I've been doing some reading about the history of AI and I found out some things that were new to me. There's a consistent pattern of over-promising, under-delivering, and ludicrous press hype, all with disastrous consequences. There have been at least two dramatic falls in funding over the last decades, called AI winters.
First AI Winter (1974-1980)
The concept of AI has been around for a while, but the first real demonstrations came with the perceptron experiments in the 1950s. The press ran with wild with speculation and massively over-hyped the technology, famously, the New York Times forecast that conscious, self-replicating robots were just around the corner. The perceptron was a great start, but the technology didn't progress very far or very quickly. Of course, it couldn't live up to the hype.
In the early 1960s, researchers spent a great deal of time and money on machine translation, most notably from Russian into English, with the Cold War obviously providing the money and the motivation. Unfortunately, the methods and the computing power just weren't there and the results were very disappointing, certainly nowhere near the level needed to be useful and nowhere near the level needed for funding to continue.
Despite these setbacks, money still flowed into AI research. Eventually, governments started to take an interest in whether their money was producing results, which, frankly, it wasn't. In 1973, the British Government published the Lighthill report which was a devastating assessment of the whole field, and as a result, the UK Government withdrew almost all funding. In the US, government agencies produced similar analysis with similar effects.
Over the next few years, no breakthroughs came, which seemed to justify official skepticism. Despite the lack of breakthroughs, AI research continued, with membership of AI research organizations increasing.
Second AI Winter (1987-2000)
By the early 1980s, things had changed. Japan had risen as an industrial power and it's powerful industry ministry, MITI, had made waves in the west. When MITI decided to fund "Fifth Generation Computer Systems" that were supposed to deliver AI, western governments got a dose of FOMO. At the same time, LISP was having its moment in the sun, driven by early successes in expert systems. Government funding came back and entrepreneurs founded companies to exploit the new technology. Notably, there were a number of companies producing LISP-specific hardware.
Once again, it was a false dawn. By the late 1980s, general-purpose cheaper and capable workstations had arrived, and LISP was ported to these machines. In turn, this led to the collapse of the LISP-specific machine market and to the collapse of the companies making these machines. Investors took note.
Expert systems generally ran into trouble. Outside a few domains, they weren't that successful and there were no breakthroughs. The idea limped on with a few variants, but only had minor successes.
The mighty MITI suffered a setback when progress on Fifth Generation systems was a lot slower than it had expected or wanted. In 1992, it quietly closed the project.
By the early 1990s, AI had a bad reputation again. It had suffered two hype-driven booms and had failed to deliver twice. Investors were skittish, so investment dried up. Governments spent their research money elsewhere and universities focused on other areas. But there were people still working in the area and working on new ideas. Later on, those ideas would bear fruit spectacularly.
What does this mean?
It's a cliche to say "this time, it's different", but so far it is. Yes, the technology is hyped, but the business benefits are obvious, the skeptical voices are louder, and the hype isn't as foolish as it was before. Comparing the press coverage from the late 1950s to now, you see hype, but it's more grounded in reality and there are fewer flights of fancy (no talk of self-replicating robots).
Gartner have a nice model of adoption called the hype cycle. Here's a typical chart used to explain it, taken from Wikipedia. The chart's pretty self-evident, so I won't explain it.
AI's path is more complex than the simple hype cycle, but you can see the same general pattern. We're in high growth now, so it's likely we're in the "Slope of Enlightenment". Are things likely to slow down as we reach the "Plateau of Productivity"? Not any time soon.
Are we likely to see another AI Winter? Probably not, but if it does happen, my guess is it will be a combination of data center constraints plus government action plus human revolt.
No comments:
Post a Comment