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

I genuinely don't understand why some people are still bullish about LLMs

twitter.com

Post reply on HN
Print view View original

601–610 of 1001 posts

Re: I genuinely don't understand why some people are still bullish about LLMs

#601

Post by bni »

gilbetron wrote: post #109

I get so confused on this. I play around, test, and mess with LLMs all the time and they are miraculous. Just amazing, doing things we dreamed about for decades. I mean, I can ask for obscure things with subtle nuance where I misspell words and mess up my question and it figures it out. It talks to me like a person. It generates really cool images. It helps me write code. And just tons of other stuff that astounds me…

I don't like tools that can't be trusted to work 100% of the time. Is this hard to grasp?

Re: I genuinely don't understand why some people are still bullish about LLMs

#602

Post by otabdeveloper4 »

wruza wrote: post #402
adastra22 wrote: post #317

Earlier quoted context omitted.

It really depends on the task. Like Sabine, I’m operating on the very frontier of a scientific domain that is extremely niche. Every single LLM out there is worse than useless in this domain. It spits out incomprehensible garbage. But ask it to solve some leet code and it’s brilliant.

At non-extremely niche tasks they fail as well. I should start collecting examples, if only for threads like this. Recently I tried to llm a tsserver plugin that treats lines ending with "//del" as empty. You can only imagine all the sneaky failures in the chat and the total uselessness of these results. Anything that is not literally millions (billions?) of times in the training set is doomed to be fantasized about…

LLMs are basically a search engine for Stack Overflow and Github that doesn't suck as bad as Google does.

If your job is copy-pasting from Stack Overflow then LLMs are an upgrade.

Re: I genuinely don't understand why some people are still bullish about LLMs

#604

Post by otabdeveloper4 »

Animats wrote: post #178
cogman10 wrote: post #161

Earlier quoted context omitted.

> Wah, it can't write code like a Senior engineer with 20 years of experience! No, that's not my problem with it. My problem with it is that inbuilt into the models of all LLMs is that they'll fabricate a lot. What's worse, people are treating them as authoritative. Sure, sometimes it produces useful code. And often, it'll simply call the "doTheHardPart()" method. I've even caught it literally writing the wrong algor…

> inbuilt into the models of all LLMs is that they'll fabricate a lot. Still the elephant in the room. We need an AI technology that can output "don't know" when appropriate. How's that coming along?

> How's that coming along?

It isn't. LLMs are autocomplete with a huge context. It doesn't know anything.

Re: I genuinely don't understand why some people are still bullish about LLMs

#606

Post by DrScientist »

readthenotes1 wrote: post #23
jongjong wrote: post #7

People who don't work in tech have no idea how hard it is to do certain things at scale. Skilled tech people are severely underappreciated. From a sub-tweet: >> no LLM should ever output a url that gives a 404 error. How hard can it be? As a developer, I'm just imagining a server having to call up all the URLs to check that they still exist (and the extra costs/latency incurred there)... And if any URLs are missing,…

"How hard can it be?" If I recall correctly, that is one of Dilbert's management axioms: if I don't understand it it cannot be difficult

Indeed.

And I have used the following response to pointy haired bosses on a couple of occasions ( though I don't recommend it ).

'If it's so easy - feel free to do it yourself'.

Re: I genuinely don't understand why some people are still bullish about LLMs

#607

Post by poincaredisk »

ThrowawayTestr wrote: post #339
adastra22 wrote: post #317

Earlier quoted context omitted.

It really depends on the task. Like Sabine, I’m operating on the very frontier of a scientific domain that is extremely niche. Every single LLM out there is worse than useless in this domain. It spits out incomprehensible garbage. But ask it to solve some leet code and it’s brilliant.

Surely you understand why an LLM that has no knowledge of your niche wouldn't be useful right?

I work in a field that is not even close to a scientific nishe - software reverse engineering - and LLM will happily lie to me all the time, for every question I have. I find out useful to generate some initial boilerplate but... that's it. AI autocompletion saved me an order of magnitude more time, and nobody is hyped about it.

Re: I genuinely don't understand why some people are still bullish about LLMs

#608

Post by snitty »

squiggleblaz wrote: post #567
NoGravitas wrote: post #316

Earlier quoted context omitted.

I frequently see news stories where attorneys get in trouble for using LLMs, because they cite hallucinated case law (e.g.). If they didn't get caught, that would look the same as using them "productively".

Asking the LLM for relevant case law and checking it up - productive use of LLM. Asking the LLM to write your argument for you and not checking it up - unproductive use of LLM. It's the same as with programming.

>Asking the LLM for relevant case law and checking it up - productive use of LLM

That's a terrible use for an LLM. There are several deterministic search engines attorneys use to find relevant case law, where you don't have to check to see if the cases actually exist after it produces results. Plus, the actual text of the case is usually very important, and isn't available if you're using an LLM.

Which isn't to say they're not useful for attorneys. I've had success getting them to do some secretarial and administrative things. But for the core of what attorneys do, they're not great.

Re: I genuinely don't understand why some people are still bullish about LLMs

#609

Post by fifticon »

somenameforme wrote: post #554
gilbetron wrote: post #109

I get so confused on this. I play around, test, and mess with LLMs all the time and they are miraculous. Just amazing, doing things we dreamed about for decades. I mean, I can ask for obscure things with subtle nuance where I misspell words and mess up my question and it figures it out. It talks to me like a person. It generates really cool images. It helps me write code. And just tons of other stuff that astounds me…

It's mostly because of how they were initially marketed. In an effort to drive hype 'we' were promised the world. Remember the "leaks" from Google about an engineer trying to get the word out that they had created a sentient intelligence? In reality Bard, let alone whatever early version he was using, is about as sentient as my left asscheek. OpenAI did similar things by focusing to the point of absurdity on 'safety'…

I am worried about them as a substitute for search engines. My reasoning is that classic google web-scraping and SEO, as shitty as it may be, is 'open-source' (or at least, 'open-citation') in nature - you can 'inspect the sh*t it's built from'. Whereas LLMs, to me seem like a chinese - or western - totalitarian political system wet dream - 'we can set up an inscrutable source of "truth" for the people to use, with the _truths_ we intend them to receive'. We already saw how weird and unsane this was, when they were configured to be woke under the previous regime. Imagine it being configured for 'the other post-truth' is a nightmare.

Re: I genuinely don't understand why some people are still bullish about LLMs

#610

Post by matkoniecz »

CamperBob2 wrote: post #151
zozbot234 wrote: post #124

Earlier quoted context omitted.

> I mean, I can ask for obscure things with subtle nuance where I misspell words and mess up my question and it figures it out. If you're lucky it figures it out. If you aren't, it makes stuff up in a way that seems almost purposefully calculated to fool you into assuming that it's figured everything out. That's the real problem with LLM's: they fundamentally cannot be trusted because they're just a glorified autocom…

they don't come with any inbuilt sense of when they might be getting things wrong Spend some time with current reasoning models. Your experience is obsolete if you still hold this belief.

Can you be more specific than "current reasoning models"? Maybe I missed it, but I have not yet seen any that would not hallucinate wildly.
Post reply on HN