Do deep learning models understand language?

Language modeling (such as GPT or Gopher) has been used to perform hundreds of tasks, including some with very few training examples. So remarkable are these systems, that the philosopher David Chalmers, not generally known for hyperbole, has said that “GPT-3 is instantly one of the most interesting and important AI systems ever produced. … More remarkably, GPT-3 is showing hints of general intelligence.”

But Herbert Roitblat, data scientist and author of the book Algorithms are not Enough, has a different perspective on large language models. In an essay for TechTalks, Roitblat explains the mechanics of language modeling, from tokenization to inference, and discusses their limits with some practical examples.

“With more parameters and more input data, these models get better at predicting the next word in a sequence,” Roitblat writes. “But that improvement does not necessarily indicate a substantial breakthrough in artificial intelligence. These models do just one thing—they model the probability of text tokens conditional on the context provided by other text tokens.”

According to Roitblat, “language models produce examples of language, but it is the people who read those examples who inject (most of) the meaning.”

Read the full essay on TechTalks.

For more:

• Why AI can’t solve unknown problems

• How can we tell if artificial intelligence understands our language?

• AI’s struggle to reach “understanding” and “meaning”

• Why we must rethink AI benchmarks