Google released a new generation of large language model PaLM2, and Bard ushered in a major update

The Google I/O 2023 conference was held today. In addition to releasing a series of hardware products, the highlight of this conference is of course in the field of AI. Let's take a look at what big moves Google has made this time!

A new generation of large language model PaLM2

Google introduced that they trained PaLM 2 with a large number of mathematical and scientific datasets. The extensive dataset of Palm 2 includes scientific papers and web pages that contain mathematical expressions. As such, it demonstrates improvements in logic, commonsense reasoning, and mathematics, and PaLM 2 has a large boost in multilingual processing, reasoning, and coding capabilities compared to the PaLM model. According to Google's official test, some results of PaLM2 (such as mathematics) are better than GPT-4.

PaLM 2 supports more than 100 different languages and more than 20 programming languages, including JavaScript, Python, Prolog, Verilog, Fortran, etc., showing powerful multilingual capabilities.

PaLM 2 not only understands text understanding and generating content, but also supports multimodality, can understand and generate audio and video content, Google showed a model called Med-PaLM that is fine-tuned for the medical field, which can be analyzed based on X-rays Patient injuries, Google says it has reached the level of human doctors in this regard.

The PaLM 2 model can also be "split" into 4 versions of different sizes and deployed on various devices. For example, a relatively small Gecko model can be run on a smartphone, so that the mobile terminal can also have a large language model. It has been used by more than 70 product teams within Google, showing a wide range of application value and practicality. Integration into Gmail, Docs, Sheets and more - Support for 25 new Google products/features! The Google chatbot Bard has used the PaLM 2 model to enhance its answering ability.

Sec-PaLM is fine-tuned for security use cases and can detect malicious scripts! Through Google Cloud, it uses artificial intelligence to help analyze and interpret the behavior of potentially malicious scripts and better detect which scripts are actually threats to people and organizations in unprecedented time.

Bard is greatly upgraded

Bard has several upgrades due to the move to the more powerful PaLM2 large language model:

1. Eliminate the waiting list and open to use in more than 180 countries and regions.
2. Added picture question and answer function to make the dialogue more intuitive. The answer can output picture information, so that the answer can be combined with pictures and texts!
3. Integrating Adobe Firefly, you can quickly generate pictures.
4. Updates to details such as source citations, dark mode, and a code export button.
5. Bard now supports Japanese and Korean, and will soon expand to support 40 languages, including Chinese.
6. Combined with Google Lens, Bard can understand pictures and generate content based on picture information.

Through these upgrades, Bard has become more powerful and comprehensive, and can provide useful functions for more people.

Android 14, the next generation mobile operating system

Magic Compose, which will launch in beta this summer in Message, Google’s default texting app, is a tool based on AI generative models that gives you tons of hints based on the content of your chat when you reply to a text message. It learns how you speak, and you can make your replies appear to have a different tone, or be written in Shakespeare's style.

In addition to introducing the model, it also specifically introduced Google's sense of social responsibility in developing AI technology, including two tools for identifying AI-generated content:

- Embed watermark

- Embed metadata / meta data

In order to identify synthetically generated content and ensure security, Google announced that it will use watermarks and metadata to mark images generated by Ai. Watermarking technology embeds information directly into content and can remain unchanged after editing.

Google plans to incorporate watermarking and other identification techniques into the models from the start.