In the middle of the 20th century, large gas stations in the US began to be built based on a new self-service model. Since the early 1900s gas station attendants had always come out to pump gas for each customer. Small gas stations were against it, not being able to afford the changes required to support self-service.

By 1968, most gas stations in the US had switched to self-service but in 23 states bans on self-service were in effect after human-service gas stations (with support from local fire marshals) played up safety hazards around self-service, arguing that untrained drivers would overfill their tanks and start fires.

Eventually, human-service gas stations were no longer as profitable as self-service (even for small stations) so gas station operators lobbied to reverse the bans. By 1992, 80% of gas stations in the US were self-service.

In New Jersey, human-service is still mandated by law. Thousands of gas station attendants are still employed to do work that has been long since automated almost everywhere else.

With the accelerating release of ever-more-impressive AI models such as GPT-3 (and it’s close cousin ChatGPT), Dall-e, StableDiffusion, MidJourney, and Github Copilot there are growing concerns amongst the public about the impact that automation will have on artists, designers, writers, programmers, lawyers, drivers, and workers in nearly every other industry.

These were all generated by wavymulder’s AnalogDiffusion checkpoint of StableDiffusion

In the past, when automation displaced jobs, it generally did so in specific industries making heavy use of manual labor, and primarily at the lower end of the income spectrum. Today, it isn’t only narrow AI systems automating specific tasks. In addition, general AI systems are demonstrating exponential growth in core competencies such as reasoning, writing, fact-checking, humor, and conceptual understanding of images and text.

ChatGPT has a sense of humor

It isn’t just the generating of text and images that is undergoing rapid progress. Some models are now outperforming the median human on tasks like trivia, SAT tests, IQ tests, and standardized math tests. ChatGPT, a model not specifically trained for law, even recently passed a practice bar exam.

In addition, some of the major challenges in machine learning are being rapidly solved. It used to be that trained AI models had difficulty remembering and integrating disparate relevant information, were inflexible, had difficulty doing simple physical tasks, and required thousands of examples to learn patterns that humans only needed a few examples to learn. Today, even in models with fixed weights, prompt engineering, fine-tuning, and new ML methods have shown that these problems can be overcome.

The transformation of our society caused by automation has been happening for a long time, but what’s different now is exactly where we are on the exponential curve. The rate of progress is now significantly faster than a human career, and it can only get faster from here. As can be seen on the following graph, compute for AI models has been doubling every 16 months since 2010, with an accelerating exponential growth.

Log-scale graph (every vertical tick represents 100 times more compute)

This isn’t years away any more. It is here, and what we do now will define our society for the rest of the century.

Our world is changing forever, the question we need to be asking ourselves right now is: what do we want it to change into? In a decade or two, do we all want to work like NJ gas station attendants, earning money from unnecessary legally protected busy-work? Or do we want to do the hard work of changing our society so that displacement of work caused by AI leads to all of us being wealthier, healthier, and happier?

ChatGPT waxes poetic about the risks and rewards of AI

What about the crisis of meaning? Lee Sedol, the world’s best (human) Go player, famously resigned from Go following his loss to DeepMind’s AlphaGo, stating that AI would never be beaten by humans. But, even though chess engines have performed better than world champions for decades, players still love to watch the best humans compete in chess. Just because someone can take a helicopter to the top of Everest, doesn’t diminish in the slightest the achievement of those who climb it.

While modern capitalism has resulted in the massive technological progress that has enabled huge wealth creation and the advent of AI, it may cease to be the right tool for the job, if left unmodified. It isn’t AI that is the real problem here, it is capitalism as we currently know it requiring everyone to either work or suffer, whether their labor is really required to keep society running or not. Our government is not set up for overnight change, nor would we likely want such changes if they were possible.

So what’s the solution, from here? One possible solution would be a steep progressive tax on large companies profiting from AI. Funds from the tax would go to funding minimum basic income for anyone earning less than a certain salary threshold. Many people will still want to work, and earn more. But many who are not able to work or don’t wish to, freed from the terrible burden of work, will finally be able to spend their time as they choose.

It’s that future that we should be imagining when we see the incredible achievements of AI.

ChatGPT imagines a specific progressive tax proposal