What Computers Can't Do, and Arguably Never Will Be Able to Do

|Peter Boettke|

Listening to NPR one morning, this story about soccer playing robots came on and I was intrigued by the discussion.

When I entered graduate school and was assigned to Don Lavoie, my expectation would be that I would be working on questions related to the problems of economic calculation under socialism.  Of course, I wasn't disappointed as in the first 2 years of graduate study Lavoie was finishing up Rivalry and Central Planning, and National Economic Planning: What is Left?  But most of my work for Don at that stage of those projects was reference checking, his argument was already completed and written down.  So besides learning by reading his books, I learned how to work the open stacks at the Library of Congress and browse the University of Maryland library shelves to double check and track down accurate references that Lavoie was citing.  But in the process, Lavoie was also exploring the full implications of Hayek's argument about "knowledge of time and place" and the Hayek-Polanyi argument about "tacit knowledge".  This led him, and those close to him, into an examination of the nature of learning, the differences between technical knowledge and practical knowledge, and ultimately into an examination of both cognitive sciences and the philosophy of mind.

So in addition to learning Gould and Ferguson (Micro), Henderson and Quandt (Micro), Varian (Micro), Branson (Macro), Barro (Macro), Chiang (math econ), Silberberg (math econ), Theil (econometrics) and Kmenta (econometrics) during my first year, I was spending weekends reading not only Michael Polanyi but Herbert Dryfus, John Searle, etc.  One of the reasons this was so relevant to my core interests in the problems of economic calculation is that many economists at the time thought that super computers would one day be able to solve the problem. (e.g., Lange 1967).

The relevant point for economics turns on the differences between information and knowledge in understanding the economic forces at work under alternative institutional arrangements.  It is not the costliness of information that is the critical issue, and thus the incentives that actors face in acquiring and utilizing the bits and pieces of information that are scattered throughout the economy, it is instead the contextual nature of the knowledge that exists only within certain contexts and outside of that contexts does not exist at all -- and therefore cannot be discovered and utilized by actors in the economic system not matter how hard they search.

The enthusiastic hard AI would-be economic planner is ultimately conflating syntax with semantics in 'the grammar of the economic calculus'*.  Formalism can provide clarity with respect to syntactic knowledge but remains silient with respect to semantic knowledge.  In the affairs of men, however, it is semantics that we must uncover for understanding how the world works.  As Searle argued long ago now, machine langue is not the same as human language no matter how much we try to mimic it.

Well now go back to think about our soccer playing robots.  The effort is to mimic human playing soccer players, with the ultimate goal of competing against elite soccer players.  But in this effort, I would argue that the analogy to chess is misplaced.  It is not just the physical dexterity required to play a sport such as soccer at an elite level, but the decision making capacity of the players with so many degrees of freedom.  Chess is no doubt an extremely taxing game of mental capacity (perhaps the hardest such game), but it is also played within an extremely structured environment with only so many permutations and combinations.  In a sport such as soccess, or basketball, or tennis, the permutations and combinations are in some practical sense far more open-ended, and thus leaving greater scope for, and necessity of, "entrepreneurial alertness" to hitherto unrecognized opportunities for winning plays, counter-strategies, and new techniques and tactics.

It is not just physical dexterity that presents a challenge to what computers can do, but mental nuance.  Nuance and semantic knowledge go hand in hand, without access to semantic knowledge nuance is not possible.  So the quest for soccer playing robots will continue to be frustrating if judged against the standard of elite human players.  This does not mean that (a) robots will not play soccer (or any other sport) against each other in the future, they just will not successfully mimic elite human play, and (b) some sort of singularity momement where man and machine merge will produce a hybrid human player that will be stronger, faster, and more capable than the current human elite player.  All I am saying is that the computer will not substitute for the cognitively nuanced aspects of the human player -- computers will not take over those parts of the playing of the game that require nuance; that require the skillful utilization of the semantic knowledge of context; that are grounded in the unique knowledge of time and place.

Bob Subrick and I explore the implications of this sort of philosophy of mind argument in our paper "From the Philosophy of Mind to the Philosophy of the Market."

 

*This is the proposed title of an unpublished book of Hayek's that he was working on in the decades after WWII.

Comments

[Brandon]

Once computers can simulate a neural network, I believe your argument falls apart. Yes there are many many more permutations in an open field of soccer than there are on a chess board, but eventually machines will be capable of running the neural net that's in your head. It's an extremely difficult problem (you're going to need billions and billions running simultaneously, some with millions of connections to others, all being rewarded and reweighted in real time), but perhaps one that's not further than a century or so off.

[Barkley Rosser]

I mostly agree with Brandon and think you overstate the problems involved with soccer-laying robots. I would note that there is still on purely intellectual game that humans can beat computers at. That is Go. But the day will almost surely come when the computers will beat even the best Go masters, just as they pretty much can do now with the top chess players.

The problem is not just the locality of tacit information, which is a problem, but the one that Roger Koppl and I have written about that involves deeper levels of computability, putting the limit on what computers can do more seriously. In effect this involves the infinite regress models of non-computability that arise when the agent/[planner attempts to account for the impacts of his/her actions on the system. This runs into Godel/Hayek problems of self-referencing, which are also not unrelated to the Lucas Critique in its higher formulation, trying to take account of how behavioral response functions change in response to policy changes.

Lucas cheated in how he got out of that by assuming rational expectations, which is a cheap cheat: just assume that people expect what will happen, although there even with that remains the problem of multiple self-fulfilling prophetic equilibria, something Roger E.A. Farmer has been recently stressing in the econoblogosphere. How does one select which one to coordinate on, the focal point problem that got Tom Schelling his trip to shake hands with the King of Sweden given especially that he played such a role in focusing the nuclear armed part of the world on a point that says "No first use of nuclear weapons."

Anyway, these infinite regress problems associated with seriously trying to figure out best responses to a system that is itself trying to formulate best responses to what you do end up in infinite do-loops with halting problems. Oooops! This is a deep problem that real world central planners (and others following them) resolve by at some point simply throwing up their hands and rolling dice or throwing darts at boards.

[Roger Koppl]

Yes to all that Barkley said. On the soccer pitch there are only three dimensions of movement for the ball at a given moment. No innovations will change that. The rules are fixed. The players can move in only so many different ways. Soccer is a static universe. The space of possible games is vast but unchanging over time. There is no particular reason robots could not be programmed to make good choices consistently, and to do so in such a way as to regularly defeat fallible humans. Robots would be immune to mental fatigue, overconfidence bias, and egoism.

The "Austrian" argument on socialist calculation assumes a dynamic world in which you always have something new to learn. There is no calculation problem in a static world. Soccer is a static world.

[Peter Boettke]

Roger,

I don't disagree with either Barkley or you, but I think you are both not seeing the arbitrage opportunity that exists between the literatures on computability and the philosophical points about the contextual nature of knowledge, and the distinction between syntactic knowledge and semantic knowledge. I'd like to explore the possible complementarities between these arguments.

However, I do want to disagree a bit about soccer (or any advanced sport) that requires on the spot judgements constantly --- this is my points about degrees of freedom (in the machine sense, obviously not statistical sense).

Pete

[Roger Koppl]

Pete,

I think you may be right about the arbitrage opportunity. Certainly, computability is about syntax and not semantics.

On judgment in sports:
Humans have to use judgment. Machines can use algorithms. In some cases of judgment or intuition the unconscious part of our thinking (be it pre-conscious, meta-conscious, or something else) may be running an algorithm and just reporting the output to the conscious mind. Thus, some cases of intuition may be perfectly algorithmic. It does seem unlikely that most of the judgment in sports would be algorithmic in that sense. But even assuming zero algorithmic content to such judgments, it may still be possible for a computer with its algorithms to make better decisions that an athlete, just as a computer can make better chess moves than a human chess master.

[Roger McKinney]

I think Boettke is right, and I would add that very important aspects of sports is deception and being able to anticipate what other players do in response to what you do.

I was the neural net expert for a small software company years ago and was disappointed in what they could do. They no where come close to the hype of the industry.

I would guess that a lot of the computer models used to forecast weather are NN type and look at how bad they are. I would think predicting weather would be a lot simpler than playing soccer.