Google’s annual conference, Google I/O 2023, was the stage for the announcement of exciting new developments in the field of generative artificial intelligence. One of the big protagonists was the release of PaLM 2, the latest version of Google’s big language model.
This new LLM, which builds on its predecessor, is already powering 25 of the company’s services, including the Bard chatbot.
Google I/O 2023: PaLM 2 and its Great Advances in Generative Artificial Intelligence
During the event, Google shared the improvements and features that PaLM 2 brings with it. One of them is his improved ability to handle multiple languages, having been trained in more than 100 languages.
Additionally, Google claims that your LLM can pass advanced language proficiency exams at the “proficiency” level. PaLM 2 has also been trained on a wide variety of scientific articles, magazines, and websites, giving you outstanding research and math skills.
Lastly, the model has been trained on a large codebase, making it proficient in programming languages such as Python, JavaScript, Prolog, Fortran, and Verilog.
Google has implemented PaLM 2 in 25 of its services, including popular productivity tools like Google Docs, Sheets, and Gmail.
In addition, the model is also being used in specialized Google services, such as Med-PaLM 2, which has shown excellent results in medical competencies and is the first big language model to reach the “expert” level in the medical license exam in the United States.
Another specialized example is Sec-PaLM, which uses PaLM 2 and Google Cloud to analyze malicious scripts and detect cyber threats, among other things.
Google highlighted PaLM 2’s ability to be fast and efficient, outperforming previous models in terms of performance. Furthermore, PaLM 2 is available in different sizes, from the smallest, called Gecko, to the largest, Unicorn.
This variety of options makes it easy to implement in a wide range of use cases, including on mobile devices, allowing interactive applications even without an internet connection.
Google highlighted the versatility of PaLM 2 and its ability to adapt and support a wide range of products and services, with the goal of benefiting a greater number of people.
In addition to PaLM 2, Google also anticipated the arrival of Gemini, a new multimodal model that is highly efficient in terms of tool and API integrations.
Although it is currently in the training phase, Gemini will be available in different sizes and capacities, just like PaLM 2, to ensure its implementation in a variety of products, applications, and devices, with the aim of benefiting a broad spectrum of users.
