AI And The Economy: What Has Changed?

The economy will change markedly due to ChatGPT and the similar large language models. We have seen gradual advances in automated information systems in recent decades, but the latest tools are different, pointing to a huge shift in the economy in the coming years.

Smarter tools have come to us for years. Amazon has told us that people who bought this book also bought that book. That’s not too complicated an idea, but it’s a huge step up from a clerk in a bookstore saying that she loved a particular novel. Netflix has recommended movies to us based on our reviews of other movies. Chatbots on a customer service line have asked us, clumsily, why we’re calling. And Alexa stands ready for our to command to play some Jimmy Buffett.

Natural Language

The latest intelligent computer programs, however, are a big step forward. First, they have a much greater ability to understand us when we speak or write normal English—or most any other common language. In contrast, the old customer service chatbot looked for certain words, such as “order,” “ship date,” or “invoice.” The new approach is called natural language processing. The key is that ChatGPT and its cousins have been trained on the general language we speak and write, not on a narrow group of words of interest to that particular program.

AI Remembering What We Said

The large language models also can remember our recent conversations (with some limits). After asking for key issues about, say, business strategy for a recession, the user can say, “Tell me more about the third point you mentioned.” Older models would start from scratch, not knowing what its last answer was. That enables us to drill down to our specific need. How many times have we used Google to search the web and received many results that ignore a key element of our query? With ChatGPT, we can clarify that a particular concept is really important to us.

Summarize or Elaborate

The large language models can summarize long text blocks or elaborate on short text. Although there are limits, a user can submit, say, a 1000-word essay and ask for a one-paragraph summary. Or the user can submit one paragraph and ask for a more detailed explanation of the points made in that paragraph.

Fine-Tuning AI for Business Applications

The next two differences will be hugely important for business use in the near future. The model can be fine-tuned with specific information aimed at a particular company, industry or process. Suppose a company makes washing machines and offers telephone support for its customers. The customer with a problem probably does not know the name of the different parts of the washer, but the large language model can figure out what the customer means. Then the model takes advantage of a history of past customer service calls to tell the customer what to do.

The large language models get their initial training on a huge amount of text gathered from the internet and from digitized books, which is really expensive. But the fine-tuning to specific subjects requires much less data and costs much less in computing time than the initial training.

Connecting to Applications

Combine that with another difference: the ability to connect the new large language models to another computer program through an application programming interface (API). That washing machine company can create a specialized chatbot to help its customers. It can also connect the AI to its customer database to know which model a particular user owns, and connect to the inventory system to know if a replacement part is in stock, and to the service database to schedule an appointment.

Or a company creating social media posts could add some additional training in search engine optimization and then link to major social media sites so that the user need not copy and paste messages across platforms. We will soon have AI-powered virtual assistants with APIs that connect to our email, our calendar and our telephone, fine-tuned on our speaking and writing style, that quickly handle the mundane tasks that occupy much of our workday.

Rapid Improvement

The last notable difference is the recent pace of improvement, which has been huge. AI has been discussed for decades, with mostly slow progress. Since its development in 2018, ChatGPT has made great strides. This is not the same plodding, gradual rate of improvement we’ve seen. Going forward, though, the growth rate is uncertain. We may have picked the lowest hanging fruit, with future gains smaller. But AI itself may help to improve AI, leading to exponential advances in capability. This uncertainty must be considered by business leaders. The landscape may change incrementally as current technology is applied to a wide range of problems, or it may change cataclysmically due to massive gains in AI capabilities in the next few years.

AI Challenges

All of the AI applications will confront challenges, such as false assertions by the AI, data security, the willingness of people to adapt to different ways of working, as well as legal and regulatory changes. But some people will take risks, and many of the problems will be overcome.

As this article is written (June 2023), writers are already using ChatGPT, either by editing existing drafts or to create entire articles. Computer programmers get help writing code. Other programmers are writing “front ends” to help people use AI for specialized tasks. However, much greater benefits will accrue to businesses as people fine-tune large language models to particular needs and add APIs to make usage much easier for specific tasks.

Incidentally, every word in this article was written by me, but when ChatGPT was asked to evaluate accuracy, it suggested a few clarifications and elaborations, half of which I incorporated—in my own words—into this article. Then an AI expert reviewed the draft and offered two additional suggestions. The final result benefitted from both pieces of advice.

This is just the beginning. Connecting large language models to the rest of our business activities will pay off greatly. The models themselves will continue to improve, but there’s plenty of benefit coming from the customized uses in which we put the models to work.