Why certain GPT use cases can't work (yet)

Some weeks ago I built the "Muslim ChatGPT". From user feedback, I very quickly realised that this is one use case that absolutely won't work with generative AI. Thinking about it some more, I came to a soft conclusion that at the moment there are a set of use cases that are overall not well suited.

Verifiability

There's a class of computational problems with NP complexity. What this means is not important except that these are hard to solve but easy to verify. For example, it's hard to solve a Sudoku puzzle, but easy to check that it's correct.

Similarly, I think that there's a space of GPT use cases where the results can be verified with variable difficulty, and where having correct results is of variable importance. Here's an attempt to illustrate what some of these could be:

The top right here (high difficulty to verify, but important that the results are correct) is a "danger zone", and also where deen.ai lives. I think that as large language models become more reliable, the risks will be mitigated somewhat, but in general not enough, as they can still be confidently wrong.

In the bottom, the use cases are much less risky, because you can easily check them, but the product might still be pretty useless if the answers are consistently wrong. For example, we know that ChatGPT still tends to be pretty bad at maths and things that require multiple steps of thought, but crucially: we can tell.

The top left is kind of a weird area. I can't really think of use cases where the results are difficult to verify, but also you don't really care if they're super correct or not. The closest use case I could think of was just doing some exploratory research about a field you know nothing about, to make different parts of it more concrete, such that you can then go and google the right words to find out more from sources with high verifiability.

I think most viable use cases today live in the bottom and towards the left, but the most exciting use cases live in the top right.

Recall vs synthesis

Another important spectrum is when your use case relies on more on recall versus synthesis. Asking for the capital of France is recall, while generating a poem is synthesis. Generating a poem using the names of all cities in France is somewhere in between.

At the moment, LLMs are clearly better at synthesis than recall, and it makes sense when you consider how they work. Indeed, most of the downfalls come from when they're a bit too loose with making stuff up.

Personally, I think that recall use cases are very under-explored at the moment, and have a lot of potential. This contrast is painted quite well when comparing two recent posts on HN. The first is about someone who trained nanoGPT on their personal journal here and the output was not great. Similarly, Projects/amarbot used GPT-J fine-tuning and the results were also hit and miss.

The second uses GPT-3 Embeddings for searching a knowledge base, combined with completion to have a conversational interface with it here. This is brilliant! It solves the issues around needing the results to be as correct as possible, while still assisting you with them (e.g. if you wanted to ask for the nearest restaurants, they better actually exist)!

Somebody in the comments linked gpt_index so you can do this yourself, and I really think that this kind of architecture is the real magic dust that will revolutionise both search and discovery, and give search engines a run for their money.