In today’s column, I identify how the use of world models is radically reshaping the nature of generative AI and large language models (LLMs). Here’s the deal. The usual approach to data training for generative AI and LLMs consists of using a vast corpus of data such as text found across the Internet and having the AI pattern-match on that encountered data. An additional and quickly emerging augmented approach involves establishing a so-called world model that interacts with the budding LLM. During those interactions, the LLM further refines its pattern-matching and logic-based reasoning capabilities accordingly.
Voila, via interacting with suitable world models, generative AI and LLMs can be substantively boosted.
Let’s talk about it.
This analysis of an innovative AI breakthrough is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
Human Use Of World Models
Before we leap into the AI side of things, it would be instructive to consider how humans make use of world models.
Imagine that you are learning about baseball. Assume that you’ve never played baseball and only vaguely know what it is about. Meanwhile, you are familiar with many other sports such as football, basketball, and others.
If a friend wanted to tell you about baseball, they would probably start by briskly outlining the key rules and precepts of baseball. This would likely include scenarios of situations that might be seen at a baseball game. The friend would explain that the batter attempts to hit the baseball and then tries to run from base to base. Etc.
In a sense, all this information is feeding into your mind, and you are in the midst of crafting a mental model of the game of baseball. The friend is depicting a type of world model that specifies the contours and particulars of baseball. You, in turn, or seeking to formulate a world model in your own mind that becomes specific to your thinking processes.
You are presented with a world model entailing baseball, and you formulate in your mind a personalized world model that represents baseball.
You and your friend then go to an actual baseball game. As you sit in the stands watching the game, your baseball-envisioned world model is being refined. Witnessing actions taking place on the field allows you to find gaps in your knowledge. Observing the game also reinforces what might have been construed as abstract and now is directly understood via what is occurring during the game.
That’s a quick rundown of world models in the context of human learning, and we are ready to proceed to the next consideration, namely adding AI into the mix.
Classic Paper On AI And World Models
Shift gears to the AI realm.
There is a considered classic research paper covering the intertwining of AI and world models entitled “World Models” by David Ha and Jurgen Schmidhuber, arXiv, May 9, 2018, and the researchers made these salient points (excerpted):
- “Humans develop a mental model of the world based on what they are able to perceive with their limited senses. The decisions and actions we make are based on this internal model.”
- “We have demonstrated the possibility of training an AI agent to perform tasks entirely inside of its simulated latent space dream world.”
- “Training agents in the real world is even more expensive, so world models that are trained incrementally to simulate reality may prove to be useful for transferring policies back to the real world.”
- “By training the agent through the lens of its world model, we show that it can learn a highly compact policy to perform its task.”
The upshot is this.
Suppose you wanted generative AI to become versed in baseball. The odds are that during the initial data training, a budding LLM encountered lots of online text that described the sport of baseball. Based on that text, it is possible to discuss baseball in general with just about any generative AI and get a reasonably sensible response.
If you wanted to go further and have the generative AI be deeply responsive about baseball, you could hook up the AI to an online video game involving playing baseball. I’m sure you’ve seen or used such online video games. You usually can use online controls to swing a virtual bat, run the bases, catch a pop-up flyball, and otherwise be immersed in a simulated baseball game.
What if we had the AI do the same?
The AI would act somewhat as a human would in terms of playing the online baseball game. By doing so, the AI is augmenting the “book learning” associated with the text about baseball that was scanned and patterned on at initial training. This gives the AI a type of “experience” related to playing baseball. It’s clearly not the same as if the AI was a robot playing on a true baseball field, but it offers a handy and easy alternative to such an experiential-oriented activity.
World Models Are Incomplete
We ought to acknowledge that though using and formulating world models is certainly advantageous, there are some gotchas and issues that arise.
A famous quote from the 1970s by one of the luminaries in a field known as system dynamics well-illustrates a big weakness or limitation of world models:
- “The image of the world around us, which we carry in our head, is just a model. Nobody in his head imagines all the world, government or country. He has only selected concepts, and relationships between them, and uses those to represent the real system” (as cited from “Counterintuitive Behavior Of Social Systems” by Jay Wright Forrester, Technology Forecasting And Social Change, Volume 3, 1971).
Any world model that we use to train a generative AI app is indubitably going to be limited and not cover everything there is to know about the domain or topic being covered. Thus, it is construed as incomplete. The AI is using an incomplete representation to figure things out, and likely the result will be incomplete too.
Another qualm is that the world model might have errors or contain incorrect aspects. The AI is patterning on what the world model is showcasing. That could lead to the AI getting things wrong. For example, suppose the online video game depicting baseball has a glitch that allows a runner to run faster than the speed of light when running the bases. That doesn’t match with the real world of baseball.
The AI might simply accept as true that base runners can run at the speed of light. The AI now has a falsehood that was carried over from the world model being used to instruct the AI.
The world model that the AI itself devises could also be flawed.
Perhaps during the playing of the online video baseball game, the AI determines that a good strategy in baseball seems to be that if you hit the umpire with your bat, the ump will let you automatically take first base. The AI has gleaned something about baseball that doesn’t align with reality.
Virtual Environments Allow Experimentation
The use of world models to augment the training of generative AI provides an important means of the AI carrying out experimentation, doing so at scale.
Allow me to elaborate on that point.
Remember that I mentioned earlier that we were pretending that you aren’t familiar with baseball and a friend was trying to explain the sport to you. The friend took you to a baseball game so you could see with your own eyes how baseball is played.
How much could you glean about baseball by attending one game?
I dare say that you’d observe a lot of the fundamentals, but the chances are that more arcane aspects might not perchance occur during that one game. Would you see a triple play? Probably not. Would you see a home run? Maybe.
The use of a computational world model such as a baseball video game allows for a multitude of runs to be undertaken by the AI. We could set up the AI to play the video game and keep doing so, over and over. The AI would readily play thousands of such instances. If we made sure that the world model was constantly changing up the showcased scenarios, the AI would likely encounter almost all the circumstances that might happen in real life when watching baseball.
I’ve discussed in prior column postings that this same conception has been used to develop AI for self-driving cars, see for example my discussion at the link here. It goes like this. Before an AI automated driving system is given access to a real autonomous car, the AI can be run through thousands or millions of scenarios in an online video game depicting driving a car.
This prepares the AI and does so with essentially no risk of harm to anyone. The moment you put the AI into an actual car, there is a chance that the AI might steer wrong or otherwise drive incorrectly. The computational world model approach is a lot less risky, and a lot less costly, versus once you put the AI onto the roadway.
Models, Models, Models
I’ve got a question for you to mull over.
How many world models and what kinds of world models should we be applying to AI?
Think of it this way. I so far discussed using a world model depicting baseball. I’ve also mentioned a driving simulator that would be a world model for operating a car. Those are two quick examples of world models used to boost AI.
The list of potential world models is seemingly endless. We might have a world model for doing heart surgery and use that to train AI on performing or aiding heart surgeries. We could have a world model on how to calculate your taxes. On and on this could go.
Setting up a world model and attaching AI to it does require a certain amount of time and effort. Running the AI and having it use the world model chews up computational resources. Validating that the AI gleaned the right stuff, and avoided patterning on the wrong stuff, that’s a big cost to the whole kit and kaboodle.
The gist is that we aren’t likely to haphazardly opt to use world models with AI. This is something that you do when the purpose is worthy, and an ROI is satisfactory for doing so.
It’s a crucial matter that is starting to gain widespread traction. You can bet your bottom dollar that we will see a lot more of these efforts. The conventional method of scanning lots of data to get AI up-to-speed is beginning to reach its boundaries.
What else can we do to advance AI?
I’ll give you three words, namely models, models, models.
