Google's Latest AI Model Can Be Taught How to Solve Problems It can be taught how to solve math-word problems, and even how to translate a different language.
The latest from GoogleThe Pathways language model, dubbed the Pathways Language Model, uses thought chains to learn how to solve problems, the company announced on its website Event Wednesday.
“The chain of thought prompt allows us to describe multi-step problems as a series of intermediate steps,” said Sundar Pichai, Google CEO. “A bit like your teacher giving you a step-by-step example so you understand how to solve a problem.”
Pichai said normally you would give an AI a question and an answer that they could use as a key to any future questions you might ask. In the chain of thought prompt, you’re still giving an AI a question and answer, but you’re also showing the steps you took to get to that answer. For example, when you had to show your work on a math word problem at school.
This technique can also be used to teach the AI different languages, so someone who speaks a lesser-represented language can still access information. In one example, Pichai showed how the AI can learn Bengali from English, just as it learned to solve math word problems.
“As before, we present the model with two sample questions, in Bengali with Bengali and English responses,” Pichai said. “That’s it. Now we can start asking questions in Bengali.”
While this technique increases the likelihood that the AI will get a correct answer, it’s not perfect. Research from Cornell University found that this technique improves an AI’s chances of answering a complex problem by about 58%. Previous models that did not use this technique were expected to give a correct answer about 18% of the time.
“We’re so optimistic about the potential for language models,” Pichai said. “We hope that one day we’ll be able to answer questions on more topics in every language you speak, making knowledge in search even more accessible across all Google searches.”
See more new articles in category: Other