Many people fear that AI will one day automate away everyone’s jobs, maybe even someday soon, but I’m hopeful the true power of AI will be solving problems that have proven too difficult for humanity to tackle on our own.

If you’re reading this, you’ve almost surely had the chance to interact with a Large Language Model (LLM) via tools like OpenAI’s ChatGPT by now. The recent progress in natural language models — mostly using a type of deep neural network known as a “Transformer” — has some declaring the imminent arrival of Artificial General Intelligence (AGI).

AGI is the ability of an intelligent agent to understand or learn any intellectual task that a human being can (per Wikipedia). Although the new class of language models are super impressive, I don’t think we’re as close to general intelligence as some people are claiming. An LLM may have read the manuals of every modern coffee maker during training, but it still can’t walk into the kitchen and make a cup of coffee. That is, LLMs will fail the Wozniak test of intelligence, because lots of human intelligence is non-verbal.

Regardless of the true timescale in which we’ll reach AGI, AI is now having a real impact in a lot of different applications. There are concerns that AI will soon be used to automate many jobs, making the people who do those jobs now obsolete. Indeed, an AGI would be able to do any job that a human being can do and it would never need to sleep, take time off, or break for lunch. That would be pretty difficult to compete with. But, I think automating tasks that human beings do well is a poor use of these new technologies; society will get a bigger lift from creating AI to solve problems human beings can’t. That is the domain of narrow superhuman intelligence.

Biology and medicine are filled with problems that humans find challenging, or impossible, to solve. Take, for example, the protein folding problem.

A protein is a string of amino acids, in which the amino acids have different properties (e.g., electric charge) that cause them to interact with each other and the environment in various ways. These interactions cause the protein to spontaneously fold itself into a 3-dimensional structure that’s important for determining its function.

The protein folding problem is conceptually simple: we want to predict the 3-dimensional structure of a protein from its amino acid sequence.

Let’s try an experiment. Without using a computer, see if you can predict the fold of the protein with this amino acid sequence:

AAYKVTLVTPTGNVEFQCPDDVYILDAAEEEGIDLPYSCRAGSCSSCAGKLKTGSLNQDDQSFLDDDQIDEGWVLTCAAYPVSDVTIETHKKEELTA

Of course you can’t do this! The human brain hasn’t yet evolved the ability to predict protein structures.

But AlphaFold, an AI model by DeepMind trained to predict protein structures, can predict the structure of this protein. Here it is:

Now, perhaps you’ll say that this example is a strawman. Of course human beings can’t fold proteins in their heads. Fine. Let’s compare the performance of AlphaFold to the previous generation of human designed algorithms for protein folding that didn’t use AI.

The results speak for themselves. Humans had been attempting to design algorithms to predict protein structures for decades, but none of those algorithms ever came close to achieving the accuracy of the new AI-based methods.

AlphaFold is an example of a narrow superhuman intelligence. It is not a general intelligence. It can’t go into the kitchen and make a cup of coffee. It can’t recognize a picture of a cat, drive a car, or return search results from the internet. There are many things it can’t do. In fact, it only does one thing; predict the structures of proteins from their amino acid sequences. But, it does that one thing with superhuman performance. Not only is it laughable to compare AlphaFold to the performance of a single person, it substantially outperforms all previous human designed algorithms. This is the real promise of AI.

I think I understand why so much of AI research centers on computer vision and natural language tasks that humans do routinely. For one, there’s lots of image and text data. In addition, it’s a lot easier to tell if your algorithm is working well when you’re trying to solve one of these problems because you can just look at the output and judge for yourself. By contrast, figuring out if a prediction from AlphaFold is correct requires someone to spend a huge amount of effort crystalizing the protein and determining its structure with X-ray crystallography. I suppose that’s the cynical view. The more hopeful view is that providing computers with human-like vision and language capabilities will create a new paradigm for human-computer interaction, enabling people and computers to collaborate more effectively. Still, I wish there was more research on the truly hard problems.

In my view, there are only two kinds of problems in this world: problems human beings have solved, and problems we haven’t solved yet. While there’s a lot of AI research on automating things humans do pretty well, there’s much more to be gained by using AI to solve the unsolved problems. I think we’ll soon find that some of problems we thought were so complex that they are impossible, aren’t actually that difficult for a narrow superhuman intelligence.