The technology world is in constant flux, and Google is no stranger to innovation. Let’s take a close look at one of their latest developments: the PaLM 2, Google’s new large-scale language model deployed in Gmail, Docs and other popular Google services.
The PaLM 2, whose name stands for Pathways Language Model, signals the use of Google’s Pathways AI architecture. This supports training machine learning models for a variety of tasks and is a key part of Google’s vision to achieve general artificial intelligence (AGI). The goal is multimodal AI that goes beyond text and integrates visual, auditory, and linguistic understandings simultaneously.
Looking more closely at the technical aspects of PaLM 2, it is clear that Google has used a new technique called “compute-optimal scaling.” This allows for better results with fewer parameters and a smaller training data set compared to previous models such as LaMDA, which had 137 billion parameters and a training size of 1.56 trillion words.
PaLM 2’s improvements have given Google’s Bard chatbot a significant performance boost. PaLM 2 was trained on a mix of languages and its training data included many non-English text samples. It also shows better logical reasoning capabilities than its predecessors and can even keep up with GPT-4 in tests. In addition, the Bard chatbot can now generate and debug code, which was not possible with Google’s previous language models.
How does the PaLM 2 compare to OpenAI’s GPT LLMs? According to Google’s technical report, it scores better on common sense tests. But synthetic benchmarks don’t necessarily translate to the real world. In an informal “puzzle challenge” with Google Bard and Bing Chat based on OpenAI’s GPT-4, both models were able to solve the puzzle correctly. However, Google’s Bard failed on a different puzzle, while Bing Chat provided detailed reasoning and the correct solution.
At Google’s I/O developer conference, we got a glimpse of the company’s future plans for PaLM 2 and its overall AI development. The new language model will power everything from a new “help me write” feature in Google Docs and Gmail to AI image generation in Slides. In addition, software developers can now access PaLM 2 programmatically, which means we may see other companies relying on Google’s big language model in much the same way Duolingo and Khan Academy have adopted GPT-4 for their AI-driven features.
Another notable feature is a special version of PaLM 2 developed for smartphones. Codenamed “Gecko,” this compact language model could one day allow generative AI to be used completely offline on a typical Android smartphone.
Finally, some frequently asked questions: what exactly is Google PaLM 2? It’s the big language model that powers Bard, Google’s AI chatbot, and other features like “Help me Write” in Gmail. Is Med-PaLM 2 available? Google has not yet released Med-PaLM 2. Initially, the company will grant access to a small group of trusted testers, and it’s unclear if it will ever be publicly available.
Overall, PaLM 2 represents an exciting innovation in artificial intelligence that not only improves Google services, but also opens up new opportunities for the tech industry as a whole. It is another step towards Google’s vision of achieving general artificial intelligence.
