We all know the first kind of centaur. The other kind is a human-computer hybrid, and comes from freestyle chess. This quote from The Economist explains the origin of this second meaning:
[In] chess, […] the best players in the world are not machines however, but what Garry Kasparov, a grandmaster, calls “centaurs”: amalgamated teams of humans and algorithms.
I predict human-computer hybrids will be a powerful force in the near future, for many domains. (In a way, this is not much of a prediction, if you count me using this computer to write an essay as a human-computer hybrid.)
Pandora’s Computer
An atmosphere of anxiety pervades society these days, with many wondering about the threat of automation. What strikes me as slightly odd is that nobody discusses past case studies in automation. We’ve consistently invented things that could put people out of jobs. We know a lot about what typically happens. Perhaps AI is different, but if it is, it’s still nice to have a reference for how different. In any case, if AI is comparable to technology we’ve seen before, then history is quite comforting. Typically what happens is: workers get much more productive and efficient, prices drop, and demand increases, and then economies of scale multiply the effect and prices drop more. Sometimes the effect of all this price dropping is a surge in demand large enough to increase employment in the relevant industry. A second, frequent impact is availability for humans to get away from boring and soul-sucking tasks. This is also good. On the downside, some employees are hurt in the short term, and they are the ones who have focused expertise in the skill that gets automated. For society at large, this is a step forward, and the gains are often large enough that we could compensate those who are hurt in the transition, and still be at a major advantage.
Humans are not pushovers when it comes to difficult environments. We are pretty adaptable, and we are good at using tools to make ourselves better off. I am one who believes AI is manageable by humans, but I won’t try to convince you. There are many who fear that AI could easily get away from humans, and I’ll leave it up to you to decide. My focus is the short-term. I think what we’ll see soon is humans figuring out more and more ways to use increasingly capable machines to clear todo lists of overhead, so they can focus on the fulfilling stuff: art, strategy, philosophy, research.
Hybrid Systems
Hybrid systems are excellent for many contexts, but certain contexts especially favor them. Those are situations in which the two component systems have crossing performance curves, and when the cost of switching between them is low relative to the performance differential. Consider this graph:
I made this up, but I think it captures the idea pretty well. At low quantities, 3D printing is great. It’s easy and cheap to make a prototype. As quantities grow, 3D printing is costly, mostly in terms of time. Injection molding, on the other hand, is costly to set up, but for large quantities it’s much better.
In situations like these, the idea is to get the curve defined by the lowest of either one at any point (in the above, the kinked curve that is blue then green).
For human-computer hybrids, the performance curves may not be so organized always, but there are clearly some cases that aren’t too far from the above, just with new labels. Replace ‘Quantity’ with ‘Degree of Abstraction’, for example, and the human and computer performance graph would be pretty similar to the one above: for concrete, well-specified, repetitive tasks, machines are great. As the tasks become more abstract, like, say, presenting new research to an audience, humans are much better.
This is encouraging. Hybrids between humans and computers ought to work, but we need to better understand how the curves look, and even what the relevant axes are.
Creative Centaurs
I’ve written about this before, but one hybrid between computers and humans that I find particularly intriguing is the brainstorming tool. Humans are very good at associating concepts to one another, translating ideas into concrete specifications, and explaining ideas to other people. Computers are good at producing random stimuli, exploring large quantitative spaces by brute force, and carrying around large amounts of data in memory.
There should be a fantastic hybrid system, a creativity centaur, if you will, waiting to be discovered. I’m hoping to find it. I’ve done some experiments already, but there is a long way to go yet.
Centaurs Elsewhere
Though I have an interest in human-computer hybrids for increasing creativity, I don’t believe that is actually the most useful application at the moment. I think it’s healthcare.
The late Robyn Dawes, while a researcher at The University of Oregon, wrote a marvelous (in my opinion) paper called ‘The Robust Beauty of Improper Linear Models in Decision Making.’ The main idea is that statistical models implemented by computers are much better at prediction than doctors are, but that doctors are much better at deciding which variables should enter the model, and how they should enter the model. To make it concrete, doctors should decide which tests to run after talking to a patient, and the computer should determine whether, given the test results, the patient actually has the disease. Currently, this doesn’t happen. It would probably improve health outcomes, and efficiency, if it did.
A wrinkle in this discussion is that the model can only process data it has been trained on, so anything that is hard to measure, or just hasn’t been measured, won’t enter into the model’s considerations. A great doctor would know when to override an uninformed model. On the other hand, a not-so-great doctor would probably deviate from the model too much. These problems seem surmountable, given time to work them out. It’s a moot point anyway, since statistical models aren’t even included in most diagnoses.
Conclusions
I’ve meandered a bit, but the main point is that centaurs are coming. You should try to embrace them by understanding the performance curves of computers, and getting good at the parts they are bad at. Don’t try to beat computers where they excel, just get out of the way. The place you end up will be more enjoyable and fulfilling anyway, since it will be a place more uniquely human.
Note: Of course, some would argue that there really is no safe place to go, nothing humans are better at. Maybe at some point in the future this will be true, but given the state of the art right now, that’s ridiculous. Humans are much, much better than computers at a lot of things. As a rule, if it involves heavy amounts of interpersonal interaction, communication, persuasion, abstraction, or diversity of skill, humans are currently better at it, and will be for the foreseeable future. Think sales, teaching, research, art, healthcare, and entrepreneurship, for a few examples.