or how we can assemble troops and how we can dispatch troops, all that intelligence, right? That. Lies at the foundation of these videogame playbacks.
And funnily enough, I think that a very strong bet I can make is going to be that these video game companies are going to be at the frontier of world model training, and they're gonna be the 1st guys to market with a commercially available world model, which they can license to robotic companies, depending on what sorts of games they made. They really have a vast collection of a lot of data, which is essentially going to be the mode for becoming a frontier world model.
Similar concept can be applied to car driving or bike riding gameplays and training models for autonomous vehicles, although the players drive extremely rash and irresponsibly while playing this games for a lack of any physical consequence, I still think some arrangements can be made towards that end as well.
And historically, chicken and egg problems have always had a tendency to get solved, mostly mathematically, so I think we are just probably one Fable 5 prompt away from solving this problem.
And it's going to be interesting to see how this entire vertical shapes up when that is solved.