Google was one of the first companies to turn to generative artificial intelligence, launching its own chatbot called Bard to compete with competitors like ChatGPT and Bing Chat. The company has also started testing the ability to display AI answers alongside search results.
But Google is also doing more important things in the field of artificial intelligence and wants to develop its own machine learning models. You probably heard about the Google LaMDA model a few years ago, but Google is now using the more advanced PalM 2 model in its AI products.
Here we’re going to take a closer look at the PalM 2 and how it differs from Google’s previous models. We will then compare it with GPT-4 as the best language model available.
What is PalM 2 and how does it work?
PalM stands for Pathways Language Model, which means using Google’s Pathways AI to help train machine learning models. This model is actually part of Google’s big plan to achieve general artificial intelligence or AGI.
Google has said that the Pathways model can provide multifaceted AI beyond text. According to this company:
Pathways can simultaneously enable multimodal solutions that include visual, auditory, and language comprehension. Therefore, the platform can simultaneously process the word “leopard” or the voice of someone saying “leopard” or a video showing a “leopard” running and arrive at the concept of “leopard”. Therefore, this model is less prone to errors.
Google first talked about PalM in 2022. At that time, Google had not yet released a product using this model. But now, with the intensification of the competition in the field of artificial intelligence, this company has upgraded the Bard chatbot to the PalM 2 model. Unlike the LaMDA model that replaced it, this model has been trained in more than 100 living languages of the world and even performs much better in areas such as coding, logical reasoning and mathematical calculations.
Google’s LaMDA machine learning model has a total of 137 billion parameters and 1.56 trillion words, while PalM 2 uses a new method called Compute-optimal scaling, which leads to better results with fewer parameters and smaller training data.
Google has not yet promised multimodal capabilities for its AI products like Bard. But the Gemini language model will likely support mixed image, audio, video, and other formats.
PalM 2 against PalM and LaMDA
The PaLM 2 model has been able to add more capabilities to Bard. This chatbot didn’t have many features when it was first introduced in early 2023 with the LaMDA language model, but now with the transition to PalM 2, it has been greatly improved.
By publishing a 92-page report, Google has explained how PalM 2 is improved over its previous models. We have briefly looked at the improvements of this model compared to previous models:
- PalM 2 is trained on a mixture of languages, and its training data includes many non-English texts. This model also performs better in foreign language tests than LaMDA and first generation PalM.
- PalM 2 has a better logical reasoning ability than its previous models and even managed to compete with GPT-4 in some tests.
- The Bard chatbot is now able to write and debug code using the PalM 2 model. Previous Google models did not have such a feature.
- PalM 2 can provide better translations than previous models. It can even be said that it performs better than Google Translate in some specific languages such as Portuguese and Chinese.
- The Pathways platform is made more efficient by taking advantage of the new PaLM 2 language model. This feature has made it more affordable to use it in various Google services such as Gmail.
- Google is able to customize PalM 2 for specific use cases. For example, the Sec-PaLM model is capable of detecting security bugs or the Med-PaLM model is able to answer medical questions.
How does Google’s PalM 2 compare to OpenAI’s GPT LLM?
OpenAI, as the creator of ChatGPT, has been able to provide more features than other language models. But the question is, will Google succeed in stealing the ball from this company by introducing PalM 2? According to Google’s technical report, PalM 2 performed better in various tests. But it should be noted that laboratory tests do not necessarily mean better performance in the real world.
That’s why we asked the Bard and Bing chatbots two questions to find out which one actually performs better. As you know, Microsoft uses OpenAI’s GPT-4 language model in the Bing chatbot.
The result of the first puzzle we asked was like this. As you can see in the screenshots, both PalM 2 and GPT-4 were able to solve the puzzle correctly. PalM 2 even gave us an alternative solution that GPT-4 didn’t.
However, Google’s Bard chatbot refused to solve another puzzle, but Bing Chat provided an accurate solution. Although these two examples cannot test all the capabilities of PalM 2, it shows that Google still has more work to do.
How will Google use PalM 2?
At its I/O 2023 developer conference, Google explained the future of PalM 2 and AI advancements. In addition to the features mentioned, this language model will be able to improve other tasks, from the new Help me Write feature in Google Docs and Gmail to the production of images with artificial intelligence in slides. These improvements will be rolled out as Duet AI for Google Workspace in the coming months.
Google is also working on a version of PalM 2 for use in smartphones. This model is currently under development under the code name Gecko and will one day enable generative AI to be used offline on smartphones.
Also, application developers can now access PalM 2 for their programming tasks via API. This means that other companies may use Google’s big language model in their products.
Frequently Asked Questions
What is Google Palm 2?
Google Palm 2 is the major language model used by the Bard AI chatbot and other Google services such as Gmail.
Is Med-PaLM 2 available for use?
Google has not yet released Med-PaLM 2 publicly. Apparently, a small group of trusted testers will be allowed access to it first.
Source: Androidauthority
RCO NEWS