This commentary is by Tom Evslin of Stowe, an entrepreneur, author and former Douglas administration official. It is republished from the Fractals of Change blog.

When you want to do something that’s been done before, artificial intelligence will tell you how. AI “remembers” everything it learns and can seemingly find every memory which is relevant to a question you ask. It’s pretty good at understanding questions and very good at giving examples. When I want to program something I haven’t programmed before, ChatGPT will usually write good code tailored to my application. It’s already done all my Google searching before I even ask a question.

Tom Evslin

If you want to do something new, AI is not nearly as helpful. It can’t remember what it never knew. ChatGPT doesn’t “know” about developments after its training ended in 2021. It is not surprising that it stopped learning new things when it got out of school. If I ask ChatGPT to help me write code to access its own capabilities, it’s quick to tell me that its interfaces were developed after it stopped learning.

You can ask ChatGPT and other large language models (LLMs) to consider new information. They will summarize that information for you. They’ll critique or rewrite it if that’s what you want. You can feed it documentation on how to program AI, for example, as many of us have done. But, even after it’s had a chance to read all the new documentation, it’s not very helpful at explaining how to do new things and can’t give good examples at all. Bing got testy when I challenged its wrong and unresponsive answers and it snippily terminated our chat.

Why can’t AI give good examples of new things? With hindsight, it’s obvious; it doesn’t have any old examples to copy. It can’t create examples from rules. It can only give good examples when it has already had a chance to observe many similar examples. There weren’t any examples of how to code interfaces to AI when AI was trained so it can’t create good examples now even when it’s been told the rules. It can learn but it can’t think!

Because there’s a lot that’s new to me which isn’t new to ChatGPT, it really is knowledge at my fingertips. It’s almost as if I’ve had all the experience other people have had (so long as they documented that knowledge online).  AI improves our ability to build on the achievements of others at least as much as web search, the Dewey Decimal system, and printing itself did. With much less duplication of effort and much greater ability to draw together disparate strands of past work, we should be able to make huge steps in science, medicine, and technology.

Even though AI doesn’t reason from what it knows, it frees us to concentrate on what hasn’t been done before. We may also build on what AI knows and teach it to reason as well. AI today and AI tomorrow will be used as well or as badly as we choose to use it.