I lean toward the hype being overblown. Pundits are all agog with ChatGPT, language generation tool built using GPT-3. There are examples galore making the social network rounds demonstrating how ChatGPT has rustled up what looks like a sensible response to a question. It seems that the simple questions that stump Alexa every day could be passé, since ChatGPT can handle them all.

In fact, in the WSJ just before Christmas, there was an article titled, “ChatGPT Wrote My AP Essay – and I got a Passing Grade”. The article explains how the author asked ChatGPT to write a 500-word essay on The Great Gatsby. And Lo, the engine did the work and a passing grade was attained. For the press and pundits this was hot news. But is it really?

Deep learning, which extends neural networks to a level hard to imagine even 15 years ago, has access to so many sources of material where the Great Gatsby is mentioned, described, analyst, précised and critiqued. There are books online, papers, and other sources on the Internet. There are patterns in the sources; and so patterns can be grouped. Certain words will appear that help group the sources. This is what deep learning does and it does it far better than we can given the volume of data.

The level of complexity of the patterns are just what layer upon layer of neural networks and nodes in each layer are designed to “learn”. So, from that perspective, ChatGPT looks pretty smart. But it is not learning in the sense of developing something net new. It is compiling, connecting, merging, synthesising.  Compared to my third son, still in high school, perhaps ChatGPT does look smarter. But is that even a fair comparison?

We Have Been Here Before

The pattern discovery and assembly (in response to questions) being discussed here are not that different to the same success we have seen in the past with respect to AI and gaming. First there was chess and the now infamous Gary Kasparov story. Then there was Go! which has eminently more moves to consider. More recently we saw StarCraft in the news. It was this last game that really got me interested.

StarCraft is not like Go! or Chess. In both games all moves are perfectly known – all moves are visible. There remains, of course, many millions of combinatorial moves throughout the game.  

StartCraft too has many alternative variants with different combinations of moves, but not all moves in StarCraft are always visible. Fog of War means that some moves take place outside the visible range of the competitor.

For AI and deep learning to really excel at StartCraft it would have to learn different capabilities to those well demonstrated with Chess and Go! It would need to learn how to:

• Feint moves, to prod and to test the enemies positions and defenses
• Place bets on unknown moves (or probabilities)

While these are quite different to Chess and Go!, it turns out that ML can in fact model these unknown moves – they are just another complex form of nodes on the network. So AI won again. But we have yet to see ML learn to lie or cheat in such a way that is not akin to a bluff under fog of war conditions.

Not the Breakthrough We Need

So, while the press suggests that ChatGPT is smart, I accept that it is smart, up to a point. There are other examples that appear almost as smart as ChatGPT in related fields.

My colleagues who cover AI are replete with similar examples that help with grammar, compiling slides and responses to queries of varying complexity, or taking text and turning them to video.

ChatGPT might be the best of the bunch, but how smart is it? The use cases seem pretty similar: Using queues, compile a response by collecting related information from a previously synthesized, huge collection of text that relate to the queues given.

Clearly what we can do with deep learning is impressive. Clearly the ability for such technologies to help with productivity-induced growth exist. But even today, several years into the most recent hyped-cycle for AI, that future remains elusive.

Did you see this article from the Economist recently: Triumph of the Luddites: Covid-19 was meant to lead to job-killing automation. It seems that over the years (the chart in the article runs from 2005 to 2022) shows that the number of “routine jobs” has declined continuously. As such, the opportunity for automation is declining – who would have thought of that?

As the article suggests, “Rather than workers complaining about shortage of jobs, bosses complain about a shortage of workers.” ChatGPT does look impressive, but it does not seem to herald (yet) the kind of breakthrough pundits are looking for.