What is PaLM 2, the model Google announced at I/O, and what does it power?

Yesterday, 10 May 2023, at its annual developer conference, Google I/O, Google introduced PaLM 2, the successor to its Pathways Language Model. It is the engine behind a long list of announcements made on stage, from a more capable Bard to writing help in Gmail and Google Docs.

The headline is not just a bigger model. Google says its research shows that "bigger is better" is too simple a rule, and that how a model is designed and trained matters as much as its size. PaLM 2 is pitched as more capable and also faster and more efficient than the models before it.

What happened, and who is behind it

The announcement was made at I/O and explained in a post by Zoubin Ghahramani, a vice president at Google DeepMind. The original PaLM, presented in April 2022, had 540 billion parameters, the adjustable numbers a model learns during training. The announcement gives no parameter count for PaLM 2.

The same day, Google said it is combining its two big research groups, Google Brain and DeepMind, into a single unit called Google DeepMind. It also said it is already training its next model, designed from the start to be multimodal, meaning able to work with images and other kinds of data as well as text.

What is new in PaLM 2

Google highlights three areas of improvement, each tied to what the model was trained on:

• Languages. PaLM 2 was trained on much more multilingual text, covering more than 100 languages. Google says it handles idioms, poems and riddles, and passes advanced language proficiency exams at the "mastery" level.
• Reasoning. Its training data includes scientific papers and web pages full of mathematical expressions, which improves logic, common-sense reasoning and math.
• Coding. It was trained on a large amount of publicly available source code. It is strong in popular languages like Python and JavaScript, and can also write in specialized ones such as Prolog, Fortran and Verilog.

The most practical change is that PaLM 2 comes in four sizes, named after animals:

Google gave few details about Otter, Bison and Unicorn beyond their order. The point of the range is flexibility: a small model costs less and answers faster, while a large one handles harder tasks. Picking a size is like choosing a vehicle: a scooter for quick trips in town, a truck for heavy loads.

What it powers

Google says over 25 new products and features announced at I/O run on PaLM 2. Some examples:

1. Bard uses PaLM 2's language skills to expand to new languages, and it powers Bard's recently added coding help. Google also removed the waitlist and opened Bard to over 180 countries and territories in English, with Japanese and Korean rolling out and plans to reach the top 40 languages.
2. Google Workspace writing help in Gmail and Docs, and organizing help in Sheets.
3. Med-PaLM 2, a version trained by Google's health teams, which Google says reached "expert" level on questions in the style of the US medical licensing exam. A small group of cloud customers will test it this summer.
4. Sec-PaLM, a version trained for cybersecurity, which helps analyze and explain potentially malicious scripts.
5. Developers and businesses can sign up for the PaLM API or use the model through Vertex AI, Google Cloud's platform for building AI applications, and it powers Duet AI for Google Cloud.

Why it matters

PaLM 2 arrives two months after OpenAI's GPT-4, and it is Google's clearest answer so far. What stands out is the strategy: rather than one model behind one chatbot, PaLM 2 is a family of sizes that Google is spreading across its products, from phones to email to the cloud.

The Gecko size is especially interesting for builders. A model that runs on a phone without an internet connection could make AI features faster, cheaper and more private, because your words would not need to leave your device. Google has not yet said when, or in which products, Gecko will appear.

How can you try PaLM 2 today?

The easiest way is Bard. As of yesterday, there is no waitlist: anyone in more than 180 countries and territories can visit bard.google.com and start using it in English, with Japanese and Korean arriving now. Google is clear that Bard is still an experiment, so check important facts with a search.

If you use Google Workspace, watch for new writing features in Gmail and Docs as they reach accounts. Google has not given a single date for everyone.

Developers can sign up for the PaLM API to build with PaLM 2 directly, and companies on Google Cloud can use it through Vertex AI with business-grade privacy and security controls. A good first experiment is to give the same task, for example summarizing a document in two languages, to both PaLM 2 and another model you use, and compare accuracy, speed and cost.