Google opens up about PaLM 2, its new generative AI LLM

Google kicked off its annual I/O conference today with a core focus on what it’s doing to advance artificial intelligence (AI) across its domain. (Spoiler alert: It’s all about PaLM 2.)

Google I/O has long been Google’s primary developer conference, tackling any number of different topics. But 2023 is different — AI is dominating nearly every aspect of the event. This year, Google’s attempting to stake out a leadership position in the market as rivals at Microsoft and OpenAI bask in the glow of ChatGPT’s runaway success. The foundation of Google’s effort rests on its new PaLM 2 large language model (LLM), which will serve to power at least 25 Google products and services that are being detailed during sessions at I/O, including Bard, Workspace, Cloud, Security and Vertex AI.

The original PaLM (short for Pathways Language Model) launched in April 2022 as the first iteration of Google’s foundation LLM for generative AI. Google claims PaLM 2 dramatically expands the company’s generative AI capabilities in meaningful ways.

“At Google, our mission is to make the world’s information universally accessible and useful. And this is an evergreen mission that’s taken on new meaning with the recent acceleration of AI,” Zoubin Ghahramani, VP of Google DeepMind, said during a roundtable press briefing. “AI is creating the opportunity to understand more about the world and to make our products much more helpful.”

Putting state of the art AI in the ‘palm’ of developers hands with PaLM 2

Ghahramani explained that PaLM 2 is a state-of-the-art language model that is good at math, coding, reasoning, multilingual translation and natural language generation. 

He emphasized that it’s better than Google’s previous LLMs in nearly every way that can be measured. That said, one way that previous models were measured was by the number of parameters. For example, in 2022 when the first iteration of PaLM was launched, Google claimed it had 540 billion parameters for its largest model. In response to a question posed by VentureBeat, Ghahramani declined to provide a specific figure for the parameter size of PaLM 2, only noting that counting parameters is not an ideal way to measure performance or capability.

Ghahramani instead said the model has been trained and built in a way that makes it better. Google trained PaLM 2 on the latest Tensor Processing Unit (TPU) infrastructure, which is Google’s custom silicon for machine learning (ML) training. 

PaLM 2 is also better at AI inference. Ghahramani noted that by bringing together compute, optimal scaling and improved dataset mixtures, as well as improvements to the model architectures, PaLM 2 is more efficient for serving models while performing better overall.

In terms of improved core capabilities for PaLM 2, there are three in particular that Ghahramani called out:

Multilinguality: The new model has been trained on over 100 spoken word languages, which enables PaLM 2 to excel at multilingual tasks. Going a step further, Ghahramani said that it can understand nuanced phrases in different languages including the use of ambiguous or figurative meaning of words rather than the literal meaning.

Reasoning: PaLM 2 provides stronger logic, common sense reasoning, and mathematics than previous models. “We’ve trained on a massive amount of math and science texts, including scientific papers and mathematical expressions,” Ghahramani said.

Coding: PaLM 2 also understands, generates and debugs code and was pretrained on more than 20 programming languages. Alongside popular programming languages like Python and JavaScript, PaLM 2 can also handle older languages like Fortran.

“If you’re looking for help to fix a piece of code, PaLM 2 can not only fix the code, but also provide the documentation you need in any language,” Ghahramani said. “So this helps programmers around the world learn to code better and also to collaborate.”

PaLM 2 is one model powering 25 applications from Google, including Bard

Ghahramani said that PaLM 2 can adapt to a wide range of tasks, and at Google I/O the company has detailed how it supports 25 products that impact just about every aspect of the user experience.

Building off the general-purpose PaLM 2, Google has also developed the Med-PaLM 2, a model for the medical profession. For security use cases, Google has trained Sec-PaLM. Google’s ChatGPT competitor, Bard, will now also benefit from PaLM 2’s power, providing an intuitive prompt-based user interface that anyone can use, regardless of their technical ability. Google’s Workspace suite of productivity applications will also get an intelligence boost, thanks to PaLM 2.

“PaLM 2 excels when you fine-tune it on domain-specific data,” Ghahramani said. “So think of PaLM 2 as a general model that can be fine-tuned to achieve particular tasks.”

Originally appeared on: TheSpuzz