55 // A Day on the Playground

As many of you know, GPT-4 was released this week, and since I got access I of course had to play with it a little.

A couple of months ago, this GraphGPT tweet got me thinking about what could be done within Formable working with natural language.

Varun Shenoy @varunshenoy_

Can LLMs extract knowledge graphs from unstructured text? Introducing GraphGPT! Pass in any text (summary of a movie, passage from Wikipedia, etc.) to generate a visualization of entities and their relationships. A quick example:

GraphGPT turns a description into a graph, and lets you update it with further prompts. That could also be useful in Formable. (maybe watch the video to get an idea of what I’m talking about)

Back then I played around with GPT-3 and got some pretty reasonable results, but without putting in a lot of work into tuning my prompts or fine-tuning, the results were often a bit off from what I would have expected. Especially when I passed it a subgraph of existing data it should incorporate into an update.

Now, my method here is just going by feeling, but unless I got lucky I would say that GPT-4 does quite a bit better with the prompts I tried. The last time I did this, it had a hard time connecting new data to what I listed as existing blocks.

Directly turning a description into something Formable’s back end understands works reasonably well. One thing you might notice is that my prompt says “part of the project Formable. Formable is a project.” It didn’t turn the block “Formable” into a project without the last part. So that would still need improvement.

I still have to look into LangChain and LlamaIndex (GPT Index)—tools to connect data to LLMs, which might open even more doors.

I’d love to hear about your experience with those and GPT-4!

–Emil