Integrated Simulation in AI Systems

Simulations are going to be an increasingly important and common feature in AI systems. In fact, I believe that any AGI system that lacks the ability to perform simulations should be considered incomplete.

I see the future use of simulations as part of a larger trend of AI research entering the design space of cognitive systems.

(However, I do not actually believe AI research is explicitly headed in the cognitive systems direction on purpose. I think that agentic features, chain-of-thought, and attention, all of which are cognitive in design, are the natural result of attempting to improve upon previous existing systems. A true first-principles approach to designing AGI based on cognitive systems would look very different.)

I see simulations as one feature out of many in the space of cognitive design. And I predict that the more we use features from the cognitive systems design space, the more capable AI will become overall.

The first thing I would like to clarify is that when I say “simulation,” I do not intend to conjure visions of 3-D virtual worlds that can be explored. Simulations within AI do not necessarily have to have any sensory modality, and they exist on a spectrum of complexity. A simulation could be as simple as a single equation, and I think simplicity is an important point.

Simulations should be as minimal as possible in order to keep them efficient. This efficiency is important if the cognitive system is a real-time system. The simulation will have to be able to provide results within strict deadlines. Think of these deadlines as a time budget on the order of milliseconds (or less). This is related to an idea of a heartbeat or “frame rate” of a cognitive system.

By “frame rate” in a cognitive system, I mean the number of cycles the system’s main processing loop completes per second. Developers of 3D engines and real-time multiplayer game servers are well aware of the concept of a tick-rate or an internal frame rate. I believe the exact same property will be involved in a cognitive systems approach. I mentioned this because it has direct relevance to simulations and the constraints in which they must operate.

So, if a cognitive system operates at 60 Hz, the entire budget for the cognitive loop is approximately 16.67 milliseconds. That is not a lot of time to do all of the processing necessary for decision-making, reasoning, and all of the other tasks that have to be done. This tight time budget necessitates subconscious and non-conscious processing. And I think it’s worth taking a little bit of time to talk about that as well, because it also relates to how simulations might be used.

I do not think that simulation should be seen as an external tool that is called upon by the AI to answer a specific prompt. The simulations I am envisioning are integral to the very design of the AI’s functionality. So in order to set this up, I’m going to introduce a hierarchy:

• Non-conscious Processing
  • Subconscious Processing
    • Conscious Processing

I also do not want to get mixed up in the word “consciousness” for this article. For now, I use the term to refer to the information available to the agent, subject, or “identity” that makes up or resides in the cognitive system. I realize that is also something that would need to be explained, but to keep this focused, I will have to defer that to a later time. The primary goal here is to discuss the different levels of informational access involved in processing.

(It is also entirely possible that a cognitive system doesn’t have to be designed based on this hierarchy. This is a hierarchy that appears to correlate with our own minds. That may not be the best way to create a cognitive system, but it is a good place to start.)

There is a level of non-conscious processing that supports both the subconscious and conscious levels of processing. This information on the non-conscious level is not accessible to either the conscious or subconscious levels of the cognitive system. This would form the boundary between the “outside” of the mind within the cognitive system and the “inside” of that mind. This distinction between “inside” and “outside” is one of informational access. This might be related to the concept of access consciousness¹, but I do not want to conflate the two.

Simulation may occur at any one of these levels, and I think that it is deeply fundamental to the very architecture of a cognitive system. For example, any of us can probably create a simulation in our minds at any time by sitting back in our chair, closing our eyes, and thinking of a game of chess or a baseball game. We could visualize all the different aspects of the game and play it out in our minds. This would be a simulation that is directly available in our conscious processing. But there are other levels where simulation might take place.

At the subconscious level, it’s hard to say exactly how simulation is used. And this is where things get a little bit abstract. For example, it could be the case that simulation is involved in the subconscious for how we learn to navigate the world or interact with objects. And then we subconsciously act out on those simulations as learned behaviors, and it becomes tacit knowledge. And I think simulations also operate on a non-conscious level.

Now clearly, it’s basically impossible to report on how simulations might be working on a non-conscious level because they’re not observable to us as humans. I would speculate that simulations might be used on a non-conscious level to support learning or information processing. For example, simulations might be used to structure the layout of an artificial mind to fit the structure or shape of information in one or more modalities. Simulation on a non-conscious level could also be used to transform information between different structures. In this case, I’m thinking of simulation more along the lines of computing by analog. It’s hard to give a concrete example and I would rather leave it for now. Therefore, I wouldn’t rule out the abstract use of simulations at all three levels that I’ve described.

The last thing I’ll say related to this simulation hierarchy is that I don’t think that simulation should be something that is purely external to the AI like a tool. This is very popular right now because it’s basically the only option. What I mean by that is that in the future there may be AI systems, especially AGI, where simulation is integral to the very design of how the cognitive system operates. It could be argued that this is a false distinction and that whether the simulation is “outside” the AI is irrelevant. I would disagree, and I think complete integration of simulation will have significant impacts on future AI capabilities.

In short, I do not believe that the most effective AI systems will employ simulations like a tool. I think they will have simulations built into their design at all of the levels I mentioned previously, and possibly other ways as well. This would require a complete rethink of how to construct such AI systems.

Right now, I think the cognitive features being realized with LLMs are a form of imitation that will lead to inefficient and inflexible outcomes. I believe that the solution to this problem is to take a cognitive systems approach to designing AI, which is essentially my view on how we ought to be building AGI. So it is basically one and the same.

Simulations also need to be infinitely recursive in the limit of cognitive resources. By this, I mean the ability to simulate the act of simulation, creating a simulation within another. I think the only thing that stops this is the available resources of the cognitive system. And even in an artificial cognitive system, I think there will be practical limits based on meeting time constraints. Having only 16 milliseconds is not a lot of time to do arbitrarily nested simulations. So there will be practical limits. However, I just wanted to point out that it will be an important property of simulations to be able to be nested in this way.

The last property I’ll discuss for now is that simulations need to be generic. I focused a lot on the efficiency of simulations and how they are situated in a cognitive system, but I didn’t really spend a lot of time talking about the structure of the simulations themselves.

Whatever the structure is for a simulation, it needs to be as general as possible. I alluded to this earlier when I mentioned that I did not want people to think of simulations for AI in terms of virtual 3-D worlds. The actual simulations could be of anything, and I will list a few examples:

• Thinking about the future.
• Empathy and theory of mind.
• Estimating the force required to throw an object.
• Planning a route through an environment.
• Anticipating and avoiding future consequences.
• Replaying or reenacting past events.
• Learning through imitation.
• Mirroring behavior.
• Mental rehearsal for learning or future performance.
• Mental models of computations, algorithms, data structures, and computer programs.
• Self-models and introspection.

Many more examples could be listed. The pattern is that simulation is very generic. So it is going to be a challenge to try to come up with a general-purpose architecture that can create and manipulate simulations. And the simulations need to be an integral part of the cognitive system, not just some external tool that is called upon. I’m not criticizing the use of tools in AI, and I think it would be a great place to start. However, I really do believe that the moral of the story is that we should be working towards changing how we build AI. And I think that will be absolutely required to construct the most efficient and effective forms of AGI.

References

¹ Block, Ned. “On a Confusion About a Function of Consciousness.” Behavioral and Brain Sciences 18, no. 2 (1995): 227–47.