The source is available to view and use under a non-commercial license. It's extremely frustrating that major AI players are trying to shoehorn a redefinition of a term that's well-understood, just to further their reputation.
There's a lot of people arguing about OSS vs FOSS. opensource.org has a specific definition but this is not universally accepted (there's a lot on this in the wiki page btw). Stallman even wrote about this[0].
If you're passionately arguing and downvoting others that are claiming that ImageBind is in fact OSS then you're also actively demonstrating naivety. Definitions are not always universal and this is an easy to verify example. Instead of just posting another link to opensource.org read their about section[1] and consider why that leads to non-universality. And maybe also consider the words themselves: is the source available? Even if the only correct definition were from opensource.org, I'd expect everyone here to have some understanding why others might be confused since the words imply something different. There's some evidence to the other side's view, so it isn't obscene or idiotic. You don't have to accept the other view, just consider its existence.
Take a step back and consider that your view isn't the only possible one. Maybe we can have more constructive discussions if we do so. We don't need these holy wars, they don't accomplish anything.
It’s not about what views are possible. It’s about honesty and humility which are both things the modern world doesn’t seem to care about. The more narrow definition of oss (rms/gnu) is un ambiguous as is the meaning of source available. People choosing to use oss with the ambiguous meaning are liable for their ambiguity, and therefore worthy of op’s criticism.
>> Why do people misunderstand it that way? Because that is the natural meaning of the words “open source.” But the concept for which the open source advocates sought another name was a variant of that of free software.
>> Since the obvious meaning for “open source” is not the meaning that its advocates intend, the result is that most people misunderstand the term.
> It’s about honesty and humility
Good faith arguments recognize this obvious meaning. I would not buy it if anyone claimed that any other definition was an obvious conclusion from the term "Open Source Software." But as you're talking about the modern world, I find that this kind of terminology and reaction is quite common. I'm not sure why, but it we see it in a lot of political sayings and these just lead to arguments. You're right, we need to have a little honest and humility and recognize that if you need a paragraph to explain a 2-5 word slogan that it isn't an obvious interpretation.
(I've advocated for abandoning some terms in favor of others and the pushback is not on proposed alternative terminology but rather someone giving me a lecture on the original phrasing and how I don't understand it. That's unproductive, as we see similar events in other comments below).
I see your point. And I understand that the term is understood differently by different people. But there is a substantial fraction of people (myself included) who understands it the way OSI defines it.
My criticism is not that Facebook chooses to release the software under any particular terms (they are well within their rights not to release it at all). But the ambiguity of the term remains, and my belief is that Facebook is well aware of the ambiguity and nevertheless chooses to use it in bad faith, and that is what frustrates me.
A less-ambiguous press release might instead say "source released under a non-commercial research license" or more succinctly "model code available for researchers," which sounds just fine.
All I'm trying to get is people to change their phrasing to something like
> Many may not be aware, because of the naming, but OSS is actually defined as having no commerce restrictions <link to OSD and DoD definitions>. Facebook should clarify that this is "source available". I'm glad the source is available, but this is extremely frustrating as we've had this discussion several times and I believe that Facebook should know better by now.
I think such a phrasing could have led to better outcomes. For example, if I saw that -- instead of a bunch of comments about OSS -- I would have commented about how there is a lack of a training script and that this makes the work difficult to reproduce. While the training looks standard sometimes there are subtle differences that make this difficult. Though I'll admit that Facebook has a decent track record for reproduction within my own experience (MSR is the worst and Google is okay but uses proprietary datasets[0]).
> or more succinctly "model code available for researchers,"
Just to be clear, the checkpoint is also available. But I should also mention that this is only the state dict and does not have other information, as is common practice.
[0] Google needs to be not allowed to use this dataset in conference publications (at least in review phase). Not only does it make the work not reproducible but it breaks double blind reviewing, thus violating conference guidelines. Any paper which mentions JFT is automatically known as a Google work and biases the reviewer. I am ashamed that both reviewers and ACs have not flagged these works.
Some argue we should acknowledge the important historical contribution made by the Linux project to the most widely used operating system in the world, but others argue that Linux is only a small component in a larger system, especially when they count lines of code, so it does not merit mention.
/s
(Above is a paraphrase of the now old argument why GNU did or did not merit mention in names of Linux-based operating systems. Now we have reached the point where Linux doesn't get a mention either).
Definition or not this is much better than nothing. Personally I'm interested in how did they do it. They have resources to experiment and try different ways. And they share the ideas. Big thanks for that.
This is absolutely open source and will enable a lot of academic, governmental, and nonprofit uses. Just because it doesn't allow someone else to make money from it doesn't make it not open source.
> The license must not restrict anyone from making use of the program in a specific field of endeavor. For example, it may not restrict the program from being used in a business, or from being used for genetic research.
It's true that people are starting to make a distinction between "free software" and "open source," but generally the distinction is smaller than what you're saying.
The broader class of programs whose code is available for viewing and (potential) use or modification subject to restrictions, has been called "source available," and I think that's a great description of what it is and sets expectations well. You may or may not have a license to use it, but you can see the source code.
For example, the Business Source License is generally not considered open-source. It places restrictions on what you can do.
I kind of hate being pedantic about this, but I also dislike the deliberate misuse of the term, especially by a company that is obviously quite familiar with publishing large amounts of high-quality software that's genuinely open source.
> Source-available software is software released through a source code distribution model that includes arrangements where the source can be viewed, and in some cases modified, but without necessarily meeting the criteria to be called open-source.
> Given both
the competitive landscape and the safety implications of large-scale models like GPT-4, this report
contains no further details about the architecture (including model size), hardware, training compute,
dataset construction, training method, or similar.
Totally, I think open ai believes they will build an AGI and rule the world with importance before everyone realised they just steal and compress everyone’s stuff and sell it via hosted database.
Ha ha they're trying to revive the metaverse with the one killer app that'll save it: immersive AI porn. Probably a good time to buy some meta shares in case they "pull it off"
I see this sentiment of “the metaverse is dead” very often on HN but VR usage is at an all time high and meta has the leading hardware. Genuinely wonder: Why do you think the metaverse is dead, and by metaverse do you mean “horizon worlds”, VR, “metaverse apps” or something else?
I’ve seen many, many people mistake Horizon worlds for the metaverse. I guess it’s Meta’s fault for not being more clear, but I doubt those deriding the tech have actually seen anything directly from them anyways.
Because Meta have had to drastically reduce the amount they're investing in their metaverse because they didn't have the user figures to reassure investors it wasn't just a waste of billions of dollars.
I dunno, everyone pivoted to AI because it’s clearly a massive force multiplier but they’re all bringing those gains back to whatever they were working on previously. I’ve already seen some great uses of AI generated voices, scripts, textures, and sky spheres in VR. AI is also a huge resource for learning how to do things you’ve never attempted before. Shit’s going to be popping off all over the place this year and I don’t think it’ll stop.
Are multi-format embeddings — like this one, or simpler ones like text-to-image — inherently bijective? I.e. could a model like this be configured to use any of its trained sense data types as either an input or output? Or is there some bias in the way the training data is ingested that makes the model better at transforming in certain directions?
I haven't yet read this paper, but other multimodal models have different layers for mapping the input/output of each modality to the common embedding space. So, to simplify, there is some high-dimensional coordinate that the audio sample for the morphemes for "frog," the text word "frog" and maybe the same in multiple languages like, "rana" - they all basically have the same point in the abstract feature space. The additional layers of the model then provide "translation"
Maybe, I take anything LeCun says with many grains of salt. Much like Elon Musk, any time he says something that appears to make sense I find myself asking "what don't I know enough about, to know why he's wrong?"
In the quest to placate shareholders, one of the failing FAANGs will make a panicked announcement they have partnered with Boston Dynamics, to ensure AI is democratized and equitable for all (technically true when we're all on the run).
For training, it probably looks very different, but the usage example in https://github.com/facebookresearch/ImageBind#usage seems to use just one cuda device, so you might be able to run it on a consumer GPU
https://github.com/facebookresearch/ImageBind/blob/main/LICE...
The source is available to view and use under a non-commercial license. It's extremely frustrating that major AI players are trying to shoehorn a redefinition of a term that's well-understood, just to further their reputation.