Google has introduced PaLM (Pathways Language Model) 2, an update to its next-generation large language model with improved multilingual, coding, and reasoning capabilities.
For multilingual tasks, PaLM 2 was more heavily pre-trained on multilingual text, spanning more than 100 languages, thus improving the software’s ability to understand and translate nuanced text including idioms, poems, and riddles.
In the coding realm, PaLM 2 was pre-trained on a large quantity of available source code data sets. The model “excels” at popular programming languages such as Python and JavaScript, Google said, but is also capable of generating specialized code in languages such as Fortran, Prolog, and Verilog.
PaLM 2 also digested a wider-ranging data set as part of its pre-training than its predecessor including scientific papers, web pages, and mathematical expressions, Google said. As a result PaLM 2 demonstrates improved capabilities in common sense reasoning, mathematics, and logic.
Developers can sign up to use the PaLM 2 model or use the model in Vertex AI.
Google noted that PaLM 2 is also faster and more efficient than previous models, while coming in four sizes for a range of use cases, allowing PaLM 2 to be fine-tuned to support entire classes of products. PaLM 2 already powers more than 25 Google products and features. These include:
- The Bard generative AI platform. PaLM 2 also powers the Bard coding update, helping developers code.
- Workspace features for Gmail and Google Docs, and for organization in Google Sheets.
- Med-Palm 2, a large language model for use in healthcare.
- Sec-PaLM 2, a version of PaLM 2 for security use cases.
- Duet for AI Cloud, an AI-powered collaborator.
PaLM was unveiled in April as a 540-billion-parameter, dense decoder-only Transformer model trained with the Pathways system.