News for hackers. Stuff that matters.
Re:Update on Llama adoption by codingwagie on Thursday August 29, 2024 @11:15AM (#41391659)
Probably will get flagged, but I get so annoyed by the cynical takes on Meta and their open source strategy. Meta is the only company releasing true open source (React, pytorch, graphql) and now LLama. This company has done more for software development than any other in the last decade. And now they are burning down the competition in AI, making it accessible to all. Meta software engineering compensation strategy pushed up the high end of developer compensation by almost twice. Enough with the weird cynicism on their licensing policy.
Re:Update on Llama adoption by SushiHippie on Thursday August 29, 2024 @11:22AM (#41391750)
The llama models use a non open source license [0].

Yes it is still better than not being able to access the weights at all, but calling these weights open source is not correct.

[0] https://huggingface.co/meta-llama/Meta-Llama-3.1-8B/blob/mai...

Re:Update on Llama adoption by codingwagie on Thursday August 29, 2024 @11:59AM (#41392186)
Dude they spent billions on the model and then just open sourced it
Re:Update on Llama adoption by koolala on Thursday August 29, 2024 @12:18PM (#41392411)
Look at Apple spending a billion on ads to say they respect your privacy or the Earth. Meta is buying / licensing a market sector in an industry they dominate where they have full control of our data. Our data is what got them that billion dollars.
Re:Update on Llama adoption by martindevans on Thursday August 29, 2024 @12:04PM (#41392240)
No, they spent billions on a model and released the weights, and that's fantastic! It's not not open source though.
Re:Update on Llama adoption by koolala on Thursday August 29, 2024 @12:14PM (#41392366)
True? The cynicism is arguments over "true". If you trick yourself into believing this is what Open Source looks like (no source data) then you lose out on imagining what a real open source AI with open source data would be like.
Re:Update on Llama adoption by JumpCrisscross on Thursday August 29, 2024 @12:17PM (#41392393)
> you lose out on imagining what a real open source AI with open source data would be like

Zero privacy?

Re:Update on Llama adoption by koolala on Thursday August 29, 2024 @01:08PM (#41393004)
Sounds like God. What is private? What is public? There is a line between an omnipresent God that sees everything (zero privacy) and an Open Source AI that collects and indexes all public information. I want an AI to be programmed / trained on our conversation here. (if your reading this God / FB / OpenAI / 2025 ai Source Dataset <3)
Re:Update on Llama adoption by YetAnotherNick on Thursday August 29, 2024 @12:35PM (#41392636)
Open source and open source model is not a term that came from dictionary, and it is just based on what community thinks it means. As long as open source model doesn't cause confusion, which it does not as open source model today just means open weights model, fighting over it is not worth it.
Re:Update on Llama adoption by koolala on Thursday August 29, 2024 @01:05PM (#41392979)
If open source model means open weight model then open source model means nothing.

I want it to mean something!

Re:Update on Llama adoption by YetAnotherNick on Friday August 30, 2024 @05:27AM (#41399062)
> If open source model means open weight model then open source model means nothing.

They are synonym and mean the same thing.

Re:Update on Llama adoption by bschmidt1 on Thursday August 29, 2024 @11:32AM (#41391865)
React? Surely I'm not the only one who remembers https://news.ycombinator.com/item?id=15050841

I don't think Facebook/Meta is the beacon of open-source goodness you think it is. The main reason they created yarn instead of iterating on npm is to use their own patent-friendly license they wanted to use with React (before the community flipped out and demanded they re-license it as MIT). Early Vue adoption seemed mostly driven by that React licensing fiasco.

Re:Update on Llama adoption by codingwagie on Thursday August 29, 2024 @11:59AM (#41392190)
React runs modern web development, get real
Re:Update on Llama adoption by bschmidt1 on Thursday August 29, 2024 @12:22PM (#41392461)
Get real? How about get the point? It would not have (could not have) been so widely adopted without being MIT licensed and open source. It was huge news when Facebook attempted to change their licensing, and they issued an apology then made it MIT when the community responded. https://meitar.com/wp-content/uploads/2021/08/facebook_react...
Re:Update on Llama adoption by cmur on Thursday August 29, 2024 @11:24AM (#41391772)
if something requires an EULA it isn’t open at all, it is just publicly available. By your logic, public services are “open source.” There are myriad corporations that release actual open source software that is truly free to use. If you experience massive success with anything regarding Meta’s LLMs, they’re going to take a cut according to their EULA.
Re:Update on Llama adoption by bunderbunder on Thursday August 29, 2024 @11:36AM (#41391907)
I'm trying to figure out the logic that makes "free for commercial use with less than 700 million monthly active users" less open than "free for non-commercial use", which is the traditional norm for non-copyleft open source machine learning products. But I just can't get there. Could somebody spell it out for me?
Re:Update on Llama adoption by koolala on Thursday August 29, 2024 @12:20PM (#41392441)
Ideals vs. Gut Instinct
Re:Update on Llama adoption by eduction on Thursday August 29, 2024 @01:03PM (#41392967)
You’re certainly entitled to the opinion that an agreement (as in EULA) is distinct from a license (as in GPL, MIT etc).

But many legal minds close to this issue have moved to the position that there is no meaningful distinction, at least when it comes to licenses like GPL.

For example: https://writing.kemitchell.com/2023/10/13/Wrong-About-GPLs

Re:Update on Llama adoption by wilsonnb3 on Thursday August 29, 2024 @11:20AM (#41391726)
> Meta is the only company releasing true open source

What? There are so many open source projects from huge companies these days.

VSCode, .NET, typescript from MS

Angular, flutter, kubernetes, go, android, chromium from Google

Re:Update on Llama adoption by gwern on Thursday August 29, 2024 @11:40AM (#41391978)
There is nothing weirdly cynical about it. This is a fact of life in Silicon Valley - that a lot of FLOSS is released for strategic reasons (such as building up a community before enclosing it to extract a profit), and not because the Grinch's heart grew 2 sizes one day. "Commoditize your complement": https://gwern.net/complement

You can benefit a lot from it, and I have... but do be sure you know what you are ferrying on your back before you decide to offer it a ride across the river.

Re:Update on Llama adoption by xgb84j on Thursday August 29, 2024 @12:18PM (#41392413)
How do you think Meta profits off React and PyTorch? Just marketing to get good candidates?
Re:Update on Llama adoption by gwern on Thursday August 29, 2024 @03:16PM (#41394337)
No. I think they use that web & ML software to help run their $1.3 trillion marketcap online social network company, on which, I am given to understand from US Congressional hearings, they sell ads.
Re:Update on Llama adoption by Der_Einzige on Thursday August 29, 2024 @12:41PM (#41392697)
The same way that the US benefits from being the reserve currency of the world. Control of the ecosystem allows meta to define the rules of the game.

Also it's bad when HN is downvoting fking GWERN

Re:Update on Llama adoption by talldayo on Thursday August 29, 2024 @12:18PM (#41392416)
> a lot of FLOSS is released for strategic reasons (such as building up a community before enclosing it

Not only is "a lot" of FOSS not released like this, both free software and Meta's models cannot be monetized post-release. If Meta decides to charge money for Llama4, then everyone with access to the prior models can keep their access and even finetune/redistribute their model. There is no strategic flip Meta can attempt here without shotgunning their own foot off.

Re:Update on Llama adoption by gwern on Thursday August 29, 2024 @03:15PM (#41394324)
It absolutely is released like that. Please note that 'such as' does not mean 'all', 'a majority', or anything like that. It simply means 'one example'. (I eat many things 'such as' honey, yet that doesn't make me Winnie the Pooh.) And enclosures come up regularly here on HN where suddenly a company has embarked on a new incredible journey involving 'not releasing all the source code anymore under an OSI-approved license', so they are not even that exotic an outcome.
Re:Update on Llama adoption by gwern on Friday August 30, 2024 @02:06PM (#41403128)
Today, for example, on Twitter, people are pointing out that Runway, which made a big deal out of how they were releasing DL models and not keeping them proprietary, has quietly deleted all its stuff on HuggingFace. (Much like how, say, Mistral decided that it was no longer releasing all its models, after presenting itself as the big open model champion opposed to OA.) Only a five-year-old child is either surprised or angered by this turn of events.
Re:Update on Llama adoption by talldayo on Saturday August 31, 2024 @05:36PM (#41412189)
Twitter is the epitome of people who "don't get it". For one, you don't use Twitter anymore because the damn website doesn't exist. Your new website, 'X' is filled with boiling frogs that (somehow) expect things to get better. People that don't get the proverbial memo when a neo-Nazi buys their website and replaces it's logo with a half-swastika. It would not be exaggeration to say that the quality and prestige of 'X' chatter is about as worthwhile as 4chan's resident opinion.

For two - they can't do shit. Every one of those businesses can try to pivot into being the next OpenAI, but even OpenAI can't be fucked to turn a profit. They can't retract their models (remember Llama 1.0? hahahahahahahahaha) and they can't piss and moan to the authorities when their weights leak because 99% of them contain unlicensed copyrighted material in the first place.

The real twist to this Gordian knot? Open source is the only model you can release conventional AI under, because anything else precludes a Fair Use defense. Everyone that has "proprietary" models only manages it by playing keep-away with the community.

> Only a five-year-old child is either surprised or angered by this turn of events.

Five year old children know that they can keep their toys once they stop selling them at Wal-mart. Pulling a model off HF doesn't stop new finetunes or even redistribution and re-quantization of the old model. Mistral and Runway can do whatever they want, all their models are belong to us.