Google PaLM: Competitor of ChatGPT is now open access
PaLM, claimed-to-be three times powerful AI than ChatGPT, will soon be accessible to everyone.
What is PaLM?
PaLM is an abbreviation for Pathways Language Model. In a self-explanatory way, this language model is a Transformer model that has been trained with the Pathways system. Let’s go a bit further about what these two terms actually mean.
Transformer model is a neural network architecture that benefits from the entire input at once. Normally, the recurrent (RNN) and convolutional neural networks (CNN) try to parse the input and analyze them sequentially, unlike the Transformer. Therefore, what the transformer model aims is to get the meaning of the language by pushing the limits of the most powerful processors.
Pathways is a relatively new AI model that always aims for more. Known AI models are usually go for a single target to become better. However, that might create an unnecessary amount of readings for the same data if the learning process relies on similar paths. With Pathways model, Google does not just aim for one but many skills by combining existing ones and make all of them better step by step. Let’s see the other important aspect from the official blog article by Jeff Dean.
… the model dynamically learns which parts of the network are good at which tasks — it learns how to route tasks through the most relevant parts of the model. A big benefit to this kind of architecture is that it not only has a larger capacity to learn a variety of tasks, but it’s also faster and much more energy efficient, because we don’t activate the entire network for every task.
Success of PaLM
The reason why PaLM is considered three times stronger than ChatGPT relies on these abovementioned terms. Other than just claims, we can see the results if they are telling the truth.
Below is the benchmark result of famous language models such as BIG-G, GPT and PaLM. As the parameters increase, PaLM shows a significantly better result than others and even surpasses the average human rater. Reminding the fact that PaLM is trained with 540 billion parameters, this result might be expected.
Generating a document in seconds
Since it is not public yet, we have only one demo provided by the CEO of the Google Cloud, Thomas Kurian. Here you can find his article as well.
Google will soon implement PaLM into their products such as Docs, Slides and Sheets. This kind of easiness may attract many users even from the loyal Office fans — as long as Microsoft does not implement GPT-4.
AI that can repair the code
Besides the language processing stuff, we all know and get used to by ChatGPT, PaLM can write or refine the codes as well. Five percent of the training dataset belongs to coding. It looks like if it will perform as it is promised, this AI can make things much easier for developers.
Take a look at the example provided again by primary sources.
Fascinating things happened above. It tidied up the code and resolved a few possible errors such as an integer declaration before j (line 8 on right). It even corrected the if statement so that the user can get the correct output (at line 13 on right). According to Google, these sort of simple repair examples are to create “opportunities for fixing more complex errors that arise during software development”.
Conclusion
What it can fully do is still a bit mystery and we will find it out once Google releases for public usage in all manners. However, even from a few demos and experimental results, we can see that PaLM is ready to destroy the monopoly of ChatGPT and carry the rivalry between Microsoft to the AI universe as well.
Thanks for reading 😊
