GPTs will change the nature of work
ChatGPT woke the world up to the importance of artificial intelligence last year. The media has not been so full of talk about AI since DeepMind’s AlphaGo system beat the world’s best Go player in 2016. Launched at the end of November, ChatGPT wasn’t the best AI in the world, as the prominent AI researcher Yann LeCun pointed out. But it was the first time the general public got to play with such a sophisticated and capable model.
Launched on 14 March, GPT-4 is a genuine game changer. It seems likely that over the next few months, Generative Pretrained Transformer (GPT) AI models will transform the nature of work for many of us. They are still (probably) a long way from artificial general intelligence and superintelligence. They are not conscious, and they do not “understand” what they are doing. But GPT-4 is a big improvement in natural language programming. Like ChatGPT, it was trained on vast amounts of data, but unlike ChatGPT, it can also ingest a lot of customer data, which makes a big difference.
It is also capable of rudimentary logic, generating conclusions that are not explicit within its training set. For instance, given a list of homes and their features, it might show you the ones which are suitable for two room-mates, by searching its memory for what they need, and discovering that what room-mates need is two bedrooms, and at least a bathroom and a toilet.
GPT-4 is multi-modal, meaning it can ingest queries in the form of images and numbers as well as text, and it makes fewer mistakes than its predecessors, although it does still have an over-active imagination, and its outputs must absolutely be checked by humans.
A powerful productivity tool
As a result, GPT-4 is an astonishingly powerful productivity tool. It can give you a transcript of what was said in a meeting, and produce a summary of the most important statements. The summary can be as long or as short as you like, and you can determine what counts as important. It can summarise a long email thread and draft an appropriate response. It can analyse data in whatever form you have it, and create Excel charts with whatever statistical graphics you prefer. (But remember: pie charts are the spawn of the devil.) It can turn a Word document – or even a scribbled handwritten note – into a PowerPoint presentation or a fully functioning website.
It can do all this in seconds, and this is the point. Humans can do all these things, but we take much, much longer.
Automating and enhancing analysts
A tool called ChatExplore just launched by a company called Akkio illustrates this well. Founded in May 2019, with backing from Bain Capital Ventures, Akkio uses AI to provide services for analysts in banks, corporates and elsewhere. These analysts are power users of Excel, using formulas, pivot tables, statistical and visualisation tools to interrogate data, and extract and display insights that will ultimately increase revenue or reduce costs for their employers.
Akkio’s products automate much of this work, and also enhance it by making new queries possible, because the AI is accessing information outside the existing data set. Imagine you are an avocado-obsessed millennial, and you have a list of avocado prices in all the cities in the USA. You couldn’t identify which state the fruit is cheapest in without adding the knowledge of which of the cities in the list are in which state. It would take a human many minutes of tedious labour to do this, but using GPT-4, ChatExplore can do it in seconds.
The following images give a worked example, showing how raw data about a retailer’s product lines can be turned into useful charts in moments, and specific questions can be answered just as fast. The analysis is not rocket science, but it requires some thought when humans do it, and quite a bit of time.
Ten times faster, or ten times fewer?
Jonathon Reilly is the co-founder and COO of Akkio. He told me that on average, GPT-4 will probably speed up the work of most analysts by something like a factor of ten. This is likely to be replicated across most types of knowledge work, including copy writing, translating, contract drafting, creative image generation, and so on.
It is impossible to be certain today what the impacts of this will be, except that they will be profound. Will a department of ten analysts be slashed to one, or will there still be ten, but producing insights which create ten times the value? Will analysts be de-skilled, because they no longer need to know how to analyse the data themselves, or is the important consideration that they will acquire the new and more valuable skill of AI whispering? Many of us would now struggle to perform the mental arithmetic which our parents and grandparents learned, and would be stumped if asked to perform calculations with a slide rule. But we can do things with Excel that would amaze them.
Will companies like Akkio retain their role in the value chain, or will the tech giants like Microsoft hoover up these services? Microsoft has launched its GPT-based app, called Copilot with this short video, complete with annoying music, and this longer one. It’s impressive.
There is also the question of who will provide the underlying AI. Big tech has no stranglehold over the image generating systems, and they are competing fiercely over the natural language processing models. It seems unlikely that any of them can build a moat, and generate monopoly profits; instead they will become commodities, with the cost of use falling ever closer to the cost of provision, which is modest.
Whoever provides the services, it seems clear that GPT-4 and its current and future rivals, plus whatever services companies like Akkio build on top of them, will generate a huge leap in productivity in the coming year. It may be that economists will be unable to measure the leap, in the way they have been unable to measure the productivity contribution that computers and the internet have made. But it will be there nonetheless, and it will transform the working lives of millions.
The Great Churn
I have long argued that AI will eventually automate enough human jobs to create an economic singularity, with lasting widespread unemployment. And that we need a much smarter response to this possibility than just assuming that universal basic income (UBI) will instantly become affordable and take care of everyone. But I have also long thought that this future joblessness cannot happen until machines are able to do – unchaperoned – almost everything that we can do for money, and we are still far from that.
What will happen in the meantime is that we humans will have to retrain more and more frequently, and change jobs and industries more and more often. I call this the Great Churn. The automation of jobs like secretaries by Microsoft Office happened relatively slowly because the software took time to mature, and people took time to familiarise themselves with it. The revolution ignited by GPT models looks likely to happen much faster. Not enough people are thinking about this, and as a result we are not ready for it.