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

141–150 of 173 posts

Re: What we know about LLMs

#141

Post by RC_ITR »

samstave wrote: post #126
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…

This implies that any RLHF is introducing human bias into any "thoughts" the model may have?

Yes, but I think your comment has the foundational misconception that it's the first or even main place where bias is put into models.

LLMs are just pattern identifiers and repeaters. They are trained on inherently biased training datasets of inherently biased text written by inherently biased humans. Every single step of training introduces some amount of bias to an LLM.

Re: What we know about LLMs

#142

Post by jawerty »

hallqv wrote: post #127
jawerty wrote: post #63

Earlier quoted context omitted.

So I'm not doing RLHF that's how LLama is pre-trained. It's in the loss/optimization phase in their training I believe. For the finetuning i'm using LoRA to freeze most of the layers for parameter optimization. Using PEFT from huggingface

RLHF is not part of LLaMa pretraining, or pretraning of any other models for that matter. RLHF comes after pretraining. https://twitter.com/Jeande_d/status/1661833563069620247/phot...

I see, that’s my misunderstanding I was grouping all training as pretraining

Re: What we know about LLMs

#143

Post by loire280 »

Legend2440 wrote: post #132
Jevon23 wrote: post #57

Earlier quoted context omitted.

>Not one killer app has emerged. Surely the “killer app” is ChatGPT itself? ChatGPT has already put some copywriters and journalists out of work, or at least reduced their hours. The app is quite literally “killing” something, i.e. people’s jobs. For those people, it’s not just empty hype. It’s very real. Certainly it’s already more real than anything having to do with blockchain/crypto.

I'm dubious. The few news websites that started publishing LLM articles (CNET, etc) were already circling the drain. They'd probably have fired their journalists anyway because they're on the edge of bankruptcy.

I expect that, over the next few years, companies that need to lay off workers will spin their mismanagement by claiming they are replacing those jobs with "AI".

Re: What we know about LLMs

#144

Post by moffkalast »

killernap wrote: post #44
scottmcf wrote: post #41

Earlier quoted context omitted.

I agree, if it delivers on the kind of demos they showed off here: https://news.microsoft.com/reinventing-productivity/ It's going to be an absolute "killer app".

https://en.wikipedia.org/wiki/BonziBuddy My god. If we hit that bullseye, the rest of the dominoes will fall like a house of cards. Checkmate.

She's built like a clippy, but she handles like a HAL.

Re: What we know about LLMs

#145

Post by moffkalast »

esafak wrote: post #128
jannyfer wrote: post #7

Great summary. I’ve been reading a pop neuroscience book called Incognito (2011). In it, the author talks about how the brain is a group of competing sub-brains of many forms, and the brain might have several ways of doing the same thing (e.g. recognizing an object). The author also posited that the lack of AI progress back then was due to the fact that there are no constantly competing sub-brains. Our brains are alw…

> The author also posited that the lack of AI progress back then was due to the fact that there are no constantly competing sub-brains. That became popular in neural networks after the introduction of dropout regularization, which forced neurons to "co-adapt" and learn to do each others' jobs. Large, over-specified models also provide a natural setting for co-adaptation.

Isn't dropout just there to avoid overfitting? This is more like a mixture of experts type architecture.

Re: What we know about LLMs

#146

Post by bastawhiz »

apsurd wrote: post #125
bastawhiz wrote: post #110

Earlier quoted context omitted.

> Not one killer app has emerged. I'll say that I pretty firmly disagree with this. I've been using Github Copilot for about six months for my own work and it has fundamentally changed how I write code. Ignoring the ethics of Copilot, if I just need to read a file with some data, parse it, and render that data on screen, Copilot just _does_ most of that for me. I write a chunky comment explaining what I want, it writ…

My friend made a great comparison that seems to agree with your take: chatGPT for coding is like when ruby on rails came out. Or wordpress. felt magical and boosted (a certain kind of) productivity through the roof. We don't think of rails as the second coming though. same with code editors. of course a rails for all of code is cool. but iono, it's a code editor. i still use sublime.

I'd maybe make the analogy that it's like the first ORM. Sure, you could write your own DB queries, but it just does what you want, and it's usually right.

Were ORMs the second coming? Meh. But it's arguable that they're still immensely powerful and useful and the way people write apps that interface with an RDBMS is permanently changed forevermore.

Re: What we know about LLMs

#147

Post by blharr »

Arkhaine_kupo wrote: post #37
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…

> Not one killer app has emerged. I think the microsoft gpt integration on Office is probably that app. Ability to ask to have your email's summarised, or getting your excel sheets formulas configured with natural language, etc are increidbly useful tools to lower the floor of entry to tools that already speed up humans so much. I don't think the use of this tools is some life redefining feature, but a friend of mine…

Relevant comic https://marketoonist.com/2023/03/ai-written-ai-read.html

Re: What we know about LLMs

#148

Post by masswerk »

jannyfer wrote: post #7

Great summary. I’ve been reading a pop neuroscience book called Incognito (2011). In it, the author talks about how the brain is a group of competing sub-brains of many forms, and the brain might have several ways of doing the same thing (e.g. recognizing an object). The author also posited that the lack of AI progress back then was due to the fact that there are no constantly competing sub-brains. Our brains are alw…

In fact, this is what psychoanalysis and the notion of the unconscious (as opposed to "subconscious processes") was all about. (And it's also, where the "talking cure" found its leverage.)

Re: What we know about LLMs

#149

Post by esafak »

moffkalast wrote: post #145
esafak wrote: post #128

Earlier quoted context omitted.

> The author also posited that the lack of AI progress back then was due to the fact that there are no constantly competing sub-brains. That became popular in neural networks after the introduction of dropout regularization, which forced neurons to "co-adapt" and learn to do each others' jobs. Large, over-specified models also provide a natural setting for co-adaptation.

Isn't dropout just there to avoid overfitting? This is more like a mixture of experts type architecture.

That is one lens to view it through. Co-adaptation reduction is another, and it is an intuitive one: generalization ability is improved if a neuron has to support multiple contexts instead of relying on other neurons to lift the weight, if you pardon the pun.

Improving neural networks by preventing co-adaptation of feature detectors https://arxiv.org/abs/1207.0580

Re: What we know about LLMs

#150

Post by janalsncm »

> Transformers can read the entire sequence at once and learn to “pay attention to” only the values that came earlier in time (via “masking”)

Unless the text fits into the model’s context window, this is incorrect. The self-attention layers will train via a sliding window over the text.

Learning to attend only to previous tokens is also not correct. There are a lot of ways to train a transformer. BERT is bidirectional for example.

Post reply on HN