News for hackers. Stuff that matters.
Re:Zuckerberg disses closed-source AI competitors as trying ... by fluffery on Friday June 28, 2024 @11:00AM (#40821298)
Isn't his also closed src?
Re:Zuckerberg disses closed-source AI competitors as trying ... by garyfirestorm on Friday June 28, 2024 @11:04AM (#40821342)
seriously? meta llm's are on huggingface and you can download and run them on your machine along with the weights
Re:Zuckerberg disses closed-source AI competitors as trying ... by theptip on Friday June 28, 2024 @11:09AM (#40821407)
Open source != open weights != open training set.

Of course, GP is ignoring that Meta is far more open than its direct competitors (MSFT, GOOG, OpenAI).

Re:Zuckerberg disses closed-source AI competitors as trying ... by throwthrowuknow on Friday June 28, 2024 @11:26AM (#40821619)
I agree that open source or just “open” is the wrong term here although it’s hard to apply it even to a project where the training data and everything else are shared. “Free” or “libre” might be closer since the end product is free to use and modify. It’s still fuzzy though, are weights code or are they an API? Oddly enough they seem to be the embodiment of the interchangeability of code and data perhaps due to neural networks being closer to mathematics than software normally is.
Re:Zuckerberg disses closed-source AI competitors as trying ... by fragmede on Friday June 28, 2024 @01:47PM (#40823054)
it's not fuzzy. weights are binary blobs. we know what binaries are, source code gets compiled into binaries. photoshop is a binary. firmware are binary blobs. that llama3.ggml outputs text when using a wrapper program vs photoshop launches a program directly doesn't make anything fuzzy.
Re:Zuckerberg disses closed-source AI competitors as trying ... by throwthrowuknow on Friday June 28, 2024 @04:10PM (#40824736)
Not really, those weights are the same ones that are there during training getting modified by backprop. You can take them and continue training the model in order to produce different behaviour. You can also modify them directly like with LoRA or steering. You can also make bigger modifications like expanding the layers or removing nodes. It’s not a frozen binary, those numbers aren’t obfuscated or optimized into some special purpose form that prevents or makes modification difficult.
Re:Zuckerberg disses closed-source AI competitors as trying ... by fragmede on Friday June 28, 2024 @04:23PM (#40824882)
And I can edit binaries with binedit. Reverse engineering is a lost art, but reversing binaries with Ida pro and patching binaries is a time honored skill. It's a shame it's usually used to patch out license checks but there have been some truely impressive reverse engineering on video game ROMs for which the source is not available so we're left with modifying the binary. or as we're looking at it, the model.
Re:Zuckerberg disses closed-source AI competitors as trying ... by fragmede on Saturday June 29, 2024 @08:34AM (#40829995)
https://github.com/WerWolv/ImHex (https://news.ycombinator.com/item?id=40828493) is a hex editor specially made for people that do exactly that kind of binary editing.
Re:Zuckerberg disses closed-source AI competitors as trying ... by throwthrowuknow on Saturday June 29, 2024 @07:48PM (#40834130)
But you’re not reversing anything with weights. They aren’t altered at all. There’s no secret human readable code, just training/test data and hyperparameters. Even if they did give you access to the exact data set they used there’s not much chance you could do anything with it without millions of dollars worth of compute time and even then they’ve already trained the model on that data so the best you could do is start from scratch and see if you could get a better result in which case your time is better spent on building your own data set and RLHF process. Like what value do you think you’re going to get anyways?
Re:Zuckerberg disses closed-source AI competitors as trying ... by PunchTornado on Friday June 28, 2024 @11:10AM (#40821420)
just putting weights on hug is not enough for a model to be open. it is like sharing the binaries of a software.

they should put their entire code on github and show how it was trained, architecture and everything else.

also, the most important thing about a model is not it's architecture, but the dataset it was trained on. that is the "it" in a model.

Re:Zuckerberg disses closed-source AI competitors as trying ... by kredd on Friday June 28, 2024 @11:27AM (#40821637)
I'm not in the game, so pardon my ignorance, but is there any company that released the dataset the model was trained on for LLMs? To my understanding, it will lead to never-ending lawsuits due to grey areas as soon as they're released, and no company actually wants to go through that pain.
Re:Zuckerberg disses closed-source AI competitors as trying ... by fragmede on Friday June 28, 2024 @01:31PM (#40822874)
StableDiffusion 1.0 was trained on LAION-5B Dataset, but the code used to train it wasn't released, so even then it doesn't qualify as open source.
Re:Zuckerberg disses closed-source AI competitors as trying ... by jncfhnb on Friday June 28, 2024 @11:14AM (#40821458)
The preferred form for modifying an ML model is the weights.
Re:Zuckerberg disses closed-source AI competitors as trying ... by stale2002 on Friday June 28, 2024 @11:18AM (#40821518)
I am not sure why people keep pretending like this is possible or keep asking it to happen.

You know that distributing all the data would be illegal copyright infringement, right? Also, its not like they could distribute the (petabytes?) worth of data to every rando who asks.

Do you just not want open weight models to exist, and you are using a fake call for it to be more open, so as to get this stuff all shut down, or what?

Because anyone who actually supports open "anything" would understand this, instead of pedantically arguing about the definition of open source and by suggesting things that both aren't possible and would make the release of these models much less likely.

Re:Zuckerberg disses closed-source AI competitors as trying ... by throwthrowuknow on Friday June 28, 2024 @11:40AM (#40821778)
Perhaps a good solution would be to form a trust responsible for storing and distributing training data. It could be held in common by all contributors and allow anyone to access high quality curated data. It would need legal allowances and exemption from copyright which would be difficult to work out but perhaps some mechanism could be created where copyright holders could derive benefit from not seeking to have their works removed, not necessarily monetary. This would at least get data collection out of the shadows and allow companies and researchers to attest that they are using the data only for training and not distributing it otherwise. That could go a long way toward clarifying issues of whether or not a book or article was in a dataset, how often it appeared, how many epochs it was used in, whether it only appeared as referenced or quoted by other text, etc. With that kind of disclosure it would be possible to write tests to ensure that the models won’t simply regurgitate the material without attribution or unprompted and publish those tests to assuage fears.
Re:Zuckerberg disses closed-source AI competitors as trying ... by fragmede on Friday June 28, 2024 @01:41PM (#40822990)
Anyone who understands the difference between model available and open source would realize they're different things, and then also be smart enough to realize that randos arguing about the definition isn't what's going to sway someone as to wether they'd release the model or the source used to create them. You think Mark Zuckerberg, or anyone else with the $100s of million dollars to spend on training is reading these comments, seeing these dorks argue about if giving away models is open source or not, and then saying, I was going to release Llama4, but because PunchTornado said pedantic things, y'know what? I'm not gonna.

We want words to mean things. model available != open source, If you go to a restaurant and order the beef and you get the fish, what's wrong, it's still food, just eat it. Don't be pedantic, otherwise the cook isn't going to make food anymore. If you're 10 and your parents made you dinner that might pass, but if you're an adult and go to a restaurant where you're paying with your own hard earned money, that's not how it works.

They don't have to distribute the data directly, Stability AI has said they used the LAION-5B Dataset. Just like I could say I read Harry Potter without giving you a copy, and can say that it says "the boy who lived" without violating copyright.

Re:Zuckerberg disses closed-source AI competitors as trying ... by stale2002 on Friday June 28, 2024 @04:05PM (#40824693)
> I was going to release Llama4, but because PunchTornado said pedantic things, y'know what? I'm not gonna.

Its not because he said things.

Instead it is that (if implemented) the things that he is advocating for would actually significantly reduce the amount of open-whatever-you-want-to-call-it models available as most of it would all get shut down.

If these people don't want these models to exist and they are just using this concern trolling to get rid of all these AI models, then I wish they would just say so openly.

Because that would be the result of their positions. It would be a destruction of most of the more-open AI models that we have available today and an end to the open-whatever AI industry.

If thats what you want, just say it, and then we can talk about the pros and cons of a completely centralized AI ecosystem that is only controlled by OpenAI, Google, and the government and how if the destruction of the AI startup space that is reliant on the models given out by meta/ect is good or bad.

Re:Zuckerberg disses closed-source AI competitors as trying ... by fragmede on Friday June 28, 2024 @04:31PM (#40824964)
it's not concern trolling, it's that the words used don't describe what it is. what Mark Zuckerberg has done with releasing llama3 and making the model available is frickin awesome, just don't call it open source! it's model available.

Trying to paint me into a corner with a false dichotomy, that if I want it to be called model available instead of open source, and that in doing so, I want a dystopian future where megacorps have AGI and we're all at their mercy isn't going to work. I don't want the models to not be available, what I want is to see what's being used to create the models, the source, if you will, and thus the alignment side of the source, and why it won't answer "how do I make cocaine", as if I'm going to go off and somehow become Tony Montoya. We may never get that, but that's what calling it open source and have it mean that would entail. Making the models available to the plebs is a generous thing for Mark to do, but that doesn't mean we can't ask for more. Or maybe we shouldn't; and just be grateful for his benevolence, because it's not like OpenAI has done the same with ChatGPT. I will give them credit for Whisper though.