This past summer, OpenAI released GPT-3, the latest iteration of their research into AI language models. As happens every time there is an advance in artificial intelligence, right away, we saw the breathless and sensational statements in the press that over the past decades have accompanied computer wins in chess, Jeopardy and Go. These range from the passé “the future has arrived” to the clickbait scare-mongering about dystopian futures where poem-writing robots take all our jobs. Unfortunately, this misguided hype leads to costly strategic mistakes, even in investment and technology circles where folks should know better. The hype gets ahead of science, and the dumb money follows it.
Take self-driving cars. According to Uber’s own IPO filings, it can only be profitable if it succeeds in fully autonomous vehicles – replacing every public transit ride in the world with them. Self-driving cars is also an example of a use case that would require solving at least some of the very basic building blocks of artificial general intelligence (AGI). Within a constrained setting (public roads), a self-driving car needs to observe a complex environment, reason about what actions to take (don’t hit the cyclist!), and take action (swerve or stop). Without fully autonomous self-driving cars, there’s no chance of solving true AGI. Put another way; we need to see fully autonomous vehicles driving around Manhattan long before we can expect to run into Westworld’s Dolores Abernathy in Central Park.
Yet, fully autonomous vehicles in complex settings remain out of reach. This past September, The Information reported Uber’s $2.5b wasted effort and failures in its Pittsburg-based unit. Ford similarly spent $4b up till 2019 on a self-driving car effort and planned to launch a self-driving fleet this year. Large tech companies like Uber are placing their future in AGI and relying on the hype around it to retain investor confidence, hoping it buys enough runway to get them there. The problem comes when that’s decades away.
Last April, to debunk some of the hype, my colleagues and I at McKinsey published an executive primer to artificial general intelligence giving a ground-up overview of the research and state of the art in the field. More recently, the MIT Technology Review published an article to do the same. Whereas these all do a good job explaining the technology and roadmaps involved, it doesn’t address the underlying reasons and psychology around the hype.
Most of us are now aware that foreign actors are actively using disinformation campaigns to disrupt our elections (therefore, don’t get your news on Facebook!). Similarly, a better understanding of the underlying dynamics driving hype might help navigate the disinformation around complex technical topics like AGI. Here are some thoughts on the causes.
1. The False Prophets: Futurists and Journalists
In a blog post titled “The AI Misinformation Epidemic,” CMU machine learning assistant professor Zachary Lipton laments that “the outsize demand for stories about AI has created a tremendous opportunity for impostors to capture some piece of this market.” “This pairing of interest with ignorance has created a perfect storm for a misinformation epidemic.” He complains that “with startling regularity, if I tell an educated person what I do, a question quickly follows in reference to the Singularity, a quasi-religious event with no reliable definition or connection to science that has been prophesied for the coming century and popularized by Ray Kurzweil.” Indeed, futurists, like Kurzweil, has been instrumental in creating a mass of hype bait. His always moving, dated predictions of a vaguely defined “Singularity” occurring is largely debunked by most observers. In an article in the IEEE Spectrum, John Rennie states that “On close examination, his clearest and most successful predictions often lack originality or profundity. And most of his predictions come with so many loopholes that they border on the unfalsifiable.” Cognitive scientist Douglas Hofstadter is no less complimentary, referring in an interview with American Scientist to Kurzweil and Hans Moravec’s books as “an intimate mixture of rubbish and good ideas, and it’s very hard to disentangle the two because these are smart people; they’re not stupid.”
When it comes to journalists, Lipton further states that “too few journalists possess the technical strength to relate developments in machine learning to the public faithfully. As a result, there are not enough voices engaging the public in the non-sensational way that seems necessary now.” More importantly, journalists write for papers that sell advertising. They get measured and paid based on the number of clicks their stories get online. Sensationalism sells.
One of the more common sensationalist sins they engage in is misrepresenting real research advances in “narrow AI” as proof and support that AGI is around the corner. Unfortunately, even the PR departments of respected research groups like Google Brain and OpenAI are fond of doing precisely this. Going back to GPT-3, all a language model like that does is to predict the next word in a given phrase. If you train it on enough data and make a complex enough model, it gets very good at that task. To the point where it can generate entire articles. Does it have any consciousness about what it just “wrote”? Not at all. That’s the logical leap that PR departments and journalists make all the time, just like IBM did in 1997, for example. Deep Blue beats Gary Kasparov at chess; therefore, it has equal cognitive abilities to humans in all domains. It’s false every time. Deep Blue wasn’t even aware that it’s playing a game.
The solution here is simple – as we learned with politics, don’t get your news on Facebook. Find and read legitimate science from folks like Rodney Brooks or journalism from publications like the MIT Technology Review and IEEE Spectrum.
2. The Planning Fallacy:
Our ability to conceive of some technology or scientific hypothesis often far exceeds any real ability to build it or prove it. Einstein predicted the existence of gravitational waves in 1916, and it took 99 years before we were able to observe them. Arthur C Clarke predicted GPS in 1956, and it was 1978 before the first satellite launched. Nuclear fusion has been “30 years away” for decades. We still haven’t colonized the Moon or Mars despite H.G. Wells writing about a Martian invasion in 1897. Some concepts like O’Neill habitats and Dyson spheres might be tens of thousands of years away before we have the technology to realize them.
Some of the most prominent minds in machine intelligence have also substantially overestimated our ability to make significant progress towards human-level intelligence. In 1950, Alan Turing predicted in his famous paper introducing the eponymous test for AI that “in the year 2000 a computer will have a 70% chance of fooling a person”. Claude Shannon was a similar giant in the world of computing and a particular hero of mine. Even Shannon stated in 1961: “I confidently expect that within a matter of ten or fifteen years, something will emerge from the laboratory which is not too far from the robot of science fiction fame.”
Dan Lovallo and Daniel Kahneman described this propensity to overestimate as the Planning Fallacy. They defined it as the tendency to underestimate the time, costs, and risks of future actions and at the same time overestimate the benefits of the same actions. According to this definition, the planning fallacy results in not only time overruns but also cost overruns and benefit shortfalls. As with Shannon and Turing, it happens to even the smartest people of our time.
Realize that it’s human to err, particularly when it comes to complex technologies, and that humanity has engaged in those errors for at least a century.
3. Industry Incentives:
The day after Watson thoroughly defeated two human champions at Jeopardy in February 2011, IBM announced a new career path for its AI quiz-show winner: It would become an AI doctor. IBM would take the breakthrough technology it showed off on television—mainly, the ability to understand natural language—and apply it to medicine. Watson’s first commercial offerings for health care would be available in 18 to 24 months, the company promised. The reality turned out to be very different, and IBM has suffered a fair amount of reputational damage for Watson’s failures. This behavior is, of course, by no means unique to IBM. In the first six months of 2020, 638 AI-related companies in the US had raised $13 billion in venture capital, according to the NCVA. Despite COVID19 slowdowns, the sector seemed on track to surpass 2019 levels, where 1,356 companies raised $18.4 billion in venture capital. Whether it’s in raising venture capital, sales presentations, or pitching consulting contracts, the frenzy around AI remains. The industry incentives to overpromise the benefits of AI is strong. It moves stock prices and valuations – until it doesn’t.
When you see a forecast about AGI, Quantum computing, and Nuclear Fusion, as Cicero did, ask yourself, “Cui Bono?“. “Who benefits?” More often, it’s vested interests bent on talking up the market.
There is a multitude of impressive companies and startups around the world building real value through machine learning. There’s also an entire ecosystem around using AI for Social Good applying machine learning to save endangered species, rainforests and prevent human trafficking. And lastly, there are researchers and companies legitimately placing long-term bets on advanced concepts like embodied cognition. Investors will make more money, and the world will be better off if capital, talent, and attention follow the science rather than hype.