Google reveals PaLM 2, an advanced AI language model Google has recently announced the launch of PaLM 2 (Pathways Language Model) – an upgraded version of their state-of-the-art AI language model, boasting enhanced capabilities in multilingual understanding, code generation, and reasoning.
In terms of multilingual tasks, PaLM 2 underwent intensive pre-training on a vast corpus of multilingual text encompassing over 100 languages. This extensive training enables the model to comprehend and translate nuanced content, including idioms, poems, and riddles, with remarkable accuracy.
In the coding domain, PaLM 2 was pre-trained on an extensive range of available source code datasets. The model excels in widely-used programming languages like Python and JavaScript, while also showcasing its ability to generate specialized code in languages such as Fortran, Prolog, and Verilog.
Moreover, PaLM 2 underwent pre-training on a more diverse dataset than its predecessor, which included scientific papers, web pages, and mathematical expressions. As a result, PaLM 2 exhibits significant advancements in common sense reasoning, mathematics, and logic.
Developers can now register to utilize the PaLM 2 model directly or access it through Vertex AI.
Google emphasizes that PaLM 2 surpasses previous models in terms of speed and efficiency. It is available in four different sizes, catering to various use cases and allowing for fine-tuning to support entire product categories. PaLM 2 has already been integrated into more than 25 Google products and features, including:
- Bard, a platform for AI-generated content. PaLM 2 powers the recent coding update for Bard, facilitating developers in their coding endeavors.
- Enhanced workspace features for Gmail, Google Docs, and organization in Google Sheets.
- Med-PaLM 2, a large-scale language model designed specifically for healthcare applications.
- Sec-PaLM 2, a specialized version of PaLM 2 tailored for security-related use cases.
- Duet for AI Cloud, an AI-powered collaboration tool.
PaLM, initially introduced in April, is a dense decoder-only Transformer model with 540 billion parameters, trained using the Pathways system.
Nominations are now open for the 2024 Best Places to Work in IT.
