Google introduces PaLM 2 at I/O

Background

A large language model is a system trained on a very large amount of text so that it can continue a passage, answer a question, translate, or write code. Models of this kind are expensive to build, and the companies that build one generally try to turn it into a platform — something that other products inside the company and outside developers can call — rather than leaving it as a single application. Ahead of its developer conference, Google had been letting a small group of developers try a PaLM API^[1]. That preview set the stage for a wider announcement.

What Google announced

On 10 May 2023, at Google I/O, Google introduced PaLM 2^[1]. It was not presented as a single model but as a family in four sizes, named Gecko, Otter, Bison and Unicorn^[1]. A family of this kind lets a developer choose a smaller model where speed or cost matter most and a larger one where the task is harder, while all the sizes share the same underlying approach.

Google said PaLM 2 improved multilingual, reasoning and coding capabilities^[1], and that it demonstrates improved capabilities in logic, common sense reasoning and mathematics^[1]. These are the abilities that decide whether a language model is useful beyond simple conversation: whether it works in languages other than English, whether it can work through a problem in steps instead of producing a fluent guess, and whether it can write and check code.

Products and platform

Google said PaLM 2 powered more than 25 new products and features, including Bard^[1]. That number is the part of the announcement most visible to ordinary users: a model of this kind usually reaches people through the products built on it rather than on its own.

For developers, Google said customers could use PaLM 2 in Vertex AI with enterprise-grade privacy, security and governance^[1]. The terms matter as much as the model. Organisations that want to build on a language model need to know where their data goes, who can see it, and what commitments the provider makes. Offering a model through a managed platform with those conditions attached is one way a research result becomes something a company can put into production.

Specialised versions

Two variants were named. Med-PaLM 2 was trained by Google's health research teams with medical knowledge^[1], and Sec-PaLM is a specialised version of PaLM 2 trained on security use cases^[1]. The pattern is a general model with narrower descendants: start from the broad model, add material from one field, and tune the result for the questions people in that field actually ask. Medicine and security are both areas where a fluent general answer is not enough and where the vocabulary and the failure modes are specific.

A successor already in training

Google also said its next model, Gemini, was still in training^[1]. PaLM 2 was therefore presented as the model available now, alongside a pointer to a successor that was not yet finished — a reminder that in this period announcements arrive while the next generation is already under construction.

Why it matters

The I/O announcement is a compact picture of how large language models were being released at the time: as families of models at several sizes rather than single systems; through developer platforms that carry enterprise terms; bundled into a large number of consumer and workplace products; and extended into specialised versions for fields such as medicine and security^[1]. PaLM 2's place in that picture is as the model Google put at the centre of its products and its developer offering on 10 May 2023, with Gemini named as what would come next^[1].