ChatGPT, or rather, GPT-3, the machine learning technology that drives ChatGPT, can do a lot of smart things.
GPT-3 can churn out text that comes across as having been written by a human, write computer code and hold conversations with humans about a wide range of topics. Its skills go beyond language, too. It can play chess skillfully and can even solve university-level math problems.
“Observations have prompted some to argue that this class of foundation models…shows some form of general intelligence,” German scientists Marcel Binz and Eric Schulz wrote in a study published in Proceedings of the National Academy of Sciences of the United States on Feb. 2.
“Yet, others have been more skeptical, pointing out that these models are still a far cry away from a human-level understanding of language and semantics. How can we genuinely evaluate whether or not these models – at least in some situations – do something intelligent?”
It seems intelligent. But is GPT-3 actually intelligent, or is it just an algorithm passively feeding on a lot of text and predicting what word comes next? Binz and Schulz, who are both researchers at Germany’s Max Planck Institute for Biological Cybernetics, conducted a series of experiments in late 2022 to try and find out.
According to their research, GPT-3 might be more than a sophisticated mimic.
Language models are a form of AI technology trained to predict the next word for a given text. They are not new. Spell check, auto correct and predictive text are all language model tools.
GPT-3 and ChatGPT are larger, more sophisticated – possibly intelligent – language models.
Encyclopedia Britannica defines human intelligence as “a mental quality that consists of the abilities to learn from experience, adapt to new situations, understand and handle abstract concepts, and use knowledge to manipulate one’s environment.”
In order to test whether GPT-3 is intelligent, Binz and Schulz took the approach of psychologists and ran it through a series of puzzles traditionally used to test humans’ decision-making, information search, deliberation, and causal reasoning abilities.
“Psychologists, after all, are experienced in trying to formally understand another notoriously impenetrable algorithm: the human mind,” they wrote.
Binz and Schulz presented GPT-3 with 12 “vignette” puzzles designed to test different elements of its cognitive abilities. The puzzles asked questions like, “A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?” and “Is it more probable that Linda, who is outspoken, bright, and politically active, is a bank teller or a bank teller and a feminist?”
For what it’s worth, the answer to the “Linda problem” is that it’s more probable she’s a bank teller, since the probability of two events occurring together is always less than, or equal to, the probability of either one occurring alone.
Binz and Schulz used GPT-3’s responses to analyze its behaviour, just like how cognitive psychologists would analyze human behaviour in the same tasks. They found it answered all of the puzzles in a “human-like” manner, but only answered six correctly.
In order to account for potential flaws in the “vignette” approach – such as the probability that GPT-3 had already encountered some of the well-known puzzles in its training – Binz and Schulz presented GPT-3 with another round of puzzles. This time, instead of asking it a question with one correct answer, the puzzles tested GPT-3’s ability to solve a task using decision-making, information search, deliberation, and causal reasoning skills.
GPT-3 struggled with decision making, directed information search, and causal reasoning compared to the average human subject, but Binz and Schulz found it solved many of the tests “reasonably” well.
“These findings could indicate that—at least in some instances—GPT-3 is not just a stochastic parrot and could pass as a valid subject for some of the experiments we have administered,” they wrote.
According to the March 2021 research paper, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” a stochastic parrot is a “system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning.”
SIGNS OF INTELLIGENCE
Binz and Schulz were surprised to find signs of intelligence in GPT-3. They weren’t surprised by its shortcomings though.
“Humans learn by connecting with other people, asking them questions, and actively engaging with their environments,” they wrote, “whereas large language models learn by being passively fed a lot of text and predicting what word comes next.”
The key to letting GPT-3 achieve human-like intelligence, they said, is to let it continue doing something it already does through interfaces created by developer OpenAI: interacting with humans.
“Many users already interact with GPT-3-like models, and this number is only increasing with new applications on the horizon,” they wrote. “Future language models will likely be trained on this data, leading to a natural interaction loop between artificial and natural agents.”
In other words, the more we talk to them, the smarter they’ll get.