Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Life is more than an engineering problem

lareviewofbooks.org

Post reply on HN
Print view View original

81–90 of 257 posts

Re: Life is more than an engineering problem

#81

Post by emtel »

benreesman wrote: post #75
emtel wrote: post #70

I like Chiang’s stories, but he simply doesn’t know what he’s talking about when it comes to LLMs. Forget the philosophical points - he claims that LLMs are strictly worse than search engines for information retrieval. This is just obviously false, I can give you dozens of conversations from my ChatGPT history over the past two months which would have been impossible to get answers to on Google.

LLMs are generally better search tools for certain topics today , but search engines have been in serious decline for at least 15 years. IMHO old-school Google remains the high water mark of generalized information retrieval, with advantages ranging from speed to semi-durable citation. I strongly suspect there is a cohort thing going on here, many HN users today weren’t involved in technology yet back when Google wor…

I’ve been using Google since about 2000/2001 or so. In fact I worked there for a time, and I still remember Craig Silverstein’s answer to “isn’t search a solved problem?” Which was: “in the future we want Google to be able to take any question and just give you the answer. So clearly we have a long way to go!”

In any case, I do not believe there was ever a time it could answer all of the questions that LLMs can today. If the question had been asked and answered on the web, Google could (and can) find it, but many questions haven’t been asked!

Re: Life is more than an engineering problem

#82

Post by bmitc »

emtel wrote: post #70

I like Chiang’s stories, but he simply doesn’t know what he’s talking about when it comes to LLMs. Forget the philosophical points - he claims that LLMs are strictly worse than search engines for information retrieval. This is just obviously false, I can give you dozens of conversations from my ChatGPT history over the past two months which would have been impossible to get answers to on Google.

What are the examples?

I've had several LLM search result summaries contain flat out mistakes and incorrect statements.

Re: Life is more than an engineering problem

#83

Post by fmbb »

mofeien wrote: post #79
abecedarius wrote: post #68

Chiang makes some insightful points, e.g. about what we mean by magic. Then I come to > [LLMs] can get better at reproducing patterns found online, but they don’t become capable of actual reasoning; it seems that the problem is fundamental to their architecture. and wonder how an intelligent person can still think this, can be so absolute about it. What is "actual" reasoning here? If an AI proves a theorem is it only…

To me this also feels like a statement that would obviously need strong justification. For if animals are capable of reasoning, probably through being trained on many examples of the laws of nature doing their thing, then why couldn't a statistical model be?

> For if animals are capable of reasoning

Are they? Which animals? Some seem smart and maybe do it. Needs strong justification.

> probably through being trained on many examples of the laws of nature doing their thing

Is that how they can reason? Why do you think so? Sounds like something that needs strong justification.

> then why couldn't a statistical model be?

Maybe because that is not how anything in the world attained the ability to reason.

A lot of animals can see. They did not have to train for this. They are born with eyes and a brain.

Humans are born with the ability to recognize pattern in what we see. We can tell objects apart without training.

Re: Life is more than an engineering problem

#84

Post by fragmede »

ustad wrote: post #77
abecedarius wrote: post #68

Chiang makes some insightful points, e.g. about what we mean by magic. Then I come to > [LLMs] can get better at reproducing patterns found online, but they don’t become capable of actual reasoning; it seems that the problem is fundamental to their architecture. and wonder how an intelligent person can still think this, can be so absolute about it. What is "actual" reasoning here? If an AI proves a theorem is it only…

Has an AI “proven” a theorem?

Leading mathematician Terence Tao says yes, with a lot of guidance.

https://mathstodon.xyz/@tao/113132503432772494

Re: Life is more than an engineering problem

#85

Post by sleepytree »

abecedarius wrote: post #68

Chiang makes some insightful points, e.g. about what we mean by magic. Then I come to > [LLMs] can get better at reproducing patterns found online, but they don’t become capable of actual reasoning; it seems that the problem is fundamental to their architecture. and wonder how an intelligent person can still think this, can be so absolute about it. What is "actual" reasoning here? If an AI proves a theorem is it only…

If I write a book that contains Einstein's theory of relativity by virtue of me copying it, did I create the theory? Did my copying of it indicate anything about my understanding of it? Would you be justified to think the next book I write would have anything of original value?

I think what he is trying to say is that LLMs current architecture seems to mainly work by understanding patterns in the existing body of knowledge. In some senses finding patterns could be considered creative and entail reasoning. And that might be the degree to which LLMs could be said to be capable of reasoning or creativity.

But it is clear humans are capable of creativity and reasoning that are not reducible to mere pattern matching and this is the sense of reasoning that LLMs are not currently capable of.

Re: Life is more than an engineering problem

#86

Post by emtel »

bmitc wrote: post #82
emtel wrote: post #70

I like Chiang’s stories, but he simply doesn’t know what he’s talking about when it comes to LLMs. Forget the philosophical points - he claims that LLMs are strictly worse than search engines for information retrieval. This is just obviously false, I can give you dozens of conversations from my ChatGPT history over the past two months which would have been impossible to get answers to on Google.

What are the examples? I've had several LLM search result summaries contain flat out mistakes and incorrect statements.

I’ll try to dig some up soon (I’m on my phone now). But of course the output contains errors sometimes. So do search engine results. The important thing for difficult questions is whether the right answer (or something pointing toward it) is available _at all_. Of course this assumes you can verify the answers somehow (usually easy with programming questions), but again, search engines have the same limitation.

Re: Life is more than an engineering problem

#87

Post by cyrillite »

mewpmewp2 wrote: post #74
abecedarius wrote: post #68

Chiang makes some insightful points, e.g. about what we mean by magic. Then I come to > [LLMs] can get better at reproducing patterns found online, but they don’t become capable of actual reasoning; it seems that the problem is fundamental to their architecture. and wonder how an intelligent person can still think this, can be so absolute about it. What is "actual" reasoning here? If an AI proves a theorem is it only…

Even most intelligent people can hallucinate, we still haven't fixed this problem. There's a lot of training material and bias which leads many to repeat those things "LLM's are just a stochastic parrot, glorified auto complete/google search, Markov chains, just statistics", etc. The thing is, these sentences sound really good and so it's easy to repeat them when you have made up your mind. It's a shortcut.

They’re right until they’re wrong.

AI is (was?) a stochastic parrot. At some point AI will likely be more than that. The tipping point may not be obvious.

Re: Life is more than an engineering problem

#88

Post by freejazz »

abecedarius wrote: post #68

Chiang makes some insightful points, e.g. about what we mean by magic. Then I come to > [LLMs] can get better at reproducing patterns found online, but they don’t become capable of actual reasoning; it seems that the problem is fundamental to their architecture. and wonder how an intelligent person can still think this, can be so absolute about it. What is "actual" reasoning here? If an AI proves a theorem is it only…

>and wonder how an intelligent person can still think this, can be so absolute about it.

I wonder how people write things like this and don't realize they sound as sanctimonious as exactly whatever they are criticizing. Or, if I was to put it in your words: "how could someone intelligent post like this?"

Re: Life is more than an engineering problem

#89

Post by OptionOfT »

emtel wrote: post #70

I like Chiang’s stories, but he simply doesn’t know what he’s talking about when it comes to LLMs. Forget the philosophical points - he claims that LLMs are strictly worse than search engines for information retrieval. This is just obviously false, I can give you dozens of conversations from my ChatGPT history over the past two months which would have been impossible to get answers to on Google.

I think those are 2 different issues.

Searching for something, and finding it, is different from what ChatGPT / Claude does.

Google (in the good old days) is like the library. You want to search 'how to plant potatoes on Mars'. No results. Well, you split it up, maybe a book on planting potatoes, and a book about missions to Mars that describes soil composition.

Then, when you have those books you start reading, parsing, understanding, making connections, identifying what needs to be done etc.

Maybe, if you're lucky, you find a book or a web page where somebody went through the thought exercise of finding out what would be needed to make it work.

ChatGPT / Claude / ... are different in that they have the information in their corpus, and that the information they present you could actually be something that has never been written down in a book, or published on the web. That's why Google can't find it, but ChatGPT is able to present you with a satisfying answer.

Now whether the answer is correct is a different issue. Do you have the knowledge to verify this?

Re: Life is more than an engineering problem

#90

Post by bmitc »

emtel wrote: post #86
bmitc wrote: post #82

Earlier quoted context omitted.

What are the examples? I've had several LLM search result summaries contain flat out mistakes and incorrect statements.

I’ll try to dig some up soon (I’m on my phone now). But of course the output contains errors sometimes. So do search engine results. The important thing for difficult questions is whether the right answer (or something pointing toward it) is available _at all_. Of course this assumes you can verify the answers somehow (usually easy with programming questions), but again, search engines have the same limitation.

There isn't an expectation or claim that search engines answer anything. They just find things or don't find things.
Post reply on HN