Real Insights about Artificial Intelligence

by Andrew McAfee on May 7, 2012

A little while back Frank Levy, an MIT economist whose work I’ve drawn on a lot, and Seth Teller and John Leonard of the Computer Science and Artificial Intelligence Lab (CSAIL) came to an important realization: MIT is home to both a set of people exploring the economic implications of cutting-edge technologies like AI, and many of the top AI researchers themselves. So shouldn’t these two groups come together, get acquainted, and start swapping ideas?

We did so over lunch last Thursday at CSAIL. Because we didn’t discuss blogging groundrules I’ll not disclose the attendees here, except to note that my Race Against the Machine coauthor Erik Brynjolfsson was there. The conversation flowed freely for an hour and a half, and none of us felt like we’d come anywhere close to exhausting the topic, so we’ll do it again.

I left with a full brain, the lingering sense that I was the dumbest guy in the room, and a few clear learnings and impressions. The latter include:

  • The discipline of AI was Big Data before Big Data was cool. As the CSAIL researchers explained, the field of AI pivoted substantially starting in the late 80s and early 90s. Prior to that time, we were trying to teach computers rules: how language worked, how knowledge was organized, how the universe was structured. After then, we just filled them up with data and put statistical algorithms on top of them: If you see this waveform, it’s this word or this sentence; if you see this string of words in one language, it corresponds to this string in this other language; etc. This approach has only gained currency as the amount of digital data in the world continues to mushroom. (here and elsewhere I’m going to oversimplify the field of AI badly, and get important details wrong because of ignorance and/or a need for brevity. Apologies.)
  • AI algorithms get better as data gets bigger. Super large datasets let researchers test their hypotheses and ideas, finding out which ones actually work. This lets the self-improving and auto-correcting features of science kick in in a way they never could when the AI debates had to remain in the realm of pure theory.
  • We don’t need to worry about the Terminator or prepare for the Singularity. Computers are getting bigger and faster, but not ‘smarter’ in any human sense of the word. Artificial intelligence bears very little relationship to the human variety, and the two are not going to merge. One of the AI researchers referred to the idea of the Singularity as a ‘category mistake,’ which is a great academic insult.
  • Faith in human intuition remains strong. Too strong. One of my biggest surprises came as I listened to the AI researchers defend human decision-making and pattern-matching capabilities. As I’ve said repeatedly (here, here, and here, for starters), I think our intuition, while amazing and real, is also highly overrated —  clearly glitchy and biased — and should be replaced by algorithmic approaches as soon as it’s clear that the algorithms to a better job. Costs will drop and outcomes will improve as we get humans out of the loop in more and more circumstances. I found it very strange to be working to convince a bunch of world-class algorithmicists of the comparative advantages of algorithms, but I guess I should be used to it by now. Most people, including professional technologists, are enamored of intuition.
  • Nobody knows where we’re headed. When one of the AI guys walked into the room with a copy of our book, my first thought was not “Wow, that’s flattering.” It was “Uh-oh — does he expect us to tell him how the future’s going to unfold? ‘Cause that’s what I was hoping to learn from him…” I’m very sure that the economic and societal consequences of recent astonishing technical progress are going to be big, and I’m not at all sure what they are. Technology-fueled capitalist creative destruction is a messy and uneven process, and its twists and turns in coming years will leave us amazed.

I’m a huge optimist about this process overall (although I’m deeply concerned about some of the labor force implications; hence Race Against…) and I very much look forward to future conversations with leading technologists to help me understand it better. Erik and I are headed to San Francisco and Silicon Valley at the end of the month to continue these conversations. I’ll be sure to report back from there…


Steve Ardire May 7, 2012 at 11:15 am

Gee I didn’t know all the leading AI minds were at CSAIL 😉

Andrew McAfee May 7, 2012 at 11:19 am

Steve, I made no such claim. I said only that “MIT is home to… many of the top AI researchers”

Jed Harris May 7, 2012 at 11:50 am

Could you elaborate on the “category error” made by the singularity folks?  I agree there’s a category error but can’t tell if your AI folks and I see the same one. 

I think you’re missing a “not” in the sentence “Artificial intelligence bears very little relationship to the human variety, and the two are going to merge.”  You probably meant “are NOT going to merge.”  But the way you wrote it is also interesting — Vinge’s earlier position was that we’d see computers and people in symbiosis. 

Also you need to be a bit cautious here because of course AI researchers don’t want to have to worry about potential AI downsides, so they are somewhat motivated to find reasons why no problems can arise — though I also believe they are well intentioned enough to respond if such problems did seem likely.  Anyway you should check their reasoning. 

Jed Harris May 7, 2012 at 2:54 pm

Good to see you added the “not”.  I’m still interested in more details on the category error. 

Regarding “faith in human intuition”, I *guess* that you and the AI folks are talking past each other a bit.  One of the big (and continuing) lessons of AI is just how good people are at finding patterns in experience, and using those patterns to guide behavior — there are examples across the spectrum from picking up a glass of water, to using natural language, to playing chess.  Even when, as in chess, computers can do better now, they don’t do it by seeing and applying patterns. 

I’m sure all good AI researchers would absolutely love to have algorithms that do a better job.  Certainly Thrun, for example, is focused on using AI to build cars that drive themselves better than humans.  But they have also found it is extremely hard to find ways to do characteristically human tasks better than humans. 

You don’t give examples but I would guess that you are using “intuition” as shorthand for the kinds of heuristics that people use in cognitive tasks like those studied by Kahneman and Tversky — and many since.  Clearly in those cases human heuristics don’t do as well as the formal answers.  And certainly some important categories of real world decisions suffer from reliance on the same bad heuristics.  (Individual and business investment decisions are often good examples.) 

But to a large extent, these cases where humans do poorly are outside our “ecological niche” — situations very different from the ones most humans faced during the first 98% of our history.  Within our niche it is extraordinarily hard to duplicate human performance. 

Of course you should check, but I am guessing that you and the AI folks are using “intuition” and “algorithm” in very different ways — not simply different senses, but also a whole different set of background assumptions. 

Andrew McAfee May 8, 2012 at 2:05 pm

Jed, thanks for catching the typo. My follow-up post dives into the ‘category error’ in some more detail:

Anonymous October 6, 2012 at 7:24 am

intuition , IMHO , is the ability to make a decision based on limited set of datapoints. There is a strong context element involved in decision making with limited datapoints. Computers are good at decision making with a large set of datapoints.

Also – I would still go with intuition in a collaborative way with computer decisions. i.e use the results from machine learning as a datapoint to feed the intuition. this way you will get much better results

Comments on this entry are closed.

Previous post:

Next post: