What kind of simulated environment is complex enough to develop a general AI?

Question

Imagine trying to create a simulated virtual environment that is complicated enough to create a "general AI" (which I define as a self aware AI) but is as simple as possible. What would this minimal environment be like?

i.e. An environment that was just a chess game would be too simple. A chess program cannot be a general AI.

An environment with multiple agents playing chess and communicating their results to each other. Would this constitute a general AI? (If you can say a chess grand master who thinks about chess all day long has 'general AI'? During his time thinking about chess is he any different to a chess computer?).

What about a 3D sim-like world. That seems to be too complicated. After all why can't a general AI exist in a 2D world.

What would be an example of a simple environment but not too simple such that the AI(s) can have self-awareness?

Answer 1

I will skip all thematic about "what is an AGI", "simulation game", ... These topics have been discussed during decades and nowadays they are, in my opinion, a dead end.

Thus, I can only answer with my personal experience:

It is a basic theorem in computing that any number of dimensions, including temporal one, in a finite size space, can be reduced to 1D.

However, in practical examples, the 1D representation becomes hard to analyze and visualize. It is more practical work with graphs, that can be seen as an intermediate between 1D and 2D. Graphs allows representation of all necessary facts and relations.

By example, if we try to develop an AGI able to work in the area of mathematics, any expression (that humans we write in a 2D representation with rationals, subscripts, integrals, ...) can be represented as 1D (as an expression written in a program source) but this 1D must be parsed to reach the graph that can be analyzed or executed. Thus, the graph that results after the parsing of the expression is the most practical representation.

Another example, if we want an agent that travels across a 3D world, this world can be seen as an empty space with objects that have some properties. Again, after the initial stage of scene analysis and object recognition (the equivalent to the parser in previous example), we reach a graph.

Thus, to really work in the area of AGI, I suggest skip the problems of scene analysis, object recognition, speech recognition (Narrow AI), and work directly over the representative graphs.

Answer 2

General AI can absolutely exist in a 2D world, just that a generalized AI (defined here as "consistent strength across a set of problems") in this context would still be quite distinct from an Artificial General Intelligence, defined as "an algorithm that can perform any intellectual task that a human can."

Even there, the definition of AGI is fuzzy, because "which human?" (Human intelligence is a spectrum, where individuals possess different degrees of problem solving capability in different contexts.)

---

Artificial Consciousness: Unfortunately, self-awareness / consciousness is a heavily metaphysical, issue, distinct from problem solving capability (intelligence).

You definitely want to look a the "Chinese Room" and rebuttals.

---

Probably worth looking at the holographic principle: "a concept in physics whereby a space is considered as a hologram of n-1 dimensions." Certainly models and games can be structured in this way.

Another place to explore is theories of emergence of superintelligence on infinite Conway's Game of Life. (In a nutshell, my understanding is that once researchers figured out how to generate any number within the cellular automata, the possibility of emergent sentience given a gameboard of sufficient size is at least theoretically sound.)

Answer 3

I think the most important thing is that it has to have time simulated in some way. Think self aware chatbot. Then to be "self aware" the environment could be data that is fed in through time that can be distinguished as being "self" and "other". By that I suppose I mean "self" is the part it influences directly and "other" is the part that is influenced indirectly or not at all. Other than that it probably can live inside pretty abstract environments. The reason time is so important is without it the cognitive algorithm is just solving a math problem.

Answer 4

Of the answers so far, the one from @DukeZhou were the most provocative. For instance, the reference to the Chinese Room critique brings up Searle's contention that some form of intentionality might need support in the artificial environment. This might imply the necessity of a value system or a pain-pleasure system, i.e. something where good consequences can be "experienced" or actively sought and bad consequences avoided. Or some potential for individual extinction (death or termination) might need to be recognized. The possibility of "ego-death" might need to have a high negative value. That might imply an artificial world should include "other minds" or other agents, which the emerging or learning intelligent agent could observe (in some sense) and "reflect on", i.e recognize an intelligence like its own. In this sense the Cartesian syllogism "I think therefore I am" gets transmuted into: I (or rather me as an AI) see evidence of others thinking, and "by gawd, 'I' can, too". Those "others" could be either other learning systems (AGI's) or some contact with discrete inputs from humans mediated by the artificial environment. Wikipedia discussion of the "reverse Turing test"

The mention of dimensionality should provoke a discussion of what would be the required depth of representation of a "physics" of the world external to the AI. Some representation of time and space would seem necessary, i.e., some dimensional substructure for progress to goal attainment. The Blocks World was an early toy problem whose solution provoked optimism in the 60's and 70's of last century that substantial progress was being made. I'm not aware of any effort to program in any pain or pleasure in the SHRDLU program of that era (no dropping blocks on the program's toes), but all the interesting science fiction representations of AI's have some recognition of "physical" adverse consequences in the "real world".

Edit: I'm going to add a need for "entities with features" in this environment that could be "perceived" (by any of the "others" that are interacting with the AGI) as the data input to efforts at induction, identification, and inference about relationships. This creates a basis for a shared "experience".

Answer 5

Though a good answer by @pasaba por aqui, I'd agree with @zooby that a graph might be too simplistic. If humans were in an environment where the options were drown or take 5000 unrelated steps to build a boat, we'd never have crossed any seas. I think any graph, if designed by hand, would not be complex enough to call the agent within as general AI. The world would need enough in-between states that it would no longer be best described as a graph, but at least a multidimensional space.

I think there are 2 points you'd have to consider. What is "simple" and when would you recognise it as a "general AI". I don't find self aware AI satisfactory, as we can't measure anything called awareness; we can only see its state and its interaction with the environment.

For 1. I'd pose that the world we live in is actually fairly simple. There are 4 forces of nature, a few conservation laws, and a bunch of particle types that explain most of everything. It's just that there are many of these particles and this has led to a rather complex world. Of course, this is expensive to simulate, but we could take some shortcuts. People 200 years ago wouldn't need all of quantum mechanics to explain the world. If we replaced protons, neutrons and the strong force with the atoms in the periodic table, we'd mostly be fine. Problem is we replaced 3 more general laws with 100 specific instances. For the simulated environment to be complex enough I think this trend must hold. We could replace trillions of particles governed by general laws with thousands of instances that have different properties when interacting with the agent, and I think more importantly, when interacting with each other.

Which brings me to 2. I think we'd only truly be satisfied with the agent expressing general AI when it can purposefully interact with the environment in a way that would baffle us, while clearly benefiting from it (so not accidentally). Now that might be quite difficult or take a very long time, so a more relaxed condition would be to build tools that we'd expect it to build, thus showing mastery of its own environment. For example, evidence of boats have been found somewhere between 100k and 900k years ago, which is about the same time-scale when early humans developed. However although we'd consider ourselves intelligent, I'm not sure we'd consider a boat making agent to have general intelligence as it seems like a fairly simple invention. But I think we'd be satisfied after a few such inventions.

So I think we'd need a Sim like world, that's actually a lot more complicated than the game. With 1000s of item types, many instances of each item and enough degrees of freedom to interact with everything. I also think we need something that looks familiar to acknowledge any agent as intelligent. So a 3D, complicated, minecraft-like world would be the simplest world in which we would recognise the emergence of general intelligence.