Google announces PaLM 2 - The Next-Gen Language Model

A unique LLM that supports 100+ spoken languages, 20+ programming languages, powers 25+ Google Services.

PaLM 2 is Google’s next generation large language model, now much stronger in logic and reasoning. Google claims that the model is really different than anything that they’ve built so far. PaLM 2 is trained on scientific and mathematical data, with specific focus on math, coding, advanced reasoning capabilities, and performing multilingual tasks like translation. PaLM 2 also got trained on a 100+ spoken languages and over 20+ programming languages.

PaLM 2 is already being used to power 25+ Google services including BARD, Google’s Workspace products, the Palm API and more. You might already be using PaLM 2 and not even know it. PaLM 2 can translate not only spoken word languages but also programming languages. So, this could be very helpful in legacy modernization projects to migrate code from one programming language to another. Say if you wanted to migrate from Python code to R, or C++ to Java, or JavaScript to TypeScript, PaLM 2 can get you most of the work done there.

PaLM 2 is proficient at creating and understanding sophisticated language, such as idioms and riddles, which is crucial because these AI models must comprehend both the literal intent and the figurative meaning of words. Users of Google’s “Workspace” apps (Gmail, Docs, Slides and Sheets) will also be able to use the PaLM 2 (Duet AI) as a co-author of text, spreadsheets and slides.

PaLM 2’s architecture also uses “compute-optimal scaling”, which basically means that, as the size of the dataset increases, the size of the compute also increases proportionally. Which means that even though PaLM 2 is smaller than any previous large language models, it performs better than those models overall. It is also more efficient to deploy / serve and therefore more environment friendly.

PaLM2 can be deployed in four sizes, namely Gecko, Otter, Bison, and Unicorn, from smallest to largest, and can also be fine-tuned for domain-specific data to perform certain tasks for enterprise customers. Google says the lightest version of PaLM 2 – Gecko, is so light weight and can even run on mobile phones, fast enough for interactive applications even when offline. Gecko’s small size is clearly a significant advantage to Google, despite its limited capabilities.

Google has always ensured us with safe and ethical tools all these days. They say PaLM 2 is no different and Google has trained the model to de-escalate aggressive and toxic prompting. It not only attempts to avoid such conversations, but also steers the user in more positive directions.

It is said that PaLM 2 is also being used in advance research directions within Google internally from healthcare to cybersecurity. The models are already capable of doing many different tasks, it will only improve over time. With the emergence of multimodal AI tools, Google is even more prepared to set the stage with many compelling capabilities, not just being able to handle modalities like text and code, but also being able to understand and even generate video, audio and images. The most obvious way to interact with PaLM 2 will be in Google’s own BARD AI chatbot, which is opening up to the general public for the first time and rolling out globally.

With PaLM 2, Google will be hoping to close the “AI gap” between the company and its competitors like Microsoft, which has been aggressively pushing AI language tools into its suite of Office software.

Thanks for Reading. Stay Tuned!