By James Pethokoukis and Melanie Mitchell

From defeating chess grandmasters to generating images using only keywords to predicting protein shapes, artificial intelligence has come a long way since the early computer era. And in that time, the techniques behind AI systems have evolved, right up to today’s “deep learning” algorithms. And now, with some dismissing the promises of AI enthusiasts as empty hype while others predict a runaway technology that will destroy jobs or worse, is there any room for optimism? To find out more about the history of artificial intelligence research, the present state of AI, and what might be coming next, I’ve brought on Melanie Mitchell.

Melanie is the Davis Professor at the Santa Fe Institute, a non-profit research center for complex systems science. She is the author of six books, her latest being “Artificial Intelligence: A Guide for Thinking Humans,” released in 2019. In 2021, Melanie authored “Why AI is Harder Than We Think,” which describes the fallacies that underlie overly optimistic AI predictions.

What follows is a lightly edited transcript of our conversation, including brief portions that were cut from the original podcast. You can download the episode here, and don’t forget to subscribe to my podcast on iTunes or Stitcher. Tell your friends, leave a review.

Pethokoukis: At the beginning of your book, “Artificial Intelligence: A Guide for Thinking Humans,” an excellent book, there’s a portion where you talk about surveying the many opinions about the state of the technology. The listeners love when I read, so let me read just a small portion from your book. You write, “In short, what I found is that the field of AI is in turmoil. Either a huge amount of progress has been made, or almost none at all. Either we are within spitting distance of ‘true’ AI, or it is centuries away. AI will solve all our problems, put us all out of a job, destroy the human race, or cheapen our humanity. It’s either a noble quest or ‘summoning the demon.’” The book came out in 2019. Have we made any progress in figuring out which of those questions and the answers are most applicable?

Mitchell: I would say we’ve made
some narrow progress, but I think all of those questions are still wide open.
So the field moves very quickly in some ways, but in other ways it moves quite
slowly on these bigger questions.

There are times where the sector seems to be moving very
rapidly and there’s a lot of interest and there’s a lot of funding. And then
there’s times where it doesn’t. Because when I really started paying attention
to AI, it was a lot about expert systems and neural networks — not that I could
probably define any of those particularly well — and now it’s about something else.
So first, if you just walk us briefly through the kind of AI that’s been in
favor. And then today, when we talk about AI, what are we talking about? So if
I read AI in the newspaper or the business section, what are we talking about
versus maybe what we were talking about 20 years ago?

The
term AI has changed its meaning throughout its history. It started out very
much trying to use logic and logic-like deductive inference as the way to model
intelligence. So people figured out that computers could not just process
numbers the way we normally think of them doing, but they could also process
language and they could process proofs like mathematical proofs, deductive
proofs. They prove theorems and so on. These things like processing language,
processing logic, logical structures have been called symbolic AI because it
deals with symbols like words rather than numbers that are harder to interpret.
And symbolic AI be was really the first big push in AI trying to capture
intelligence by explicitly programming logical abilities into machines. And
that kind of failed. For one reason, it turned out that you couldn’t just
deduct everything about the world. We needed some kind of knowledge. You know,
humans have a lot of knowledge.

And
so that’s when people started building these so-called expert systems where
they would go out and interview experts in a field (like, for instance, medical
diagnosis or something), try to get all the rules that this human expert used
to perform the task of diagnosis or whatever the task is, and then program
those rules into a computer. So those are called expert systems. Those also
failed to a large degree because it turns out that a lot of the actual rules
that experts use, or the knowledge that they use, is not conscious. They’re
using unconsciously a lot of their sort of so-called common sense. And they
weren’t able to express that in a way that could be sort of programmed into a
computer.

If you went out and interviewed Warren Buffet the
investor, and you just sat with him for a few days and “How do you do,
what you do?” I’m sure you would capture a lot of very interesting and
useful information. But then if you program that, you wouldn’t have AI Warren
Buffet, because it’s likely that there are things that he does that he maybe
doesn’t even realize he’s doing and connections he can’t, perhaps, obviously
articulate that are fundamental to what he does.

That’s
exactly right. This became a big problem for that whole thrust of AI. I’m
talking about like the 1960s or the 1970s, 1980s, and then neural networks,
which simulated in a very rough sense the way the brain works with simulated
neurons and simulated connections between the neurons, became popular. People
had been working on them for a long time, but they became much more popular
like in the later 1980s, 1990s.

And
these were systems that you didn’t program. They learned from data, from being
exposed to data. What they were processing was not symbols, but numbers again.
This became what was called the time “connectionist AI,” and kind of
morphed into a more general machine learning approach where, instead of
programming intelligent behavior in, you let the machine itself learn from giving
lots, and lots of examples.

So in your Warren Buffett analogy, instead of interviewing Warren Buffett and trying to figure out what rules he uses to do what he does, you would instead get a lot of big data about how Warren buffet invests and how he moves money around and so on. And you try to have the machine figure out its own rules just by looking at that data. So that’s machine learning. That has been quite limited, too, until about the early 2010s when this whole area called deep learning came about, and that’s using a kind of neural network system that is much more complicated than anything people used in the 1980s. And it learns from a huge amount of data that’s now available to people because of the internet and we are able to process this using these extremely fast parallel hardware. And that’s really allowed the whole field of AI to kind of explode with a lot of new applications and successes, although, as I’m sure we’ll talk about, it has its limitations, too.

Sort of the keys here: sounds like a massive amount of
data, far more data, much faster computing power. And then, I imagine the
programs themselves is more sophisticated.

That’s
correct. Exactly.

Generally when I hear people talk about AI, they’re
talking about machine learning and then a kind of machine learning called deep
learning.

That’s
kind of what it means now for the most part, although that’s very different, of
course, from what it meant like in the 1960s and ’70s.

We’ve talked about these ups and downs and people call
them “springs” and “winters.” What is the season we’re in
right now?

We’re
now in an AI spring. So the idea with that is, it measures how optimistic
people are, how much funding there is, the predictions people are making about
near-term artificial intelligent cars and robots and so on. Often these AI
springs are followed by AI winters where the promises that people are making
like, “You don’t have to get a driver’s license anymore because you’ll be
driving around in a self-driving car” — those don’t happen. The promises
are not fulfilled, and the funding dries up and people become disappointed and
think, “Okay, AI doesn’t work.” So we get these cycles where there’s
some new technology that people use that has a lot of promise and people often
overpromise its applications, and there’s a lot of optimism until suddenly
people become disappointed. And then AI winter happens. So there’s a lot of
debate now over whether we’re going to have an AI winter after this very
exuberant AI spring.

What makes people exuberant? Is it purely a technological
exuberance, merely sort of a fascination at scientific and technological
advance? Or is it how that could be translated into other new technologies or
new jobs or other conveniences for our life?

I
would say it’s more the latter than the former. There’s not really any huge new
scientific advance beyond what you said of just like huge amounts of data and
fast computers and maybe more sophisticated machine learning programs. But it’s
not that different technology than what was used like 30 years ago. We just
have more data and faster computers. But that enables these technologies to go
from just academic exercises to being actually applied in the real world.
Because you see that all over the place: We have facial recognition systems, we
do have self-driving cars that have their limitations, but they can recognize
pedestrians and they can figure out traffic lights and traffic signs and so on,
and all kinds of applications in terms of gene translation, AI applied in
healthcare, being able to diagnose certain diseases. You just read about
something new every day that AI is doing and that it’s doing in the real world.
So I think that’s the real revolution.

And if you had exposed someone in the late ’90s to what AI
can do today, would they have been disappointed or would they think, “Wow!
Those are really great advances, things must have really progressed over the
subsequent 20 years”?

I
think most people would have been very impressed and would think that AI has
made a ton of progress, but they also have to be made aware of some of the
limitations of these systems. One of the things that happens in these AI
springs is that we get these systems that can do certain things, they’re more
successful than they were in the past, and then people start making predictions
and saying, “Well, 20 years from now, we’re going to have robot house
cleaners and we’re going to have AI systems that can function on their own and
drive around and maybe even flying cars that can fly around on their own”
and they make these great predictions for near-term technological advances. And
then the technology they promised turns out to be harder than people imagined.

That’s a fantastic segue to your paper “Why AI is Harder Than We Think,” which is an excellent companion to your book. In that paper you talk about the difficulty of common sense. First, how do you define common sense? And is common sense something we’ll be able to instill in AI at some point?

Common
sense is this kind of umbrella term that means all of this background knowledge
we have that we’re barely aware of, of how things work in the world. So here’s
an example of a common sense failure in AI: So there was a self-driving car
driving on the road and it kept slamming on the brakes and stopping at a
certain point and the human in the car couldn’t figure out what was going on.
Why was it stopping? And it turned out there was a billboard that was like an
anti-drug billboard with a picture of a sheriff holding up a stop sign saying
like, “Stop using drugs” or something. And the car was interpreting
that billboard as an actual stop sign and stopping. And no human would do that.

No five-year-old would do that.

We
know the difference between real stop signs and billboards, and we can
interpret these different things about the world. And this is an unusual case.
People will call these unusual situations “edge cases.” They’re the
things that maybe aren’t in the training examples that this car was trained on,
because you don’t encounter something like that that often. But if you think
about all the possible edge cases and all the possible cars, it’s just
unlimited. And we deal with this kind of thing all the time by using our common
sense, if you will. And so the question is how do we get common sense into
computers?

And
one of the very earliest papers published in AI (I forget exactly when, but in
the 1960s, maybe) by John McCarthy was on how to give computers common sense.
And his solution was a very logic-based approach. And over the years, many
people have worked on this problem of common sense. Some people by trying to
build these giant database bases of common sense knowledge, trying kind of an
expert system of common sense where humans type in statements like, “You
can’t be in two places at once” or “If you walk somewhere, your body
moves from one place to another” — the very basic things that are never
written down but just everybody knows. I think the biggest open problem in AI
is, how do we give machines this kind of common sense we have to deal with the
world? And no one has solved it.

In “Why AI is Harder Than We Think” you address four fallacies that can make AI seem easier than it really is. One of those fallacies is “easy things are easy, and hard things are hard.” Could you explain what you mean by that?

There’s
certain tasks that we humans think of as very hard and take a lot of
intelligence. And one example might be playing chess at a grand master level.
We deify these chess players who can play chess and we think of that as
requiring a huge amount of intelligence. And yet it turns out that the game
chess is much easier for computers than a game like tag that you might play on
a playground. Because robots have trouble navigating, they have trouble often
tracking where people are. They have trouble predicting their movements and so
on. The easiest game for a four-year-old child turns out to be much harder than
the hardest game for a human. So this is this idea that things that are easy
for us often are hard for computers. And so if a computer does something that’s
really hard for us, we assume it’s going to be able to do all the things that
are easy for us, but that’s actually not the case at all.

Another fallacy you address in the paper is the allure of
wishful mnemonics. This one really struck me because it seems that scientists,
in order to explain complicated concepts to non-experts, use terms like
“learning.” But an average person doesn’t understand learning in the
same way a computer scientist might.

We
say “machine learning,” and we anthropomorphize that term and say
it’s similar to human learning. And yet it’s really different from human
learning. Because for one thing, if you a child learns something, then you
assume they’ll be able to apply that knowledge in other contexts than where
they learned it. If they’ve only seen dogs outside and they learn what a dog
is, they can still recognize a dog when it’s inside. This is not necessarily
the case for machine learning. This is one example of a wishful mnemonic where
it’s just a term that we use to describe something in machine intelligence that
also applies to human intelligence and we assume that the meaning carries over
from one to the other. Another example is neural networks. We talk about neural
networks as being like the brain, they’re actually quite different from the
brain, but that term “neural” sometimes gives people the impression
that they’re more like the brain than they are.

When you say neural network, that’s what I’m thinking. I’m
thinking of like a brain made out of fiber optic cable. And the final fallacy,
which is that intelligence is all in the brain. What does that mean? Why is
that a fallacy?

It’s
hard to understand. It’s controversial. We talk about artificial intelligence,
and when you think of artificial intelligence, you often think, “Oh, the
computer on your desktop might be thinking. It’s reasoning. It’s trying to
solve problems and so on.” So we assume we could possibly take a brain,
put it some kind of vat away from the body and it would still be able to think.
Well, there have been a lot of people in cognitive science that say that’s just
the wrong way to think about intelligence, that intelligence is not just in the
brain. It’s also how the brain and the body interact with the world together. A
lot of the things that we think about, we can only think about because of the
kinds of physical experiences that we have in the world with our bodies that
were embodied. Our intelligence is embodied.

Now,
not everybody agrees with this view, and there’s a big kind of debate in
cognitive science about how much you can separate intelligence from the rest of
the body. I think that if we assume that intelligence is all in the brain, we
don’t have to worry about anything else about the body or social interactions
or emotions and other kinds of things we don’t necessarily associate with
intelligence but probably are, we overestimate how fast we’re going to be able to
achieve artificial intelligence.

Those are some fallacies, and in the paper and the book
you get into a lot of the remaining big, open questions. It sounds there are a
lot of open questions and there’s much further to go. And certainly, I think
after reading the book and the paper it seems we’re still very far from
human-like artificial intelligence.

Well,
this gets back to the passage you read at the very beginning in the book where
I say, “We’re either spitting distance away or we’ve made no progress at
all.” And what I meant by that is these are two different views in the
field of actual practitioners in AI. Some of them believe that we’re very close
to getting the breakthrough that will give us human-level robots. And some
people think we’re centuries away, if not that it never will happen. And so
there’s that kind of divergence of views. Nobody can say for sure. We just
don’t understand, in some sense, what we’re aiming for. What intelligence is,
is the real problem. And I think nobody really knows. So we have a science of
artificial intelligence, and yet we don’t understand what intelligence is.

So for the people who are very optimistic, maybe far more
optimistic than you might be: With that optimism, does there also naturally
come fear? If you’re optimistic about where the technology is, then you’re also
fearful that the technology could somehow get away from us or eliminate all the
jobs or whatever scenario. Does the super optimism go hand-in-hand with
concerns? Now Elon Musk is optimistic. But he also has a lot of scary stories
to tell.

I
would say that there are some people who are very optimistic of that in the
sense that they believe we’re on the brink of creating true AI in some sense.
And a lot of people who believe that are also worried about AI systems not
having the same values as we do. And they talk about alignment of AI’s values
with us. So there’s a group called the AI alignment movement. There’s also
people who are like me, not as sanguine that we’re going to get to full
artificial intelligence anytime soon, and yet still fear are some of the
current issues in AI that come up like bias in these machine learning systems
and the fact that some of the systems that are being granted autonomy really
aren’t smart enough to have that kind of autonomy. So sort of the opposite of
the alignment people. It’s not that they’re too smart, it’s that they’re not
smart enough.

Can I be excited about this science and how it can help
people, even if I think we’re not going to get to some sort of human-like, much
less superhuman-like, artificial intelligence this century? Can it still have
enough of a positive impact to think that this is an extremely important and,
if used the right way, potentially extremely beneficial technology?

I
think it has the potential to be extremely beneficial. We already see some of
the benefits of AI. For instance, just recently the company DeepMind, part of
Google, was able to train a system that was able to predict protein structure
better than anything that had come before. So predicting protein structures is
like the first step in being able to design new drugs for diseases and being
able to do a lot of good in healthcare. And also just in understanding the
science.

I
think AI will eventually revolutionize healthcare. You think of a lot of people
who are unable to obtain healthcare or good healthcare, they live in certain
areas where they can’t obtain it. I think AI could have a huge impact on that.
I think also AI could have a lot of very positive impact on automating some of
the more dangerous jobs that people really shouldn’t be doing, and more
intellectually, helping us understand our own intelligence better and maybe
understand our own biases and the cognitive traps we fall into. Helping us
think better.

And since I work for a think tank, I always like to ask
one policy question. You can give me your best answer or no answer. What do you
want government to do to help? Do you want them to help in some way? Get out of
the way in some way? Any thoughts on that?

I
think a lot of AI policy, if there is any, has been left to big companies. So
big tech companies have decided which applications to deploy and have been
tasked with making sure that they’re safe and have the properties that we would
want them to have. And they haven’t really stepped up to the plate on doing
that. So I do feel like there’s some kind of role for government in regulating
AI in a similar way that it regulates say genetic engineering or other kinds of
biotechnology. We don’t want to shut down research because it can have a lot of
benefits, even though it can have a lot of downsides, too. But we do need something
outside of our kind of corporate structures to create regulations that will
kind of keep the research on a beneficial rather than harmful track.

Melanie, thanks for coming on the podcast.

Thanks. I enjoyed it.

James Pethokoukis is the Dewitt Wallace Fellow at the American Enterprise Institute, where he writes and edits the AEIdeas blog and hosts a weekly podcast, “Political Economy with James Pethokoukis.” Melanie is the Davis Professor at the Santa Fe Institute, a non-profit research center for complex systems science. She is the author of six books, her latest being “Artificial Intelligence: A Guide for Thinking Humans,” released in 2019.

The post Why artificial intelligence is harder than we think: My long-read Q&A with Melanie Mitchell appeared first on American Enterprise Institute – AEI.