Google Unveils PaLM, a 540-Billion-Parameter Language Model
Google Research introduces PaLM, a massive 540-billion-parameter AI model that outperforms human baselines on key language benchmarks. The system demonstrates unprecedented reasoning and logical inference capabilities.
Google Research unveils the Pathways Language Model (PaLM), a massive 540-billion-parameter AI system that surpasses average human performance on the BIG-bench benchmark. The model outperforms other state-of-the-art systems on the majority of 29 evaluated natural language processing tasks. PaLM achieves these results by utilizing an autoregressive decoder-only Transformer architecture trained on an unprecedented cluster of 6144 TPU chips.
PaLM shows particularly strong capabilities in complex reasoning tasks that challenge older AI models. By pairing the massive model with a new chain of thought prompting method, the system achieves top results on several reasoning benchmarks. Furthermore, PaLM demonstrates unique abilities in two novel areas: logical inference and explaining the meaning behind jokes.
The development of PaLM relies heavily on Google's custom Pathways technology, which allows the model to scale efficiently across thousands of hardware accelerators. This breakthrough represents a major step toward Google's ultimate goal of creating a single AI system that generalizes across millions of tasks and understands various data types with remarkable efficiency.