GPT makes learning fun again • Hacker News Board

Re: GPT makes learning fun again

#231

Post by vaidhy » Tue, May 02, 2023, 6:27 PM UTC

hammyhavoc wrote: post #65

LASR wrote: post #64

Earlier quoted context omitted.

I want to use this comment to state one thing, not directly addressed to you. Stop using GPT as a database! GPT is far more useful a reasoning engine that can accumulate fuzzy data and then provide various views or transformation of that data. So asking GPT to parse a Wikipedia page and then asking it to teach you from it - this is a much more successful usage than what the author in the original article is doing. It…

Is it? I couldn't even get it to figure out an NGINX config, despite feeding it the documentation and URL rewrites it would need, and the prior Apache .htaccess file. Spent days on it. Consulted with people who know more than me about prompt engineering. Nope. Reasoning is not its strong point, IMO. It's a next-word prediction model, why would it be? It's doing what an LLM does. Frequently, nonsensically.

I will paraphrase Geoff Hinton here - Predicting next word requires understanding context. Otherwise, you cannot predict the next word. It is a big deal.

Re: GPT makes learning fun again

#232

Post by sangnoir » Tue, May 02, 2023, 6:29 PM UTC

Frost1x wrote: post #153

seba_dos1 wrote: post #140

Earlier quoted context omitted.

If "chaining some simple API calls" is "a large part of your work", I can see it being incredibly useful. I already used it in this way a few times and it saved me some time indeed. However, as a software developer I feel like it's actually a tiny minority of my work; most of it is "doing some thinking" and slowly unfolding the inner parts of systems in order to understand how they work and how to fix them or make th…

Actually this is the aspect I love about GPT. The less I have to spend dealing with idiosyncrasies of arbitrary API boiler plate and can get down to the business of leveraging said things together or designing aspects of said things to achieve a new goal, the better. As someone who is far from an expert in DevOps, this has saved me a lot of time recently, and it worked fairly well to get past duct taping infrastructu…

Who knew the next "billion dollar security mistake" after C's memory handling would be LLM-suggested misconfigurations? I bet someone will try to sell an LLM-based service to fix thr bugs caused by other code-generating LLMs.

Re: GPT makes learning fun again

#233

Post by dmbche » Tue, May 02, 2023, 6:42 PM UTC

abdou4a wrote: post #200

dmbche wrote: post #52

Earlier quoted context omitted.

Sticking to what you want to understand - for example, when reading a paper, you don't necessarily need to read the methodology, especially if it's out of your field. Read the abstract and the conclusion, identify any part of it that you are suprised by and would like further explanation, and go see that part of the paper. A lot of the paper is talking to peer and people wanting to verify the validity of the paper -…

>> you can mostly assume that the paper is valid you should check this https://en.wikipedia.org/wiki/Replication_crisis https://en.wikipedia.org/wiki/Data_dredging or P-hacking and many more...

Yes! I'm well aware of the replication crisis.

While this is valid, this shouldn't make us paranoid of all papers. If it's a paper out of your expertise, it's unlikely that you would be able to catch problems with it that the peer review process wouldn't have caught.

So in practice, it doesn't really change the way you interact with papers - it should change the way people write them and how the peer review process works.

Re: GPT makes learning fun again

#234

Post by gwd » Tue, May 02, 2023, 6:51 PM UTC

notamy wrote: post #171

lcuff wrote: post #25

Earlier quoted context omitted.

I'm curious about what topics you have knowledge of where it's failed. Does it seem like there's a pattern to the failures? I've been using GPT for coding help, and it it is very helpful in ruby and bash, though it often delivers buggy software: Badly handled non-happy-path conditions, mostly, which when I tell it to handle the case, it may. It's a huge help for me finding gems and showing me standard ruby library sy…

> Does it seem like there's a pattern to the failures? Personally, trying to use it to write code (primarily Elixir backend and Rust systems/CLI), I tend to run into: - Hallucinating APIs that don't exist - Hallucinating entire libraries, despite being told repeatedly they don't exist - Saying it will make requested changes and not doing so - Not being anywhere close to idiomatic code - Not being able to explain code…

> I use boring normal free ChatGPT so maybe it's on me for not using GPT-4 or some other model, but either way, imo it's not been very impressive in the problem spaces I find myself in.

You can use GPT-4 fairly cheap if you sign up for a developer account on platform.openai.com and then use their playground. There you pay per usage; even though I've been using it fairly heavily, my typical monthly usage is still way under $20.

Here's an example that both impressed me and saved me a load of time just yesterday:

https://gitlab.com/-/snippets/2535443

Re: GPT makes learning fun again

#235

Post by olddustytrail » Tue, May 02, 2023, 7:12 PM UTC

chx wrote: post #90

p-e-w wrote: post #45

Earlier quoted context omitted.

> The information value of answers by ChatGPT is zero. Nonsense. The only way this would be true is if ChatGPT's answers were as likely to be correct as random answers. They are much, much more likely to be correct, however, so their information value is greater than zero by definition. What you're claiming is equivalent to "search engines can find incorrect information, so search engines are worthless for informatio…

> The only way this would be true is if ChatGPT's answers were as likely to be correct as random answers. Which is the case > They are much, much more likely to be correct, They are not. It's just apophenia to find meaning in the word salad.

Of course it's not the case. You do realise that people have measured this?

The irony of people complaining that LLMs confidently assert things that are completely wrong...

Re: GPT makes learning fun again

#236

Post by rapsacnz » Tue, May 02, 2023, 8:08 PM UTC

wrycoder wrote: post #219

rapsacnz wrote: post #75

Don't forget that GPT was trained on data from all the sources that failed you. So if we all collectively neglect them, and they fall over, we will lose many valuable resources. I think we need to think about how to keep these valuable sites going, because they are ultimately providing most of the value of the various available language models.

The sources didn’t fail. The point is, especially at the beginning, it’s easy to miss seeing the forest for all the trees.

By "failed" I meant "Proved hard to parse". I don't think they failed at all.

Re: GPT makes learning fun again

#237

Post by seb1204 » Tue, May 02, 2023, 8:52 PM UTC

kccqzy wrote: post #222

seb1204 wrote: post #129

Earlier quoted context omitted.

Not that harsh but I do think that googling as well as ChatGPT will only yield in a superficial understanding of topics. For a deeper profound understanding, connections, complexities, etc. A different more holistic learning is required. Not sure how else to explain this.

Agreed. I look back to my high school and college education and can't help but think that for a deeper understanding, exercises must be needed. It is too easy to think you have learned something but then fail at the easiest synthetic problem that requires you to use what you have learned to solve a problem.

Yes, I did write 'exercise' in my original post but then deleted it as I felt it was too much like asking to make it like school.

Re: GPT makes learning fun again

#238

Post by QuantumGood » Wed, May 03, 2023, 12:12 AM UTC

LASR wrote: post #64

devjab wrote: post #7

I'm in two minds about it. On one hand the internet (in general) has become so hard to find information on, and I agree with the author that GPT is a breath of fresh air. On the other hand I've seen GPT fail so miserably at topics that I'm knowledge about that I have a very hard time trusting anything it tells me. I'm not sure what the answer really is, but I'm not sure it's GPT. I wish we could go back to having sea…

I want to use this comment to state one thing, not directly addressed to you. Stop using GPT as a database! GPT is far more useful a reasoning engine that can accumulate fuzzy data and then provide various views or transformation of that data. So asking GPT to parse a Wikipedia page and then asking it to teach you from it - this is a much more successful usage than what the author in the original article is doing. It…

I asked it give me specs for a well-known microphone. Over the course of several queries, it gave three different numbers for the requested spec, none of them matching what the manufacturer gave on their website.

Re: GPT makes learning fun again

#239

Post by throwaway2990 » Wed, May 03, 2023, 12:46 AM UTC

mrtranscendence wrote: post #229

throwaway2990 wrote: post #211

Earlier quoted context omitted.

Lol I’m not sure you understand what an example is.

OK, tell me what an example is, if not a description of something that lends -- or removes -- credence toward a position. And I do want to know if you think GPT-4 is actually incapable of making mistakes, because that's what you seem to be implying (even though it's flatly ridiculous).

And your “example” wasn’t even close to that.

No one can deduce anything from it and attempt to reproduce the problem to see where it failed.

You basically wrote “I asked for white and the mail was grape juice is the best!”

What can anyone do with that? Nothing. It’s meaningless.

This is why open source projects hate people who raise issues because they are like “error happened fix it” and say nothing more then get upset when the issue is closed.

Re: GPT makes learning fun again

#240

Post by JoshuaDavid » Wed, May 03, 2023, 1:09 AM UTC

kumarvvr wrote: post #27

This is so ridiculous. GPT is like that "know-it-all" friend we have who just has something to say about anything , with knowledge skimmed from the internet. GPT is a language model. It outputs what you want to hear , not what is correct.

Sometimes what I want to hear is "the thing that the typical person who writes about this sort of thing would say about my particular question". If I can figure out whether or not to trust an internet rando who may or may not know what they're talking about, I can figure it out when I'm talking to a simulated internet rando.