A.I. Wiki

A Beginner’s Guide to Important Topics in AI, Machine Learning, and Deep Learning.

The Next AI Winter - Deep Learning and its Discontents

An AI winter is a time when support for and interest in artificial intelligence research and commercial ventures dries up. AI went through various winters in the 20th century, when early promise turned to disillusion, and the best technologists bent their minds to other problems.

Though many definitions of AI are overly vague, it is an important technology that receives a lot of media attention. That means a lot of people have opinions about it even if they are unfamiliar with how it works and what its limits are. This post gives you a framework by which to interpret opinions about AI, to situate someone who holds an AI opinion stands and understand who their allies and opponents are. Like most divisions of opinion, these differences roughly break down into “I like AI” or “I don’t like AI”, although there are several reasons why people have chosen not to like it.

Broadly, there are two kinds of AI pessimists: One believes AI will become so strong that it destroys humanity, and another believes AI is so flawed, brittle, and overhyped as a technology that the sector will collapse. The first is afraid of killer robots, the second predicts that the next AI winter is on its way. These we could call the “winterists” and their ideology “AI winterism”. Both kinds receive a lot of attention, although their claims are mutually exclusive.

Everyone likes prophecies of doom: messages that arouse fear are more likely to go viral, because doomsday prophets play to our insecurities. Like a politician seeking re-election might launch a war on terror or a war on drugs, fearmongers standing at the edge of AI are waging a strange war on the technology. Their message of decline implies that they are superior due to their foresight, and that the rest of the world is headed blindly for disaster.1

For reasons I have never understood, people like to hear that the world is going to hell. - Deirdre McCloskey (historian)

I have observed that not the man who hopes when others despair, but the man who despairs when others hope, is admired by a large class of persons as a sage. - John Stuart Mill

Polarization of Opinion about AI

Discussions of AI are multipolar and increasingly polarized. That is, people are gravitating toward more and more extreme views about the future of AI and its impact on society. In addition to the two kinds of pessimism described above, you have AI hype. The boosters who seek to succeed alongside the technology. It is in their interest to make promises about AI, and to view a world with stronger AI through rose-colored glasses.

At the center of all this sturm and drang, far from the poles of fear or hope, the hard, practical work of researching and implementing AI continues, pushed forward incrementally and independent of opinion, by researchers who are as aware of AI’s limitations as its capabilities.

Politicization of AI

The issue of AI is becoming more and more politicized, and the political polarization has multiple axes. Political polarization has led to partisanship, paranoia and a loose relationship with the facts. In such situations, it’s easy to see how both sides are wrong, or at least using unsound arguments.

Within most developed countries, there are those who believe that AI will destroy jobs domestically through automation vs. those who believe that AI will create new industries and offset that unemployment.

Internationally, there are those who believe the major powers are in an AI arms race vs. those who believe that AI is developing fast and well based on the open sharing of knowledge among research groups.

In both cases, paradoxically, both sides are right.

AI Winterism

The main problem with AI winterism is that it seems to arise more from AI hype fatigue than actual contact with contemporary AI research. Sure, hype comes and goes, but AI research is moving so fast that most insiders have trouble keeping up, and AI capacity is tied to brute compute, which is also progressing.

While fundamental AI research research and complementary technologies like raw compute are making progress, such advances are rarely linear. They come in bursts, and so does the media attention that is paid to them. You might say that AI hype is both cyclic and recurrent. The hype cycle doesn’t happen just once. And each advance moves the base a little higher, a sinusoidal curve on an upward tilt. AI winter is a strong metaphorical frame that needs no explanation, and it implies that AI, as a technology, is doomed to be deflated. But you could reframe it by saying that each AI winter is like a cold season occurring on an ever warmer planet. Sure, the temperature goes down, but never quite as much as the last time.

A History of Winter

A series of setbacks between the mid-60s and the mid-70s led to the first, true AI winter. Serial disappointments in AI’s performance in machine vision, machine translation and basic tasks led to cutbacks in government funding in both Britain and the United States. Notable among these setbacks were:

Interest in symbolic reasoning saw a resurgence in the 1980s, a time when expert systems were commercialized, but that trend collapsed by the early 1990s, as it became apparent that immense, rules-based systems were often brittle, hard to maintain and of course, unable to learn. This period is best exemplified, perhaps, by Cyc.

By the mid-2000s, that winter, too, had ended, as Geoff Hinton and his disciples began to show the first, promising results of deep neural networks.

AI Summer

We are currently in one of the longest “summers” in the history of AI development. Performance gains, theoretical advances, and significant funding from the public and private sectors have been consistent for several years. Unlike previous AI summers, this one has delivered major breakthroughs on several fronts, enabling a variety of other technologies such as autonomous vehicles.

Those who believe the AI summer will continue fall into two camps: those who believe it will bring mostly good things (the boosters), and those who are concerned that the breakthroughs will lead to major changes that we are not prepared for (the millenarians).

AI Winterists

AI Millenarians 2

AI Boosters

  • Most AI startups and large technology companies

AI Pragmatists

Footnotes

1) In financial markets, these pessimists are called bears, and their optimistic counterparts are called bulls. Bulls and bears both have biases that color their predictions, and lead them to seek evidence to support their beliefs. Being a bull or bear is not so much a matter of being right or wrong as never being right at the same time as your opposite number. There’s usually someone who calls the market right, and they play musical chairs with everyone else. When you play the stock market, timing is everything. Being a bull during a bear market is called “trying to catch a falling knife.” It hurts. In tech, timing is fuzzier, because the investors’ time horizon is longer. That’s why VCs and futurists can go around saying augmented reality and hyperloops are going to eat the world without being laughed out of the room. They might be right in a couple years. Or they might be wrong. But their feedback loop is a lot looser, which means that tech gets more unaccountable prophets than Wall St.

2) These people and organizations believe that society will be fundamentally transformed by AI, and more specifically, that it could be negatively transformed by a superintelligence. As outlandish as their views may seem to outsiders, it is hard to argue that creating a being more intelligent than humans in every way would not affect human society. Whether creating such a being is possible, and when it might happen, are the largest uncertainties in this debate.

Chris V. Nicholson

Chris V. Nicholson is a venture partner at Page One Ventures. He previously led Pathmind and Skymind. In a prior life, Chris spent a decade reporting on tech and finance for The New York Times, Businessweek and Bloomberg, among others.