The world has been marveling at the capabilities of ChatGPT, but even the program’s creators were once left amazed at its responses.
OpenAI President Greg Brockman has revealed how he was blown away when ChatGPT was able to add together two 40-digit numbers. “Let us know something that blew your mind, something that you did not see coming,” a TED interviewer recently asked Brockman.
“You can try this in ChatGPT,” Brockman replied. “If you add 40-digit numbers, the model will do it. This means it’s really learned an internal circuit for how to do it. The really interesting thing is that if you have it add a 40-digit number plus a 35-digit number, it’ll often get it wrong. And you can see it’s really learning the process, but it hasn’t fully generalized, right?” he said.
Brockman hinted that they hadn’t expected ChatGPT to be able to multiply two 40-digit numbers. “You can’t memorize the 40-digit addition table; that’s more atoms than there are in the universe. So it had to have learned something general,” he said.
Brockman also revealed how OpenAI first realized the potential of Large Language Models, which in theory simply look to predict the next word in a sequence. “One person (within OpenAI) was working to predict the next character in Amazon Reviews. This is a syntactic process, you expect, the model will predict where the commas go, where the nouns and verbs are. But he got a result — he actually got a state-of-the-art sentiment analysis classifier out of it,” Brockman said.
Brockman said that this result made the company double down on such models. “This model could tell you if a review was positive or negative. This was the first time you saw this emergence, this sort of semantics that emerged from this underlying syntactic process. And there we knew, we got to scale this thing, got to see where it goes,” he said.
Interestingly, Greg Brockman says that now OpenAI is able to predict the advances that their programs will make. “We’re able to predict, for example, the performance on coding problems. There are these incredibly smooth scaling curves — we can look at a model that’s 1000 times smaller, or 10,000 smaller, and predict what the new model is going to do,” he said.
Multiplying 40-digit numbers isn’t the only emergent property that ChatGPT and GPT-4 have displayed. Users playing around with the product have discovered that they can recognize jumbled up words, solve mazes, create entire boardgames, solve Captchas, and even answer McKinsey case studies. Given the pace at which AI is progressing, it’s very likely that these programs will come up with even more emergent properties in coming years, but for now, as OpenAI’s Greg Brockman describes, even researchers are sometimes blown away by what these models are capable of.