The Automation of Automation
I have a new article up at Discourse making the case for “The Spectacular Promise of Artificial Intelligence.”
This topic has caught a lot of attention recently with the release for public testing of a new generation of AI chatbots, especially OpenAI’s ChatGPT, which is able to create plausible (though often hilariously incorrect) English sentences on a wide variety of topics.
There is something of a bias in the coverage of AI, because the people covering it are writers, and we naturally think that what we do is special and irreplaceable. So create an AI program that writes computer code, and we don’t care. Write one that designs molecules, and it is a mildly interesting curiosity. Write one that makes fake art, and you start to get our attention. But create an AI that writes sentences, and now it’s getting real.
My view is that this is the opposite of the order in which this new technology is going to be productive.
But the more important uses of AI are those that don’t get the big headlines. Consider Copilot, software developed in part by OpenAI, the same organization behind ChatGPT, but for writing software code. Given a prompt to write code to achieve a particular function, Copilot will churn out a workable result….
According to one estimate, with this AI assistance, a programmer can complete a task in half the time it would otherwise take. Then there is a recently announced $6 billion deal in which a major pharmaceutical company bought the rights to a molecule (the basis for a drug to treat autoimmune disorders) that was discovered using AI. If we’re looking for AI applications that are about “atoms not bits,” then here it is, and it is just the beginning.
In terms of how this new technology will actually be used, the key concept is the “work sandwich”: “human thought and effort at the beginning and the end, with automatically generated AI results as the sandwich filling in the middle.”
The best summary of the impact of Copilot is one that I heard after I wrote my article: In 2023, the world's most popular programming language will be English. In other words, you just tell the computer in plain English what you want a program to do, and it creates the code. I'm sure this will have its limits, but it has already been tested and massively increases the productivity of programmers by taking routine work out of their hands. It will eventually trickle down to a product usable by non-programmers to create our own code. This is the long-standing “Star Trek computer” ideal—replicating the kind of performance we have long fantasized about in our fictional future computers.
The promise here is what I call the “automation of automation.”
Automation of production is nothing new. It has been central to the economy since the Industrial Revolution. What the new generation of AI makes possible is the automation of automation. Up to now, when we want to automate something, that means designing a machine that will perform that task, and only that one task. If we want to make something or do something different, a human has to come in to redesign the machine or reprogram it.
But we’re on the cusp of technology that will allow us simply to decide what new thing we want to do, and as with Copilot, the AI will figure out how to do it. This is why there’s so much excitement surrounding the new AI technology: It has the potential for a great leap forward in human productivity. If a programmer can write code twice as fast, how much more work will everyone else be able to do with AI?
When it comes to writing, I think the main initial application of this sort of technology is going to be for very routine types of writing based on information that can be easily and accurately scraped off the Internet—stuff like stock prices or game results in sports, maybe even election results. Usually, there is an intern somewhere whose job is to pull down the latest numbers, put them into intelligible sentences, and post them. Now the AI will be doing that, and the intern will just be double-checking its results. (On TV, AI avatars will soon be reading those results to us.)
This might also work for a lot of political clickbait sites that just repackage somebody else's reporting and use it to stoke partisan outrage. I'm counting down the days until Townhall or HuffPost has a hit columnist who is actually just ChatGPT. People have already created random column idea generators to mock some of the hackier writers. It's a small step from there to just replicating their entire output with AI.
The AI art generators might be useful for generating original artwork to accompany columns. (See the AI-generated images of me in the article.) I don't intend to use these, because it's just another excuse not to pay actual artists. The worse examples are people using AI programs like Midjourney or Stable Diffusion to ask for artworks in the style of a living artist. So if you work to develop a distinctive and original style, it will immediately be ripped off by AI.
But like I said, I think this will have a far greater productivity impact in industrial applications. Let me draw your attention again to that story about AI being used to find therapeutic molecules for drug development.
There is a lot of work that still need to be done to bring this into wide and common use and make it a “general purpose technology” that increases productivity across many fields. But the first steps are already being taken.