Knowledge, Plugins and Understanding
In this NY Times article from July, the author says ChatGPT 4 isn’t reliable on its own, but it becomes better with plugins such as “Tasty Recipes”. The article concludes:
The main benefit of training machines with enormous data sets is that they can now use language to simulate human reasoning, said Nathan Benaich, a venture capitalist who invests in A.I. companies. The important step for us, he said, is to pair that ability with high-quality information.
For recipes, one shouldn’t really need ChatGPT-generated ones since so many recipes already exist. The issue is they can be hard to search and the web pages have too much filler content. I found that “Tasty Recipes” didn’t provide enough recipes, but I assume it (or other plugins) will improve. There’s still something more interesting about conversing directly with ChatGPT but these plugins are useful when accurate data is needed.
This issue raises an interesting question - if we end up connecting these AI tools to other sources of data, does the AI itself really need have so much knowledge “baked-in”? Even GPT-2 knew English pretty well, even if it didn’t always make sense logically. However on some level, the more one knows the better one can understand the text in front of it. If you asked GPT-2 to process an academic paper, it might “understand” the English words but it wouldn’t have enough context to really explain what’s going on.
The latest AI tools have much more training data which gives them enough “background knowledge” to decently summarize a wide range of papers. There are lots of ways to do this: ChatGPT offers a few different “PDF summary” plugins, the NY times mentioned using Humata.AI, Google is rolling out NotebookLM, and I’ve recently been using Claude to summarize medieval Hebrew texts! They’re not perfect but one can ask for citations within the text and then look it up. This is one of the most practical use cases today for LLMs.