A little over a year ago now, we echoed in Genbeta the launch of ‘PaLM’, a Google language model “so precise that it even explains the jokes to you”. Google stuck out then that this AI was “three times higher than GPT-3”at a time when ChatGPT with GPT-3.5 had not yet come to light and left the world speechless.
Well, possibly the spotlight today is on the international launch of Bard, ‘Google’s response’ to ChatGPT, the search engine company has also used its Google I/O event to announce the launch of the new generation of its PaLM model. In the words of Zoubin Ghahramani, Vice President of Google DeepMind,
“Over the past decade of AI development, we’ve learned that many things can be done as neural networks expand. […] Building on this work, today we present PaLM 2, our next-generation language model. […] with enhanced multilingualism, reasoning, and coding capabilities.”
According to Google, PaLM 2 is not only faster and more efficient than PaLM 1, but it is also available in four different sizes to facilitate its implementation for a wide range of use cases (different types of applications and platforms).
The smallest of the models —’Gecko’— is so light it can work “on mobile devices And it’s fast enough to deliver great interactive apps on the device, even offline.”
- Multilingualism: The variety of languages used in the texts with which PaLM 2 has been trained exceeds a hundred. This has also allowed him to improve his ability to understand, generate and translate what Google calls ‘nuanced texts’ (a category that includes idioms, poems and riddles).
- Reasoning: The training texts of the new model are not only varied in languages, but also in subject matter, since they have included mathematical and other science content, thanks to which PaLM 2 “demonstrates improved abilities in logic, common-sense reasoning, and mathematics.”
- Coding: Being proficient in creating/analyzing programming code seems to be a given requirement for large language models today. PaLM 2 is capable of doing this when we work with popular programming languages (such as Python and JavaScript), but also with other much more minority languages (such as Prolog, Fortran and Verilog).
‘Sgroogled.com’: when MICROSOFT launched ANTI-GOOGLE ads
It’s not just PaLM 2, it’s everything that makes it work
The novelty lies not only in PaLM 2 as a model itself, but also in the multitude of new projects that already implement it: “Today at I/O we announced 25 new products and features based on PaLM 2and dozens of teams are already using it across Google.” Here are some examples:
- The international expansion of Google Bard beyond the Anglo-Saxon sphere is due specifically to the new multilingual capabilities of PaLM 2, which allow your expected chatbot to respond in a greater number of languages.
- Google Workspaces features which help you write in Gmail and Google Docs, and stay organized in Google Sheets, are based on PaLM 2, just like in Duet AIyour own ‘copilot’ for Google Cloud.
- An API: Since March, Google has been testing the ‘PaLM API’ with a select group of developers. Starting today, developers can sign up to use the PaLM 2 model in their own applications.
- Med-PaLM 2: Google recalls that this derivative of PaLM 2 was the first large linguistic model to reach the ‘expert’ level when answering questions on the exam that allows obtaining a medical license in the United States.
- Sec-PaLM: It is a specialized version of PaLM 2 trained on security use cases, available through Google Cloud, “helps analyze and explain the behavior of potentially malicious scripts.”
Image | Generated by Marcos Merino using AI
In Genbeta | Bard finally arrives in Spain with improvements to try to defeat ChatGPT: a better design and richer responses