Hacker News Reader

Comment by simonw | Hacker News Reader

"I'm not sure what the answer really is, but I'm not sure it's GPT."

My suggestion is to stick with it and get a feel for what it's good at.

I've found that after a few months of using ChatGPT every day I've developed a pretty solid intuition for which questions are likely to get good answers and which are likely to trigger hallucinations.

It's difficult to describe what those intuitions are though!

One rule of thumb I've developed: if something is likely to be "common knowledge" - if it's something that is likely to have been discussed accurately on the internet by many different people - then ChatGPT is very likely to answer questions about it accurately.

View on HN

> One rule of thumb I've developed: if something is likely to be "common knowledge" - if it's something that is likely to have been discussed accurately on the internet by many different people - then ChatGPT is very likely to answer questions about it accurately.

If so, this information is already easy to find, making GPT redundant.

it being commonly discussed doesn't mean easily accessible. Personally ads and bad UI are main factors of my increased usage of gpt-4.

Is that just the problem that ChatGPT is solving? That we have most of the answers to everyone's questions on the web but that ads and SEO have made it impossible to access?

If so it seems like 1. that problem is of our own making and could be fixed without ChatGPT and 2. what will actually happen is that 'experts-exchange' and all the similar slightly scammy help sites and forums are going to try to stop LLMs from stealing their lunch.

I think it's more about being more confident that GPT is working with at least a base of semi-accurate knowledge.

You can ask it to "combine" knowledge in "novel" ways that are not discussed verbatim on the web. It's not groundbreaking reasoning by any means, but it can be very useful. ("My ridiculously specific question about model X83844-QQ combined with random factor X")

If you don’t mind a query language that is mostly grunting out nouns.

Let me give you a concrete counter example; I wanted to review DNS records the other day which is a topic I had covered in undergrad. Searching on google retrieved a bevy of poorly written articles that were mostly noteworthy for bow they glossed over the material I wanted to review.

I asked chat gpt to review the topic for me and it wrote a helpful summary that tracked with what I remembered and provided enough keywords that provided better search results - fact checking and digging for more detail became much easier when I was able to find the exact wikipedia page I needed (and other resources).

> then ChatGPT is very likely to answer questions about it accurately.

Uh, and sooo boringly, especially if there's even any tiny part of it that is developing or theoretical, and you want to learn about that part.

I've used the phrase/request "more obscure (perspectives/explanations/etc)" with GPT so many times.

(Try it, you might be surprised at alternative takes on things, takes which are not even necessarily conspiracy theories and such)

...Which has made me think: Maybe life is more boring, the more one thinks there are just really good one-and-done answers to most everything.

Yeah, there are so many fun ways you can make it more interesting.

One of my favourite is to ask it to explain with analogies. The other day I was digging into the attention mechanism used to train LLMs, so I asked it:

"Explain queries, keys, and values in the context of LLM attention using analogies from Terry Pratchett's Discworld"

I find sometimes I get some real gems out of this that help me remember things much more effectively than just reading the basic explanation.

(In case anyone's curious I ran that just now and got the following: https://gist.github.com/simonw/777b1d19f36beb39fb4216a0238fe... )

Jonathan Blow is spending up to a decade on a new programming language to get away from c++.

Most everything has been covered in logical extensions of concepts started in the 80s-90s. Those products aren't good for our..

Ah there's no point commenting here anymore. HN is so blinkered in it's thinking.

The feeling of flying high on lofty concepts and pretending to get a bird's eye view of tech is no longer worth the squeeze. The thought patterns here are predictable like slashdot. I don't belong here. Bye.

Yep, if you don't care about the details or scope, everything is the same, all patterns repeat, all wheels are reinvented.

You've probably heard this before, but IMO you'll never find a happy place working from that perspective.

Try switching to the OpenAI playground in the API console. You can adjust the system prompt and the temperature of the response. Higher temperatures = less predictable responses.

Thanks for the tip. I set it to 1 and still get the same effect...like should this affect e.g. book recommendations?

Recommend 10 project gutenberg books that are fun to read --> same old stuff--Huck Finn, Robinson Crusoe, War of the Worlds.

But I'm not sure if this particular use is really a temperature-relevant thing or not.

Reply on news.ycombinator.com