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

What we know about LLMs

willthompson.name

Post reply on HN
Print view View original

91–100 of 173 posts

Re: What we know about LLMs

#91

Post by bilsbie »

RC_ITR wrote: post #70
bilsbie wrote: post #59

Earlier quoted context omitted.

When you’re doing RLHF are you actually modifying the weights of llama itself? Or is something on top?

RLHF does change the parameters. The way to think about it is that backpropagation changes the parameters of a model so they get closer to some sort of desired output. In pre-training and SFT, the parameters are changed so the model does a better job of replicating the next word in the training data, given the words it has already seen. In RLHF, the parameters are changed so the model does a better job of outputting…

Thanks. That helps.

So how can you update weights without doing back-propagation? Or is it still back propagation but with a different metric?

Re: What we know about LLMs

#92

Post by obiefernandez »

killthebuddha wrote: post #75
killernap wrote: post #13

ChatGPT was announced November, 2022 - 8 months ago. Time flies. Question for HN: Where are we in the hype cycle on this? We can run shitty clones slowly on Raspberry Pi's and your phone. The educational implementations demonstrate the basics in under a thousand lines of brisk C. Great. At some point you have to wonder... well, so what? Not one killer app has emerged. I for one am eager to be all hip and open minded…

I don't want to make any real assertions but my intuitive reaction to this comment is _this person has no clue what they are talking about_. I would rather turn of syntax highlighting than turn off Copilot and I'd rather disable Google search rather than ChatGPT. And frankly, it's not even close, I use these tools "all the time for everything".

Not one killer app…

I stopped reading there. So ignorant it’s painful

Re: What we know about LLMs

#93

Post by adventured »

nathanfig wrote: post #24
killernap wrote: post #13

ChatGPT was announced November, 2022 - 8 months ago. Time flies. Question for HN: Where are we in the hype cycle on this? We can run shitty clones slowly on Raspberry Pi's and your phone. The educational implementations demonstrate the basics in under a thousand lines of brisk C. Great. At some point you have to wonder... well, so what? Not one killer app has emerged. I for one am eager to be all hip and open minded…

If you follow the Gartner model, there is usually a surge of high expectations right before a "trough of disillusionment" - but eventually the real applications do emerge. Humans are just impatient.

That whole Internet thing is never going to work out.

What's the killer app? The Web? CompuServe? AOL chat? ICQ? Blue's News? WebChat Broadcasting? Real Audio? Broadcast.com? GeoCities? It's all ridiculous.

I mean, have you tried to search for anything on AltaVista, Excite or Lycos?!? They hardly work, all you get is page after page of garbage results!

And don't even get me started with eBay or CDNow. Nobody is going to shop online.

Re: What we know about LLMs

#94

Post by bradfox2 »

bilsbie wrote: post #91
RC_ITR wrote: post #70

Earlier quoted context omitted.

RLHF does change the parameters. The way to think about it is that backpropagation changes the parameters of a model so they get closer to some sort of desired output. In pre-training and SFT, the parameters are changed so the model does a better job of replicating the next word in the training data, given the words it has already seen. In RLHF, the parameters are changed so the model does a better job of outputting…

Thanks. That helps. So how can you update weights without doing back-propagation? Or is it still back propagation but with a different metric?

It's still loss being backproped, but the loss is calculated over a different criteria

Re: What we know about LLMs

#95

Post by FishInTheWater »

famouswaffles wrote: post #82
FishInTheWater wrote: post #69

Earlier quoted context omitted.

This is where the terminology becomes a bit annoying, but there is a key difference in the kinds of reasoning at work here. When you ask LLMs to provide a reasoning, the actual reasoning performed is linguistic; The LLM has (is) a model about language and performs some (limited) reasoning on that model to get an output. But that is explicitly different from reasoning about the abstract question at hand, thus the answ…

>Thus we can observe that LLMs do not abstractly reason about the question and it's model. Your conclusion makes no sense. Humans provide increasingly wrong answers as questions get more complex too. Jumping from that to "incapable of abstract reasoning" is silly. You have not "trivially proven" anything at all >The LLM has (is) a model about language and performs some (limited) reasoning on that model to get an outp…

Humans provide increasingly wrong answers as questions get more complex too.

Human this, Human that. LLMs aren't humans. "My model is crap but the human brain isn't very good at this either" is irrelevant when we have machines that are not only very good at these tasks but almost perfect at them.

Humans make such mistakes precisely because they are not perfect reasoning machines. To compare LLMs to humans is not only disingenuous, but proves my point.

(And no, I will not humour you with an argument about how the amount of wrong answers is drastically lower from human mathematicians)

Jumping from that to "incapable of abstract reasoning" is silly.

They are language models. It is explicitly what they are designed to do.

If these LLMs are not, as I claim, reasoning on language rather than the abstract model of the query, then how come they fail miserably in exactly the ways you would expect where that the case?

LLMs generalize to non linguistic patterns.

Yes, congratulations, if you turn a problem into a linguistic one LLMs can deal with them. This does not in any way go against what I said about the capabilities of LLMs.

The same levels of actual abstract reasoning can be achieved on a graphing calculator running off literal potatoes.

Re: What we know about LLMs

#96

Post by obiefernandez »

llm_nerd wrote: post #87
NhanH wrote: post #80

Earlier quoted context omitted.

> The Venn diagram of people at the forefront of ML/LLM, and its advocates, is almost entirely separate from the web/crypto sphere. There is astonishingly little overlap That statement seems false. Especially since this was a headline I saw in half a dozen online news in my country yesterday. > OpenAI's Sam Altman launches Worldcoin crypto project[0] If anything, even without taking into account the greed stuff, peop…

Me - "Little overlap" You - "That's false because I saw a thing about this one guy doing some thing" Okay... "even without taking into account the greed stuff," What "greed stuff"? It is incredibly hard to make money in LLMs/ML. The barriers to entry are colossal, and it is technically extremely difficult. Everyone keeps talking about all the "grifters" (a go to term that usually means the speaker's arguments can be…

I can assure you that the smart folks running daily AI newsletters are currently raking in money hand over fist. It’s awe inspiring.

Re: What we know about LLMs

#97

Post by sanderjd »

killernap wrote: post #13

ChatGPT was announced November, 2022 - 8 months ago. Time flies. Question for HN: Where are we in the hype cycle on this? We can run shitty clones slowly on Raspberry Pi's and your phone. The educational implementations demonstrate the basics in under a thousand lines of brisk C. Great. At some point you have to wonder... well, so what? Not one killer app has emerged. I for one am eager to be all hip and open minded…

I'd say we're maybe half or two-thirds of the way down from the peak of inflated expectations toward the trough of disillusionment. Before long, I think maybe in the next three months or so, certainly around the time we hit the one year anniversary of chatgpt's release, we'll start seeing mainstream takes along the lines of "chatgpt and Bing's Sydney episode and such were good entertainment, but it's obvious in hindsight that it was a fad; nobody is posting funny screenshots of their conversations anymore, and all the pronouncements about a superhuman AGI apocalypse were obviously silly, it's clear chatgpt has failed and this whole thing was the same old hype-y SV pointlessness".

And at that point, we will have reached the trough of disillusionment. I think funding will be less readily available, and we'll start seeing some of the bevy of single-purpose LLM-based products start closing up shop.

But more quietly, others will be (already are) traversing up the slope of enlightenment. As others have mentioned, this is stuff like features in Microsoft's and Google's productivity products (including those for software engineering productivity like Github Copilot), and some subset of products and features elsewhere that turn out to be compelling in a sticky way.

I expect 2024 and 2025 to be the more interesting part of this hype cycle. I don't think we're on the verge of waking up in a world nobody recognizes in a small number of days or months, but I think in a few years we're going to have a bunch of useful tools that we didn't have a year ago, some of which are the obvious ones we've already seen, but improved, and others that are not obvious right now.

Not sure if this was insightful enough for you :) Apologies if not.

Re: What we know about LLMs

#98

Post by iainctduncan »

killernap wrote: post #13

ChatGPT was announced November, 2022 - 8 months ago. Time flies. Question for HN: Where are we in the hype cycle on this? We can run shitty clones slowly on Raspberry Pi's and your phone. The educational implementations demonstrate the basics in under a thousand lines of brisk C. Great. At some point you have to wonder... well, so what? Not one killer app has emerged. I for one am eager to be all hip and open minded…

I work in tech diligence so I look at companies in detail. I have seen a couple where good machine learning is going to make a massive difference (whether it will keep them ahead of everyone is a separate question). I think it really boils down to:

"Is this a problem where an answer that is mostly right and sometimes wrong is still a great value proposition?"

This is what people don't get. If sometimes the answer is (catastrophically) wrong, and the cost of this is high, there's no market fit. So I think a lot of these early LLM related startups are going to be trainwrecks because they haven't figured this out. If the cost of an error is very high in your business, and human checking is what you are trying to avoid, these are not nearly as helpful.

I looked at one company in this scenario and they were dying. Couldn't get big customers to commit because the product was just not worth it if it couldn't be reliably right on something that a human was never going to get wrong (can't say what it was, NDAs and all that.) I also looked at one where they were doing very well because an answer that was usually close would save workers tons of time, and the nature of the biz was that eliminating the human verification step would make no sense anyway. Let's just say it was in a very onerous search problem, and it was trivial for the searcher to say "wrong wrong wrong, RIGHT, phew that saved me hours!". And that saving was going to add up to very significant cash.

So killer apps are going to be out there. But I agree that there is massive overhype and it's not all of them! (or even many!)

Re: What we know about LLMs

#99

Post by swyx »

> Transformers can be generally categorized into one of three categories: “encoder only” (a la BERT); “decoder only” (a la GPT); and having an “encoder-decoder” architecture (a la T5). Although all of these architectures can be rigged for a broad range of tasks (e.g. classification, translation, etc), encoders are thought to be useful for tasks where the entire sequence needs to be understood (such as sentiment classification), whereas decoders are thought to be useful for tasks where text needs to be completed (such as completing a sentence). Encoder-decoder architectures can be applied to a variety of problems, but are most famously associated with language translation.

theres a whole lot of "thought to be"'s here. is there a proper study done on the relative effectiveness of encoder only vs decoder only vs encoder-decoder for various tasks?

Re: What we know about LLMs

#100

Post by llm_nerd »

obiefernandez wrote: post #96
llm_nerd wrote: post #87

Earlier quoted context omitted.

Me - "Little overlap" You - "That's false because I saw a thing about this one guy doing some thing" Okay... "even without taking into account the greed stuff," What "greed stuff"? It is incredibly hard to make money in LLMs/ML. The barriers to entry are colossal, and it is technically extremely difficult. Everyone keeps talking about all the "grifters" (a go to term that usually means the speaker's arguments can be…

I can assure you that the smart folks running daily AI newsletters are currently raking in money hand over fist. It’s awe inspiring.

Such as? Can you give some examples?

There are some very smart people working incredibly hard, yielding some rewards from it. It is a tiny, tiny set.

Post reply on HN