Behind the recent explosion of interest in artificial intelligence (AI) technology, there is a large dollop of magical thinking at play. To some degree, that is understandable. In the teeth of a significant economic downturn, optimism-starved investors, companies and techno-idealists have latched onto the hype around the likes of ChatGPT and Google’s Bard as a rare reason to be cheerful.
There are, however, also equally good reasons to be sceptical.
In an essay published by The New Yorker last month, science fiction writer Ted Chiang outlined a helpful way to think about Chat GPT and other large language models. Far from being something akin to an artificial brain or consciousness, it is more accurate to describe it as a kind of “blurry jpeg of the internet”. Essentially, he writes, the model compresses a huge amount of information available on the web “in the same way that a jpeg retains much of the information of a higher-resolution image”. But herein lies the rub. “If you’re looking for an exact sequence of bits, you won’t find it; all you will ever get is an approximation.” So ChatGPT and other language processors are not “hallucinating” when they occasionally churn out inaccurate, garbled or surreal answers. Rather, like an MP3 file, some of the information is lost in the act of “smooshing” the information.
Naturally, this is not a reason to totally abandon faith and the technology is likely to improve in the future. Still, a lot of investors are struggling to work out the business case for it. Better chat bots? Maybe. Or as Chiang suggests, a compressed version of the web that can be stored on a private server in the theoretical situation where a company was “losing [its] access to the internet forever”? Sure.
But whatever you think, Chiang’s line of analysis is, at the very least, a strong counterbalance to the optimism. As Yisroel Brumer, founder of venture capital fund Red Cell Partners, told The Financial Times this week, recent progress in generative AI tech is “incredibly impressive and well beyond where we estimated we’d be today”. But “life and death decisions make it very difficult to delegate to a system that makes mistakes that are hard to understand and hard to predict”.
