Google Unveils PaLM, a Highly Efficient AI Model That Rivals GPT-3
Google introduces the Pathways Language Model (PaLM), a 540-billion-parameter AI that outperforms GPT-3 in reasoning, math, and translation tasks despite not being the company's largest model.
Google introduces the Pathways Language Model (PaLM), a 540-billion-parameter AI system designed to answer questions, reason through arithmetic, and explain jokes. While it is not the largest language model in Google's arsenal, the company claims PaLM represents its most advanced AI tool to date. The model outperforms comparable systems like GPT-3 and LaMDA in key areas such as logical inference, multi-step arithmetic, and multilingual translation.
PaLM trains on a diverse dataset that includes books, Wikipedia, web documents, conversations, and GitHub code, with 22% of the data consisting of non-English text. Google highlights that the model achieves the highest training efficiency ever recorded for a language model of its size. This breakthrough allows PaLM to process complex concepts effectively while utilizing computational resources more responsibly than previous generations of large language models.
A standout feature of PaLM is its ability to chain ideas together to reach a logical conclusion, mimicking human thought processes when solving problems. For example, the model breaks down grade-school math questions step-by-step before providing a final answer. This advanced reasoning capability enables PaLM to successfully solve 58% of the questions in a popular 8,500-question grade-school math dataset.