Is AI funding overheated? Beware the hype about artificial intelligence. AI| Artificial intelligence
In 1964, an American computer scientist named John McCarthy set up a research centre at California’s Stanford University to explore an exciting new discipline: artificial intelligence.
McCarthy had helped coin the term several years earlier, and interest in the field was growing fast. By then, the first computer programs that could beat humans at chess had been developed, and thanks to plentiful government grants at the height of the Cold War, AI researchers were making rapid progress in other areas such as algebra and language translation.
When he set up his laboratory, McCarthy told the paymasters who had funded it that a fully intelligent machine could be built within a decade. Things did not pan out. Nine years after McCarthy’s promises, and after millions more had been ploughed into research around the world, the UK government asked the British mathematician Sir James Lighthill to assess whether it was all worth it.
Lighthill’s conclusion, published in 1973, was damning. “In no part of the field have the discoveries made so far produced the major impact that was then promised,” his report said. “Most workers in AI research and in related fields confess to a pronounced feeling of disappointment.” Academics criticised Lighthill for his scepticism, but the report triggered a collapse in government funding, in the UK and elsewhere. It was seen as the catalyst for what became known as the first “AI winter”, a period of disillusionment and funding shortages in the field.
More than 50 years after McCarthy’s bold predictions, technologists are once again drenched with optimism about artificial intelligence. Venture capital funding for AI companies doubled in 2017 to $US12 billion, almost a 10th of the total investment, according to KPMG. In Europe alone, more than 1000 companies have attracted venture funding since 2012, 10 times more than fields such as blockchain or virtual reality, according to the tech investor Atomico.
No end to the big claims
Giants such as Google and Microsoft are building their companies around AI. Earlier this year, Google chief executive Sundar Pichai called the technology “one of the most important things that humanity is working on”, adding: “It’s more profound than, I don’t know, electricity or fire.”
The rest of the corporate world is getting in on the act too. An analysis of investor calls by US public companies last year found that the term “artificial intelligence” was mentioned 791 times in the third quarter of 2017, up from almost nothing a few years earlier.
Significant breakthroughs are promised. Driverless cars are often predicted within a decade. Rising global tensions are boosting government investment, particularly in China. Elsewhere, economists fret about widespread unemployment. Others, such as the late Stephen Hawking, have feared that the rise of robot weapons could eradicate humanity.
But another kind of pessimism is also gaining traction. What if instead of being radically unprepared for the rise of the robots, we have drastically overestimated the disruption caused by the recent excitement? What if, instead of being on the cusp of one of the greatest breakthroughs in history, we are in a similar position to that of the Seventies, at the moment before the bubble bursts?
“The whole idea of making machines intelligent has been a long goal of computer scientists and, as long as we’ve been following it, AI has gone through these waves,” says Ronald Schmelzer of Cognilytica, an analyst firm focused on artificial intelligence. “A lot of the claims [from the Sixties and Seventies] sound very familiar today. It seems to be one of those recurring patterns.”
Indeed, many of the recent breakthroughs in AI have been along the same lines as the chess and language breakthroughs of the Fifties and Sixties, if far more advanced versions. Two years ago, Google’s AI subsidiary DeepMind beat the world champion at Go, an ancient Chinese board game that is many times more complicated than chess. In March, researchers at Microsoft said they had created the first machine that could beat humans when it came to translating Chinese to English.
Two major trends powering the hyperbole
The current excitement about AI owes largely to two trends: the leap in number-crunching power that has been enabled by faster and more advanced processors and remote cloud computing systems, and an explosion in the amount of data available, from the billions of smartphone photos taken every day to the digitisation of records.
This combination, as well as the unprecedented budgets at the disposal of Silicon Valley’s giants, has led to what researchers have long seen as the holy grail for AI: machines that learn. While the idea of computer programs that can absorb information and use it to carry out a task, instead of having to be programmed, goes back decades, the technology has only recently caught up. But while the technology has proven adept at certain tasks, from superhuman prowess at video games to reliable voice recognition, some experts are becoming sceptical about machine learning’s wider potential.
“AI is a classic example of the technology hype curve,” says Rob Kniaz, a partner at the investment firm Hoxton Ventures. “Three or four years ago people said it was going to solve every problem. The hype has gone down but it’s still way overblown. In most applications it’s not going to put people out of work.”
Schmelzer says that funding for AI companies is “a little bit overheated”. “I can’t see it lasting,” he adds. “The sheer quantity of money is gigantic and in some ways ridiculous.”
Most AI sceptics point out that the breakthroughs that have been achieved so far are in relatively narrow fields, with clearly defined structures and rules, such as games. The rapid advancement in these areas has led to predictions that computers are ready to surpass humans at all sorts of tasks, from driving to medical diagnosis.
Fatal accident proof claims detached from reality
But transposing prowess in games to the real world is another task altogether, something that became clear with fatal consequences this year. In March, a self-driving car being tested by Uber in Arizona failed to stop in front of Elaine Herzberg when the 49-year-old stepped out into the street. She became the first person to be killed by a driverless vehicle, which was travelling at 38mp/h (61km/h). The car’s systems had spotted Herzberg six seconds before the crash, but had failed to take action. The incident was the most striking example yet that the grand promises made about AI just a few years ago were detached from reality. While driverless cars were once predicted to be widely available by 2020, many experts now believe them to be decades away.
Driverless cars have not been the only setback. AI’s potential to revolutionise healthcare has been widely touted, and Theresa May said this year that AI would be a “new weapon” in fighting cancer.
The reality, so far at least, has been less promising. IBM’s Watson technology, an AI system that has promised major breakthroughs in diagnosing cancer, has been accused of repeatedly misdiagnosing conditions. Shortly after the Uber crash, the AI researcher Filip Piekniewski wrote that a new AI winter is “well on its way”, arguing that breakthroughs in machine learning had slowed down.
Schmelzer says that companies have stopped placing blind faith in AI, pointing out comparisons with the dotcom bubble when businesses demanded an internet presence even when it was unnecessary. “It was technology for technology’s sake and there was a lot of wasted money. I think we started to see that [with AI].”
Kniaz, of Hoxton Ventures, agrees that the bubble has started to deflate, saying that while companies would often attract funding merely for mentioning artificial intelligence in investor presentations, they are now having to prove that it works.
However, he says that even the narrow progress made in recent years has plenty of real-world uses, even if it is a long way from matching human intelligence. “We’re now at the point where it’s a little more sane,” Kniaz says. “It’s reaching a nice stable point now. You’re seeing it applied to better problems.”