[simonw]

As a programmer, large language models let me solve an increasing range of problems that I couldn't have solved without them. So I think they are a very big deal.

Just one example: parsing structured data out of a big pile of poorly formatted PDF documents.

That used to be too difficult and expensive for me to tackle without a small army of data entry people to help do the work.

Today I can point Textract OCR at it and then use a language model to extract structured data.

(I haven't implemented this particular example just yet, but I'm looking for an opportunity to do so.)

[6keZbCECT2uB]

I think that ChatGPT could be a big accelerator for creative activity, but I wouldn't trust any output that I've only partially verified. That limits it to human scale problems in its direct output, but there are many ways that human scale output like code snippets can be useful on computer scale data.

[simonw]

There are plenty of valuable use-cases that aren't at much risk from hallucinations at all:

- Asking it to summarize text

- Using it to extract facts from text and present them in an alternative format - turning a chunk of HTML into JSON for example

- Creative writing - poems, stories etc

- Getting feedback on your own text - asking it what should be tightened up, which bits are confusing and so on

- All kinds of code generation activities

[sarchertech]

> turning a chunk of HTML into JSON for example

I haven't done exactly that, but based on similar examples this is likely very vulnerable to hallucinations.

[simonw]

For simple things it's pretty safe. I tried pasting in HTML from the homepage of Hacker News and having it turn that into a list of JSON objects each with the title, submitter, number of upvotes and number of comments.

Here's a similar trick I did with Copilot: https://til.simonwillison.net/gpt3/reformatting-text-with-co...

[gcanyon]

There's two classes of response: those that are factually "right" or "wrong" -- who was the sixteenth president of the U.S.? And those that are opinion/debatable: "How should I break up with my boyfriend?" People will focus on the facts, and those will be improved (viz: the melding of ChatGPT with Wolfram Alpha) but the opinion answers are going to be more readily acceptable (and harder to optimize?).

[essentialfloss]

I have a project I'm working on where I need to turn a bunch of bank statements into spreadsheets, is that an example what you're talking about? If so, I am extremely interested and would love to know how to implement it.

[abrichr]

Look into PaddleOCR layout analysis.

Sounds like an interesting project! Who are the users?