Bard vs ChatGPT, each one is pretrained with different sets and subsets of semantic NERs (named entity recognizer). Yes, they both have large amounts of vectored data, but their pretraining makes all the difference, especially for the specificity that one may require.
Here’s Bard vs GPT-4, given the same question. Look at the differences in their answers.
THE QUESTION (Prompt for both): “You are a data scientist explaining the structure of a GPT to the masses. What is the optimum way of question and answer for pretraining an NLP LLM? Please answer in a 3 line length sentence”
Bard’s Answer:
“Sure, here is the optimum way of question and answer for pretraining an NLP LLM in 3 lines of sentence:
– Use a massive dataset of text and code.
– Ask the model a variety of questions, including factual, open ended, and
challenging ones
– Reward the model for providing accurate and informative answers.”
ChatGPT-4’s Answer:
“GPTs are pretrained on a massive corpus of text data in a self-supervised manner, learning to predict the next word in a sentence. It’s not trained with a traditional question-answer format, but it understands this format due to its exposure to such patterns in the data. The optimal question-answer format for further fine-tuning would mirror real-life conversation, including variability in context length, query complexity, and answer diversity.”
Evidently, GPT-4 is a scientific answer (as I requested it to be in the prompt I gave it), however, Bard gave a laiman’s answer with bullit points, showing it’s smaller depth in trained language material.
GPT-4 is trained on 500 billion parameters
Bard is trained on 137 billion parameters (short of GPT 3.5 turbo with 175 billiob parameters)
It’s obvious.