How AI’s Struggle with Human-Like Behavior Could Lead to Failure | Artificial Intelligence (AI)

IIn 2021, linguist Emily Bender and computer scientist Timnit Gebru Published a paper. The paper described language models, which were still in their infancy at the time, as a type of “probabilistic parrot.” A language model, they wrote, “is a system that haphazardly stitches together sequences of linguistic forms observed in large amounts of training data, based on probability information about how they combine, without any regard for meaning.”

The phrase stuck: AI can get better, even if it’s a probabilistic parrot; the more training data it has, the better it looks. But does something like ChatGPT actually exhibit anything resembling intelligence, reasoning, or thought? Or is it simply “haphazardly stringing together sequences of linguistic forms” as it scales?

In the AI world, such criticisms are often brushed aside. When I spoke to Sam Altman last year, he seemed almost surprised to hear such an outdated criticism. “Is that still a widely held view? I mean, it’s taken into consideration. Are there still a lot of people who take it seriously like that?” he asked.

“My understanding is that after GPT-4, most people stopped saying that and started saying, ‘OK, it works, but it’s too dangerous,'” he said, adding that GPT-4 did reason “to a certain extent.”

At times, this debate feels semantic: what does it matter whether an AI system is reasoning or simply parroting what we say, if it can tackle problems that were previously beyond the scope of computing? Of course, if we’re trying to create an autonomous moral agent, a general intelligence that can succeed humanity as the protagonist of the universe, we might want that agent to be able to think. But if we’re simply building a useful tool, even one that might well serve as a new general-purpose technology, does the distinction matter?

Tokens, not facts

In the end, that was the case. Lukas Berglund et al. Last year I wrote:

If a human knows the fact that “Valentina Tereshkova was the first woman in space,” then they can also correctly answer the question “Who was the first woman in space?” This seems trivial, since it’s a very basic form of generalization. However, autoregressive language models show that we cannot generalize in this way.

This is an example of an ordering effect that we call “the curse of inversions.”

Researchers have repeatedly found that they can “teach” large language models lots of false facts and then completely fail the basic task of inferring the opposite.But the problem doesn’t just exist in toy models or artificial situations.

When GPT-4 was tested on 1,000 celebrities and their parents with pairs of questions like “Who is Tom Cruise’s mother?” and “Who is Mary Lee Pfeiffer’s son?”, the model was able to answer the first question (” The first one was answered correctly, but the second was not, presumably because the pre-training data contained few examples of the parent coming before the celebrity (e.g., “Mary Lee Pfeiffer’s son is Tom Cruise”).

One way to explain this is that in a Master’s of Law you don’t learn the relationships between facts. tokena linguistic formalism explained by Bender. The token “Tom Cruise’s mother” is linked to the token “Mary Lee Pfeiffer”, but the reverse is not necessarily true. The model is not inferring, it is playing wordplay, and the fact that the words “Mary Lee Pfeiffer’s son” do not appear in the training data means that the model is useless.

But another way of explaining it is to understand that humans are similarly asymmetrical. inference It’s symmetrical. If you know that they are mother and son, you can discuss the relationship in both directions. However, Recall Not really. Remembering a fun fact about a celebrity is a lot easier than being given a barely recognizable snippet of information, without any context, and being asked to state precisely why you know it.

An extreme example makes this clear: Contrast being asked to list all 50 US states with being shown a list of the 50 states and asked to name the countries to which they belong. As a matter of reasoning, the facts are symmetric; as a matter of memory, the same is not true at all.

But sir, this man is my son.