PaLM (Pathways Language Model) is a large language model, i.e. LLM, similar tothe GPT series created by OpenAIorMeta's LLaMA series models. Google first announced the launch of PaLM in April 2022, surpassing 540 billion training parameters. Like other LLMs, PaLM is a flexible system that can perform various text generation and editing tasks. For example, you can train PaLM to be a conversational chatbot like ChatGPT, or you can use it for tasks such as summarizing text or even writing code. (This is similar to the features Google also announced today for its Workspace applications, such as Google Docs and Gmail.)
At the 2023 Google I/O conference, Google CEO Pichai announced the launch of PaLM 2, a large model benchmarked against GPT-4, and officially released a preview version, improving capabilities in mathematics, code, reasoning, multilingual translation, and natural language generation.
Google will offer PaLM 2 in four different sizes, from smallest to largest: Gecko, Otter, Bison, and Unicorn. Gecko is very lightweight, can work on mobile devices, and is fast enough to run excellent interactive applications on-device even when offline. This versatility means PaLM 2 can be fine-tuned to support an entire category of products in more ways, thereby helping more people.
Features of PaLM 2
PaLM 2 is Google's next-generation large language model, with improved multilingual, reasoning, and coding capabilities.
- Multilinguality: PaLM 2 was trained on more multilingual text, covering more than 100 languages. This significantly improves its ability to understand, generate, and translate nuanced text in multiple languages (including idioms, poetry, and riddles), which is a very difficult problem to solve. PaLM 2 also passed advanced language proficiency exams at the "mastery" level.
- Reasoning: PaLM 2's broad dataset includes scientific papers and web pages containing mathematical expressions. As a result, it demonstrates improved capabilities in logic, common-sense reasoning, and mathematics.
- Programming: PaLM 2 was pretrained on a large dataset of publicly available source code. This means it excels at popular programming languages such as Python and JavaScript, but can also generate specialized code in languages such as Prolog, Fortran, and Verilog.
