I read an article in the New York times this week about OpenAI’s GPT-3 — a supercomputer specifically designed to learn, without explicitly being told how, to write proper English prose.
By using a very practical form of A.I. technique called Convolutional Neural Nets (CNNs), and after training on extremely massive quantities of human prose, the program can answer all sorts questions sensibly in what looks like very cogent English.
Which is all well and good. The problem I had was with the sensational way that the article was written.
Steven Johnson, who is himself a very good writer, gave the story a persistently sensational slant. He interviewed one expert after another, and all of them said the same thing: This is not at all an example of intelligence in the human sense.
It is, rather, extremely advanced mimicry. The computer has absolutely no self-awareness or consciousness. It is simply processing data.
But that would not have made for as fun a story. So we are introduced to the tantalizing “possibility” that we are witnessing the emergence of intelligence.
But in fact we are not. CNNs, while very useful, are not in any sense sentient beings. In human terms, they have all the smarts of a doorknob.
Which could have been made crystal clear in the article, for the benefit of non-expert readers. But I guess that wouldn’t have made for as fun a story.
One point of view I’ve come across is a comparison to a dog. If a dog could give a speech as good as the GPT-3 model can, would you call the dog intelligent?
The rhetorical effect is not as great if you switch the animal to a parrot though 🙂
Those sorts of comparisons can be problematic. An airplane can fly very well, but I wouldn’t call it a bird. And a tape recorder can also give a great speech, but I wouldn’t call it intelligent.
On the other hand, I would most definitely call a dog intelligent. Dogs and birds are vastly nearer to us in intelligence than are tape recorders, airplanes, doorknobs or the GPT-3.
Mark Twain: “Never let the truth get in the way of a good story.”
I think our brains are optimized to take in stories, and stories that follow certain familiar narratives have become accepted in journalism today as frameworks: e.g., David vs Goliath. If the facts can be shoehorned into a version of this, it will be more likely be read, even though the casting of the roles of “David” and “Goliath” may not be entirely accurate.
It seems that “humans creating an intelligent being” (a.k.a. “playing God”) must be one of these classic story lines. Classic stories are such because they tend to play on universal anxieties?
What would be the viability of an service, perhaps called something like “Buzzkill News” (a play on the existing “Buzzfeed News”)? The goal would be to cover the current “hot” topics in a way that de-sensationalizes them. This is not the same as making the stories boring. Instead, the goal would be to provide background, nuance and analysis that encourages what Daniel Kahneman referred to as “slow thinking” over “fast” even if (especially if) it takes some of the “steam” out of the topic.
Would this “David” of a news service have a chance of competing successfully against the “Goliaths” of MSM?
I would recommend the podcast (and radio show) “More or Less” from the BBC. It has been brilliant at putting numbers published in stories or quoted by politicians into context.
Sort of like fact checking, but also trying to inform about the wider context.
I like Phil’s observation about the structure of human stories — and the way he neatly turned it into yet another story of David versus Goliath.
I wonder whether talking about the structured way that humans receive stories — the story about the story — is itself one of the classic stories.
In which case Phil is also telling yet another classic story. Oh wait, I think I just told it too. 🙂
One year later: the sensational style seems to be the motto of GPT technology.
Which doesn’t change the fact that GPT is not an example of sentience, but rather of useful mimicry.