Google announces large-scale language model 'PaLM 2', already installed in 25 Google services



Google has announced ' PaLM 2 ', the next-generation version of the large-scale language model (LLM) 'PaLM'. PaLM 2 has already been introduced in various Google services, and it is said that PaLM 2 is already being used in Bard, a chat AI.

Google Japan Blog: Bard supports Japanese

https://japan.googleblog.com/2023/05/bard.html

Google AI PaLM 2 – Google AI
https://ai.google/discover/palm2



Google AI: What to know about the PaLM2 large language model
https://blog.google/technology/ai/google-palm-2-ai-large-language-model/

PaLM API | Generative AI for Developers
https://developers.generativeai.google/products/palm

MakerSuite | PaLM API | Generative AI for Developers
https://developers.generativeai.google/products/makersuite

Google announces PaLM 2 AI language model, already powering 25 Google services - The Verge
https://www.theverge.com/2023/5/10/23718046/google-ai-palm-2-language-model-bard-io

Google launches PaLM 2, its next-gen large language model | TechCrunch
https://techcrunch.com/2023/05/10/google-launches-palm-2-its-next-gen-large-language-model/

At Google I/O 2023 , Google announced PaLM 2. This is a competing LLM such as OpenAI's GPT-4.

PaLM 2 can demonstrate excellent performance in various tasks including inference, multilingual translation, and coding.

·inference
PaLM 2 can decompose complex tasks into simpler subtasks and is better at understanding the nuances of human language than previous LLMs such as PaLM. For example, they are good at understanding riddles and idioms that require an ambiguous, metaphorical meaning rather than a literal meaning.

・Multilingual translation
PaLM 2 is pre-trained on parallel multilingual texts and a much larger multilingual corpus than PaLM. This makes PaLM 2 excellent for multilingual tasks, which Slav Petrov, senior research director at Google, explains as 'a significant improvement over PaLM.' Specifically, PaLM 2 can understand idioms in various languages, and can correctly understand German-specific phrases as well as literal translations.

·coding
PaLM 2 is pre-trained with large amounts of web pages, source code and other datasets, so it not only excels in common programming languages like Python and JavaScript, but also in languages like Prolog, Fortran and Verilog. It also supports special coding of .



Such PaLM 2 has already been introduced in 25 types of Google services including Google search.




In a research paper describing the function of PaLM 2, Google engineers appealed that PaLM 2's language proficiency is 'enough to teach the language'. The reason is that PaLM 2 training includes a lot of non-English texts.

PaLM has a version of ` ` Med-PaLM 2 '' trained with health data, which can answer questions at the same level as questions asked in the American doctor's license exam. There's also Sec-PaLM 2 , which was trained on cybersecurity-related data to help explain the behavior of potentially malicious scripts and detect threats in code. pattern.

Both of these models will be available via Google Cloud.

One of the services in which PaLM 2 has been introduced is Bard, a chat AI. For Bard, the introduction of PaLM 2 has improved coding capabilities and enhanced language support. In addition, Bard supports more than 20 programming languages with the introduction of PaLM 2.




Additionally, both Bard's output and user-side prompts now support images as well as text.




In addition, with the upgrade of Bard, it is possible to import the output contents into Gmail and Google Docs.



In addition, Bard plans to release extensions soon, making it possible to use services other than Google. For example, integration with Firefly , Adobe's generative AI, has already been announced, and Firefly will be available via Bard.




Bard is available in over 180 countries and regions and will soon support 40 languages, including Japanese and Korean.




PaLM 2 is also used to enhance the functionality of Google Workspace apps such as Google Docs, Google Slides, and Google Sheets.

Gmail will introduce a reply function 'Help me write' using AI. Simply put, this is a function that allows AI to create the wording used in emails while communicating with chat AI on Gmail.




For example, the AI functions added to Google Docs are as follows.




With Google Spreadsheet, you can let AI create the table.




With Google Slides, it is also possible to let AI brush up the presentation materials, or to create a script for the speaker based on the created slides.




With Google search, it is possible to flexibly respond to search words in the form of questions such as 'I have a child under the age of 3 and a dog. Which is better to live in, town A or town B?' You can ignore the output or continue interacting with the AI in such a way that the output is displayed on top of traditional search results.




Like other LLMs that require huge amounts of time and resources to develop, PaLM 2 is more like a family of products than a single product, with different versions available for consumer settings and enterprise will be expanded according to the settings of Specifically, it will be deployed in four different sized versions: Gecko, Otter, Bison and Unicorn, which will also be fine-tuned based on domain-specific data to perform specific tasks for enterprise customers. It has been.

The lightest 'Gecko' in PaLM 2 is small enough to operate offline on terminals such as smartphones, yet it processes 20 tokens (equivalent to about 16 to 17 words) per second. possible. Google didn't disclose what hardware it used to test Gecko, only stating 'on the latest smartphones.'

Related Posts:

in Software, Posted by logu_ii