Intelligence as software engineering

Humans can learn to become competent at playing board games such as Chess and Go, but humans are phenomenally excellent at learning how to play such games quickly. Recent approaches to using reinforcement learning to play games hardcode simulators for the games, and mostly ignore what I think is the more important challenge on the path to AI.

You can teach a child how to play Connect Four in a few minutes, at which point the child can already perform some kind of model-based lookahead search. Yet even a game as seemingly trivial as Connect Four is not trivial to implement a simulator for. Even the code that checks whether either player has won the game is a little tricky, and there are ample opportunities for off-by-one errors and out-of-bounds errors.

We can describe variations of the game (e.g. use a diamond-shaped grid where pieces fall on either diagonal) and instantly build bug-free internal simulators in our minds that permit lookahead search. Yet even such a simple change might require a tricky refactor of the implemented simulator, and the refactoring would be even trickier if the implementation needed to be general enough to support both versions, in which case it might even require new classes and abstractions. And this is a trivial game. We humans can build simulators in our heads for much more complicated domains, such as mathematics, that might be legitimately difficult for expert software engineers to implement at all, let alone in a flexible and maintainable way.

By what magic are humans so good at this?

It is tempting to think that a human mind has the equivalent of a large codebase with reusable libraries, and that it builds its Connect Four simulator by gluing together a generic Game class with a Grid class and a enumerate_connected_sequences function that can operate on Grids and that can deal with diagonals and edge cases correctly, and so forth. However, every software engineer knows how incredibly hard it is to design such flexible and reusable abstractions.

Every programming language makes some kinds of compositions and abstractions easier than others. Traditional languages such as C++ and Python make this kind of software engineering possible in principle but even expert software engineers must design abstractions carefully for specific problems and find it hard to extend systems for new, unanticipated use cases. Other kinds of programming languages have different trade-offs. The Game Description Language (GDL) allows simulators for many finite games like Connect Four to be synthesized from declarative descriptions and has built-in support for several common game concepts such as players, legal moves, and goals. However, it is otherwise a very low-level language and has very limited support for building new abstractions. For example, there is no easy way to declare an \( m \times n \) grid; the the GDL description of Connect Four requires a distinct logical atom for the existence of every slot in the grid. Even more exotic kinds of languages such as probabilistic programming languages (e.g. Church) and differentiable programming languages (e.g. Tensorflow) make other kinds of software engineering dramatically easier, but don’t seem to help with this kind of programming.

I do not know of any existing kind of programming language that could make it so easy to build simulators for new games that building such simulators would be a plausible search problem for an AI system. I expect the following idea to be controversial, but I sometimes suspect that building AI may require inventing a new kind of programming language, vastly superior to those we have developed so far, that makes it much easier to engineer, refactor and extend complicated software systems. I think we may need such a language both as a resource to provide our AIs with, and as an encoding of the inductive biases we want to embue them with.


QuickSpec and the quest for good lemmas

9 minute read

Given a fixed set of proofs, it is natural to consider a lemma to be good to the extent it enables compression of the proofs. But how might we know if a lemm...

The general intelligence hypothesis

4 minute read

A friend of mine recently proposed the General Intelligence Hypothesis (GIH), which is a restatement of an old idea, and which has both a weak and a strong v...

Search vs meta-search

8 minute read

Consider the following puzzle, from Martin Gardner’s My Best Mathematical and Logic Puzzles (1994):

Mathematics: our overlooked ability

3 minute read

Suppose you have a list of 25 characters, 13 of which are ‘a’, and no character ever appears twice in a row. What can you tell me about the list?

Intelligence as software engineering

6 minute read

Humans can learn to become competent at playing board games such as Chess and Go, but humans are phenomenally excellent at learning how to play such games qu...

Abstractions in mathematics

5 minute read

The word “abstraction” is vague in most contexts within the AI community. Has AlphaGo learned abstractions? I don’t think the question is relevant. AlphaGo ...

Issues with AlphaZero

6 minute read

The AlphaZero algorithm developed by David Silver, Julian Schrittwieser, Karen Simonyan and others at DeepMind has proved very successful in the board games ...

Back to top ↑