3 ms·

> Source? The popular narrative, at least a few months ago was that Chinese AI models like deepseek were not far behind models from OpenAI, hence the stock mark

> Source? The popular narrative, at least a few months ago was that Chinese AI models like deepseek were not far behind models from OpenAI, hence the stock market crash Like I said, I was talking from my personal experience. I was comparing the on-prem models I know with the large, proprietary cloud models. This includes variants of DeepSeek, but not the largest one that they host online. > Is there any evidence this actually caused open weights models to be worse? As much as you hate this "rent seeking", it's mostly a sideshow. If open weights models are competitive, then the proof is in the pudding, no need to fret about the morality platforms "rent seeking" over UGC. AI performance is famously hard to benchmark, so no, I don't have "evidence". But it is not unreasonable to assume some generality of my experiences with OpenAI's, Google's and other cloud models vs on-prem self-hosted coding agents models, that is, that the really large proprietary models perform best. Also, I'm not "fretting about morality", I'm relating to the comment you replied to. The point was a monopolization of monetization for knowledge and skills by AI companies, when the data they build on was created for free by humans contributing to social media, or people paying to use cloud AI models. It's possible for me to think about thinks beyond my control without claiming moral authority. > This makes no sense. You can't take some copyrighted content you scraped (eg. some blog post), relicense it as "CC0", and then redistribute it. You are distorting my point and building a strawman. Yes, people willingly give away their work for free, either a) because they willingly accept CC0 or CC-SA (like with SO): https://stackoverflow.com/help/licensing https://stackoverflow.com/help/licensing Most people who participated in these communities did it for fun, reputation, etc, not anticipating someone monetizing it by interpolating public content using AI or b) because they use Social Media like Reddit and simply don't know or don't care > That's straightforwardly copyright infringement What about a coding agent being able to do something "in its own words" because it sucked up lots of GitHub projects solving this problem, where people spent a lot of time on proper documentation etc > There was a recent court ruling in the US that basically ruled it's fine to train on copyrighted books without permission, but you can't pirate books to obtain them. This hasn't reached the supreme court yet, so it's not final. Wasn't this the same ruling that said it's enough to buy one copy of a book in order to basically sell paraphrased excerpts of said book? Please correct me where I'm wrong, but I have not yet conceived any new information from your comment, and it seems that you distorted some of the previous discussion to make your point, which I'd paraphrase as "current copyright covers AI and there is no threat that AI companies would monetize public content that was put out under licenses meant to prohibit monetizing derivative work"

gruez 1y ago
>I was comparing the on-prem models I know with the large, proprietary cloud models. >This includes variants of DeepSeek, but not the largest one that they host online. AFAIK you can get the same model that they host online, it's just that there's so many distilled/quantized models floating around that the average person using ollama or whatever think they're running the same thing as the one on chat.deepseek.com, but they're not. >AI performance is famously hard to benchmark, so no, I don't have "evidence". But it is not unreasonable to assume some generality of my experiences with OpenAI's, Google's and other cloud models vs on-prem self-hosted coding agents models, that is, that the really large proprietary models perform best. Right, but the claim wasn't open source models were on par, only that they're "competitive", which allows for some amount of subperformance. >Also, I'm not "fretting about morality", I'm relating to the comment you replied to. The point was a monopolization of monetization for knowledge and skills by AI companies, when the data they build on was created for free by humans contributing to social media, or people paying to use cloud AI models. 1. how is it "monopolization", when AI companies are also providing open weights models? 2. why are you giving "AI companies" flak for this, when it's companies like stackexchange and reddit that are instituting such policies? Characterizing them as "AI companies" seems like a stretch. >You are distorting my point and building a strawman. Yes, people willingly give away their work for free, either You're conflating licensing with distribution. Yes, the authors licensed their comments or answers as CC0 or whatever, that doesn't mean companies can't charge for access for it. In fact, the use of permissive licenses explicitly allows for this, in contrast to creative commons non-commercial licenses. >Please correct me where I'm wrong, but I have not yet conceived any new information from your comment, and it seems that you distorted some of the previous discussion to make your point, which I'd paraphrase as "current copyright covers AI and there is no threat that AI companies would monetize public content that was put out under licenses meant to prohibit monetizing derivative work" If you're talking about the legality of training on CC0 comments then there's even less basis to object on AI companies' training on them.
- moritzwarhier 1y ago
  > AFAIK you can get the same model that they host online, it's just that there's so many distilled/quantized models floating around that the average person using ollama or whatever think they're running the same thing as the one on chat.deepseek.com, but they're not. No, I was perfectly aware of this difference when writing my comment. I just don't have the hardware, time, and skill/motivation to do a scientific comparison. OpenAI used to use static weights, but it's pretty clear that they're doing more by now (and no, I don't mean just RAG). How would you go about verifying the weights used by cloud DS are the same as the ones you get when using full local DS with beefy hardware? Running millions of prompts at temp=0? > 1. how is it "monopolization", when AI companies are also providing open weights models? Please refer to the rest of my comment for my response to this. Also, AFAIK, DS was trained on ChatGPT conversations for RILHF? Regarding the licensing of training data, I feel like we are not communicating on a common basis anymore. Do you think that training needs no commercial content licensing, or do you think that most AI companies have commercial licenses (not just a paid copy!) for their training data?

gruez 1y ago

>I was comparing the on-prem models I know with the large, proprietary cloud models. >This includes variants of DeepSeek, but not the largest one that they host online. AFAIK you can get the same model that they host online, it's just that there's so many distilled/quantized models floating around that the average person using ollama or whatever think they're running the same thing as the one on chat.deepseek.com, but they're not. >AI performance is famously hard to benchmark, so no, I don't have "evidence". But it is not unreasonable to assume some generality of my experiences with OpenAI's, Google's and other cloud models vs on-prem self-hosted coding agents models, that is, that the really large proprietary models perform best. Right, but the claim wasn't open source models were on par, only that they're "competitive", which allows for some amount of subperformance. >Also, I'm not "fretting about morality", I'm relating to the comment you replied to. The point was a monopolization of monetization for knowledge and skills by AI companies, when the data they build on was created for free by humans contributing to social media, or people paying to use cloud AI models. 1. how is it "monopolization", when AI companies are also providing open weights models? 2. why are you giving "AI companies" flak for this, when it's companies like stackexchange and reddit that are instituting such policies? Characterizing them as "AI companies" seems like a stretch. >You are distorting my point and building a strawman. Yes, people willingly give away their work for free, either You're conflating licensing with distribution. Yes, the authors licensed their comments or answers as CC0 or whatever, that doesn't mean companies can't charge for access for it. In fact, the use of permissive licenses explicitly allows for this, in contrast to creative commons non-commercial licenses. >Please correct me where I'm wrong, but I have not yet conceived any new information from your comment, and it seems that you distorted some of the previous discussion to make your point, which I'd paraphrase as "current copyright covers AI and there is no threat that AI companies would monetize public content that was put out under licenses meant to prohibit monetizing derivative work" If you're talking about the legality of training on CC0 comments then there's even less basis to object on AI companies' training on them.

moritzwarhier 1y ago
> AFAIK you can get the same model that they host online, it's just that there's so many distilled/quantized models floating around that the average person using ollama or whatever think they're running the same thing as the one on chat.deepseek.com, but they're not. No, I was perfectly aware of this difference when writing my comment. I just don't have the hardware, time, and skill/motivation to do a scientific comparison. OpenAI used to use static weights, but it's pretty clear that they're doing more by now (and no, I don't mean just RAG). How would you go about verifying the weights used by cloud DS are the same as the ones you get when using full local DS with beefy hardware? Running millions of prompts at temp=0? > 1. how is it "monopolization", when AI companies are also providing open weights models? Please refer to the rest of my comment for my response to this. Also, AFAIK, DS was trained on ChatGPT conversations for RILHF? Regarding the licensing of training data, I feel like we are not communicating on a common basis anymore. Do you think that training needs no commercial content licensing, or do you think that most AI companies have commercial licenses (not just a paid copy!) for their training data?