We started working on what turned out to be What to Do When Machines Do Everything in 2015. Debates were beginning to heat up at the time amongst those who thought the rise of artificial intelligence presaged an era in which humans need no longer apply and those that saw such fears as overblown. If you’ve read our book, you’ll know we came down somewhere in the middle – that automation would undeniably replace some human labor but that machines would also create new work for people that would more than backfill what was lost.
To get to that point, I and my fellow authors spent a lot of time (and I mean a lot of time) arguing out both sides of the issue. All three of us could make the case for a period of adjustment in which societies would go through difficult transformation but ultimately work things though ok, and for a jobs apocalypse that would lead to war and bloodshed.
The consensus we reached, which resonated with readers and audiences all around the world, was hard fought but was all the better for it. In the five years since those working sessions, we’ve not once been blind-sided by a perspective that we hadn’t considered during the development of our manuscript.
Fast forward to the summer of 2020 though and the specter of human job displacement is once again emerging center stage. After grappling with the implications of AlphaGo and coming to the conclusion that Demis Hassabis’s baby was astonishing but wasn’t going to hollow out jobs in the processing department of Acme Insurance Inc., Anytown, anytime soon, we (the authors, society, all of us) now have to grapple with another astonishing piece of technology, GPT-3, that raises the same questions that we thought we were done with back when Bruno Mars was in an Uptown Funk.
Generative Pre-trained Transformer 3 is an “autoregressive language model that uses deep learning to produce human like text”, created by the Elon Musk founded OpenAI research laboratory. Like AlphaGo, GPT-3 has been trained on billons of examples - words of text this time, not Go moves - and over time has figured out the underlying rules of language and thus how to use it. This is a pretty good primer on GPT-3 and here are some fun examples of just what it can do.
GPT-3 is generating written text so human like that the Turing Test is an irrelevancy lost in the mists of time.
In What to Do When Machines Do Everything we wrote about software from companies like Automated Insights and Narrative Science which was being used by news agencies and sports new sites to automatically create stories without human hands on the keyboard. These programs could be fed scores and other details from a game and then craft a write up indistinguishable from something Frank Bascombe would turn in https://nyti.ms/31yBeOr.
Those programs seemed amazing at the time. Seen from 2020 though, they appear like nothing more than the first 12 second flight Kittyhawk took in in 1903.
It is not just that GPT-3 seems like a step-function leap forward that draws our attention again. It is that, combined with other recent technology advances, the concerns that we tamped down before come back even more insistent.
Seriously. Stop reading this and imagine that …
Recently, I spent two hours trying to fix an issue with my broadband provider. Without naming names, and with all due respect to the person on the other end of the call who did their best to help me, the experience was completely sub-optimal/frustrating/painful.
I don’t think the current human bar is going to be that hard to leap …
And when it is, when a photo realistic digital human can talk to me in my dialect (let alone my language), and look at me with the color eyes that I prefer (which I pre-selected when I signed up for the service), and solve my problem before my hat has hit the ground, well, what then?
Millions and millions of people who work in call center roles – ordinary people, not making much, just trying to get by – are going to be surplus to requirements. It will literally be a no-brainer to replace those ordinary people with extraordinary machines; machines that can learn and become smarter through the collective intelligence of the network in the way that Tesla cars are all upgraded at once and Waze collects real time data from all of its users to inform all of its users. That can auto-generate a conversation from the billions of conversations that have come before. Monkeys, typewriters, and Shakespeare be damned …
Last summer we invited Carl Benedikt Frey, co-originator in 2013 of the famous claim that 47% of human work could be done by machines “over the next decade or two”, to our annual thought leadership event. In front of our most important clients Carl and I walked through the two sides of the AI and jobs debate, and in a good humored way he accepted that there was little evidence to suggest his hypothesis was coming true - this at a time when pre-Covid, employment levels in most G-7 countries were at record highs.
It felt to me then, that the latest automation scare had been dealt with. Carl’s book, The Technology Trap, that he talked about with our guests, examines in great detail the history of such scares since Aristotle’s day and how they are overcome through the reality of progress as opposed to the fears of the future. As I made my way back to the US from Barcelona, I felt pretty certain we had, once again, looked that future in the face and realized that everything was going to be, as we’d argued, ok.
Now, to be honest, I’m not so sure. Now, I have an image of King Canute standing on the seashore, shouting at the waves to stop crashing on the beach. Now, I see AI and automation as waves coming relentlessly towards us, one after another, unstoppable, uncontrollable, forever. Canute took his courtiers to the ocean to show them that no one, not even a King, was mightier than God and nature. The latest wave of AI shows us - not Kings, just regular people - that we are puny in the face of the advance of the machines of loveless grace.
Perhaps the error we’ve made in looking at AI is at looking at it in isolation, rather than as a piece of something bigger. In Machines we laid out a logical architecture of a “system of intelligence”, which sketched out a combinatorial vision of different technologies working together (see page 52). In our vision and analysis though the other elements of the system were not yet as advanced as the AI “engine”. With innovations in natural language processing and graphic rendering since 2015, this engine (GPT-3) can be the heart of something more than just an engine. Now, a “Machine” is truly emerging.
When machines can write better than humans – or simply as well as humans and without any of the costs and complications of humans – and can talk better than humans and can look better than humans and can work all of the time (and not be worried about Covid) and can get better and cheaper every day, and when more and more of our life and work is online so that the human advantage of presence and reality has diminished, then we will, again, have to ask, as we do on page one of our book, when machines do everything, what am I going to do?
AI and automation are waves that are increasing in strength in incredible ways. We may be approaching a seventh one soon - one that crushes every one of us that write and talk for a living. That’s quite a lot of us. Including me.
Carl’s “couple of decades” still have 13 years to run. Carl - you may still be right.
In the meantime, my name’s Ben Pring and I wrote this message.