When Taylor Webb played around with GPT-3 in early 2022, he was blown away by what OpenAI's large language model appeared to be able to do. Here was a neural network trained only to predict the next word in a block of text—a jumped-up autocomplete. And yet it gave correct answers to many of the abstract problems that Webb set for it—the kind of thing you'd find in an IQ test. "I was really shocked by its ability to solve these problems," he says. "It completely upended everything I would have predicted."
[...] Last month Webb and his colleagues published an article in Nature, in which they describe GPT-3's ability to pass a variety of tests devised to assess the use of analogy to solve problems (known as analogical reasoning). On some of those tests GPT-3 scored better than a group of undergrads. "Analogy is central to human reasoning," says Webb. "We think of it as being one of the major things that any kind of machine intelligence would need to demonstrate."
What Webb's research highlights is only the latest in a long string of remarkable tricks pulled off by large language models. [...]
And multiple researchers claim to have shown that large language models can pass tests designed to identify certain cognitive abilities in humans, from chain-of-thought reasoning (working through a problem step by step) to theory of mind (guessing what other people are thinking).
These kinds of results are feeding a hype machine predicting that these machines will soon come for white-collar jobs, replacing teachers, doctors, journalists, and lawyers. Geoffrey Hinton has called out GPT-4's apparent ability to string together thoughts as one reason he is now scared of the technology he helped create.
But there's a problem: there is little agreement on what those results really mean. Some people are dazzled by what they see as glimmers of human-like intelligence; others aren't convinced one bit.
"There are several critical issues with current evaluation techniques for large language models," says Natalie Shapira, a computer scientist at Bar-Ilan University in Ramat Gan, Israel. "It creates the illusion that they have greater capabilities than what truly exists."
That's why a growing number of researchers—computer scientists, cognitive scientists, neuroscientists, linguists—want to overhaul the way they are assessed, calling for more rigorous and exhaustive evaluation. Some think that the practice of scoring machines on human tests is wrongheaded, period, and should be ditched.
"People have been giving human intelligence tests—IQ tests and so on—to machines since the very beginning of AI," says Melanie Mitchell, an artificial-intelligence researcher at the Santa Fe Institute in New Mexico. "The issue throughout has been what it means when you test a machine like this. It doesn't mean the same thing that it means for a human."
[...] "There is a long history of developing methods to test the human mind," says Laura Weidinger, a senior research scientist at Google DeepMind. "With large language models producing text that seems so human-like, it is tempting to assume that human psychology tests will be useful for evaluating them. But that's not true: human psychology tests rely on many assumptions that may not hold for large language models."
Webb is aware of the issues he waded into. "I share the sense that these are difficult questions," he says. He notes that despite scoring better than undergrads on certain tests, GPT-3 produced absurd results on others. For example, it failed a version of an analogical reasoning test about physical objects that developmental psychologists sometimes give to kids.
[...] A lot of these tests—questions and answers—are online, says Webb: "Many of them are almost certainly in GPT-3's and GPT-4's training data, so I think we really can't conclude much of anything."
[...] The performance of large language models is brittle. Among people, it is safe to assume that someone who scores well on a test would also do well on a similar test. That's not the case with large language models: a small tweak to a test can drop an A grade to an F.
"In general, AI evaluation has not been done in such a way as to allow us to actually understand what capabilities these models have," says Lucy Cheke, a psychologist at the University of Cambridge, UK. "It's perfectly reasonable to test how well a system does at a particular task, but it's not useful to take that task and make claims about general abilities."
[...] "The assumption that cognitive or academic tests designed for humans serve as accurate measures of LLM capability stems from a tendency to anthropomorphize models and align their evaluation with human standards," says Shapira. "This assumption is misguided."
[...] The trouble is that nobody knows exactly how large language models work. Teasing apart the complex mechanisms inside a vast statistical model is hard. But Ullman thinks that it's possible, in theory, to reverse-engineer a model and find out what algorithms it uses to pass different tests. "I could more easily see myself being convinced if someone developed a technique for figuring out what these things have actually learned," he says.
"I think that the fundamental problem is that we keep focusing on test results rather than how you pass the tests."
(Score: 0) by Anonymous Coward on Wednesday September 06 2023, @07:22PM
In one of Steve Wolfram's videos on the innards of ChatGPT he mentions that it builds sentences going forward and can't answer a simple question of the form:
"How many words will it take you to answer this question?"
Maybe by now the developers have special cased this? An awkward sentence could be output with the number at the end, like:
"I can answer your question with a word count of eleven."
Inside Wolfram's Mathematica, a smaller (earlier) version of ChatGPT is available directly and it allows the user to output any/all intermediate calculations that the model does. He uses it in the video and it makes for a good explanation of the tricks required to get this to work--including some randomness.