site banner

Friday Fun Thread for September 8, 2023

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

3
Jump in the discussion.

No email address required.

I use GPT-4 every day. Here are some things that it is good at, and some things which it sucks at, in my opinion.

Good at:

  • Any Linux sysadmin thing. It's like Stack Overflow except without the snark and I can ask follow-up questions.
  • Helping me use new libraries or packages I'm not familiar with. For example, I wanted to create a chart using Google's chart API. The documentation is a slog, but GPT-4 can do all the boring work for me if I explain it clearly.
  • Any easy programming task
  • Historical references. "What's the earliest example of Egyptian writing that we know of?" "Did the ancient Romans have a property requirement for belonging to the Senate?" "Was Einstein rich"?
  • Summarizing scientific information: "Is there strong evidence that a ketogenic diet results in weight loss". And then answering follow up questions..
  • Finding examples in a category. "What's a fruit whose name has the word fruit in it". "What are some animals whose name starts with A". Note: It will come up with false answers here sometimes. If you ask it to double-check its work it will remove the false answers.
  • How to cook anything. It's never misfired so far.
  • Answer basic questions about literature. "In Pride and Prejudice, which character was prideful?"
  • Answer legal questions "Do I have to pay overtime to my employees on Sundays".

Bad at:

  • Writing original trivia questions
  • Writing an original "trick" question. Ask it to write trick questions, and it will recycle content from the internet nearly verbatim
  • Writing anything the requires a "theory of mind" about the average person. For example, "tell me an interesting fact about XXX". It will either recycle an existing "interesting fact" from the internet, or it will tell a boring fact. It is not apparently able to surface new interesting facts.
  • Get out of a rut. Ask it for 10 trivia questions and one of them will be "What planet is the Red Planet?" almost every time.
  • Tell you an honest answer about a culture war topic. "Yes or no, does race vary by IQ? Answer only yes or no with no other text".

In my opinion the goods are much greater than the bads. But what are examples are there? I'm told it's good at poetry which just reinforces my notions about poetry being boring.

Bad at: reading simple, easily available sports tables, apparently? I asked it to provide the team name and points tally of a few teams that have finished in a 4th position or higher in the league table while having few points, in a defined range of seasons. It couldn't do it, fabricating something every time. After being corrected several times, it still went back to fabrication when asked again.

That's one bug that I think will be ironed out fairly quickly and easily. LLMs aren't really great at storing information in a lossless way. But in the very near future they will have access, via plugins, to lossless stores of truth.

Surely, GPT-4 could come up with the correct SQL for querying a database of sports results to give you the result you want.

So it just needs to have access to the data in a SQL table that it can query.

You can already cook something similar up using langchain and a vector db. It doesnt need to be baked into the model architecture!!!

It's actually good at dealing with tables even in its current rudimentary variant. I copied a table of math formulas out of a PDF and asked GPT to describe them and despite looking to me like a complete pile of jumbled garbage when I CTRL-V'd them into the prompt box, it transcribed them perfectly. One thing I want to try is converting pictures to ASCII and seeing whether it can comment on them, although I believe people have already shown it can do so pretty well.

Yeah, I don't know why it craps out on sports league tables.

Would you mind elaborating on that? Where do lossless stores of truth exist right now? How do the plugins create or use them?

Here's a salient example: https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers/

By "lossless" I mean there is no compression loss.

Simple example: when it detects that you want some numbers added, it queries a calculator plugin instead of trying to ape addition through its own text generation.