I’ve been slowly reading Brian Cantwell Smith’s “The Promise of Artificial Intelligence” recently. I haven’t finished reading it yet, and like much of BCS’s writing, it’ll probably take me 3 or 4 read-throughs to really understand it, but there’s one point that I want to pick up on. It is the idea that “Good Old-Fashioned AI” (GOFAI), in the famous term of John Haugeland, failed because it assumed that perception is an easy problem (relatively) that can be solved separately from cognition (intelligence):

“It is not random that classical first-wave AI was based on symbolic representation. Its approach, immortalized in Haugeland’s indelible phrase “good old-fashioned Artificial Intelligence” (GOFAI), grew out of four vaguely Cartesian assumptions:

C1 The essence of intelligence is thought, meaning roughly rational deliberation.

C2 The ideal model of thought is logical inference (based on “clear and distinct” concepts, of the sort we associate with discrete words).

C3 Perception is at a lower level than thought, and will not be that conceptually demanding.

C4 The ontology of the world is what I will call formal: discrete, well-defined, mesoscale objects exemplifying properties and standing in unambiguous relations.”

—The Promise of Artificial Intelligence, ch.2

BCS argues that machine learning (“second-wave AI”) is a better way of tackling intelligence in general. (His arguments go way beyond the narrow point I’m going to pick up here, with some very interesting insights—this is quite a minor issue in his general thesis. He attacks C4 particularly powerfully).

Firstly, BCS takes Haugeland’s characterisation of GOFAI as unproblematic, but see for example Drew McDermott’s argument that it was essentially nonsense: GOFAI Considered Harmful (And Mythical).

Secondly, if machine learning (by which we both mean neural nets specifically) is an argument for anything at this point, it is that perception is a separable concern! The things that machine learning excels at are precisely turning messy perceptual inputs into neat symbolic categories. From the very first triumphs of deep learning to the recent applications across a range of fields, neural nets are fantastic at perception. But they perform much less impressively on any tasks involving reasoning or deliberation. LLMs are bullshit machines (in a strict sense of that term). Even apparent successes of machine learning to the traditional realm of GOFAI, such as AlphaGo and its successors, rely on machine learning primarily to pattern-match and score board positions, and use classic symbolic state-space search (in the refined form of Monte Carlo tree search) to actually choose what move to make.

(Update January 2025: As Rodney Brooks points out, even for perception, modern ML systems are less than ideal, citing this quote:

“The kinds of things that tend to go wrong with these systems are things like it was not trained on, pictures of an overturned double trailer. It just didn’t know what it was. There were some lights there, but the lights were in unusual positions. A person would have clearly said something big is in the middle of the road. But the way machine learning works is it trains it on a bunch of examples and if it encounters something it doesn’t have a bunch of examples for it may have no idea what’s going on.”)

It’s been 15 years since I finished my PhD in AI, and I haven’t been involved in the field for most of that time. But when I talk to academics still in the field, it seems there is a lot of talk about hybrid approaches that combine symbolic and ML techniques. Essentially, the ML does the messy perceptual stuff and then the symbolic part does the reasoning. If this is not a paradigm of GOFAI, I’m not sure what would be!

None of this is to say that I think GOFAI (if it exists at all) is a good approach, but I think perception is the wrong lens to view it through. The real “second wave” AI was not neural nets (which have been around since the start), but rather Brooks’s “Nouvelle AI”. This approach rejected explicit symbolic representation in favour of tight feedback loops between perception and action (“the world is its own best model”). This approach also ran into limitations, and hybrid approaches emerged as the solution there too. But the idea that intelligence is grounded in embodiment has had a much more powerful and lasting impact on how we conceptualise intelligence than anything since. It seems likely that the very first nervous systems evolved to support movement, and that almost all subsequent developments were to support more complex interaction with the outside world (finding food or mates, avoiding predators, social interactions, etc). Perception is in service of interaction. But progress in explicitly embodied enterprises like self-driving cars also seems to have stalled, running into the hard problem of general intelligence and “common sense” reasoning.

Russell and Norvig describe three broad types of representations in AI:

• Atomic representations like those of finite-state machines and state-space search. The state is an indivisible whole and you can’t say anything about its composition.
• Factored representations break up the state into a finite set of variables, each taking on a simple (scalar) value of some kind. Boolean circuits and SAT solvers fit in this category, as do constraint solvers, Bayesian networks, and neural nets.
• Finally, structured representations such as formal logic, most programming languages, and relational databases, where the world is modelled in detail as collections of entities with properties and relationships between them.

McDermott argues in the essay I linked earlier, that early AI attempts were based on the latter, most complicated, representations. But actually (in 2015) all the recent successes of AI were based on factored representations: SAT solvers, neural nets, Bayes nets, etc. I think this is basically true, but that everywhere we seem to be running into the limitations of this approach. Personally, I don’t think we can tackle intelligence without structured representations, and maybe we are ready to look anew at that challenge. SAT solvers evolved into SMT solvers, which combine the strengths of the former with more complex representations (still largely factored). Logic programming and Datalog (Prolog-lite) is seeing something of a resurgence. These kinds of logical, symbolic approaches have plenty of conceptual and technical difficulties, but I think we need to tackle those rather than declaring them insurmountable. The history of AI is throwing babies out with the bath water. Hopefully this time we can learn the lesson that hybrid architectures that combine a range of techniques are the best.