I firmly believe it will slow down significantly. My prediction for the future is that there will be a much bigger focus on a few “base” models that will be tweaked slightly for different roles, rather than “from the ground up” retraining like we see now. The industry is already starting to move in that direction.
While I agree in principle, one thing I’d like to clarify is that TRAINING is super energy intensive, once the network is trained, it’s more or less static. Actually using the network isn’t dramatically more energy than any other indexed database lookup.
When I read the headline I just assumed this must be an onion article. That is not good.
You don’t need to exercise your stock options to access their value. It’s common practice to take loans out against their value, which allows you to access your money effectively tax free by instead paying interest against the loan. This is (again) a fairly commonplace practice used to make collecting tax difficult, and allow them to make the argument to regulators that they aren’t actually being paid that much, it’s totally just options they would never sell off. That’s why C suite has such a “burn everything to the ground, as long as our stock price goes up” mentality, because if it doesn’t, they have to start worrying about interest on their loans— because they have fairly low liquidity (percentage wise).
Base pay is a meaningless figure in tech. Total comp is all that matters. The true high rollers intentionally keep their base pay low to dodge taxes. This doesn’t mean they aren’t still getting and using this money, it’s just an intentional straw man to allow them to make this exact argument.
That would be a dream come true for Youtube. Think about how that would affect them financially, they would be serving significantly less load, but they would still rake in all those Youtube premium subscriptions. It would effectively be campaigning to remove the users who aren’t monetized from the platform.
I would love to switch, and I tried to the other day, but I discovered that Firefox still doesn’t support integrated WebAuthn tokens (I.E. using Touch ID in lieu of YubiKeys). That is (unfortunately) a non-starter for me, as I use that technology everywhere, and I’m not intentionally weakening my security posture to switch. I’m honestly really surprised to find this feature disparity, as this feature has been generally available elsewhere for years. I’m a developer, so maybe I’ll take a crack at implementing it myself sometime, but it’s a big enough deal that I genuinely can’t switch yet :(
I’m out and about today, so apologies if my responses don’t contain the level of detail I’d like; As for the law being collective morality, all sorts of religious prohibitions and moral scares HAVE ended up in the law. The idea is that the “collective” is large enough to dispel any niche restrictive beliefs. Whether or not you agree with that strategy aside, that is how I believe the current system works in an ideal sense (even if it works differently in practice), that’s what it is designed to protect from my perspective.
As for anti-AI artists, let me pose a situation for you to illustrate my perspective. As a prerequisite for this situation, a large part of a lawsuit, and the ability to advocate for a law is based on standing, the idea that you personally, or a group you represent has been directly, tangibly harmed by the thing you are trying to restrict. Here is the situation:
I am a furry, and a LARGE part of the fandom is based on art and artists. A core furry experience is getting art commissioned of your character from other artists. It’s commonplace for all these artists to have a very specific, identifiable signature style, so much so that it is trivial for me and other furs to be able to identify artists by their work alone at just a glance. Many of these artists have shifted to making their living full time off of creating art. With the advent of some new generational models, it is now possible to train a model exclusively off of one singular artists style, and generate art indistinguishable from the real thing without ever contacting them. This puts their livelihood directly at risk, and also muddies the waters in terms of subject matter, and what they support. Without laws regulating training, this could take away their livelihood, or even give a (very convincing, and hard to disprove) impression that they support things they don’t, like making art involving political parties, or illegal activities, which I have seen happen already. This almost approaches defamation in my opinion.
One argument you could make is that this is similar to the invention of photography, which may have directly threatened the work of painters. And while there are some comparisons you could draw from that situation, photography didn’t fundamentally replace their work verbatim, it merely provided an alternative that filled a similar role. This situation is distinct because in many cases, it’s not possible, or at least immediately apparent which pieces are authentic, or not. That is a VERY large problem the law needs to solve as soon as possible.
Further, I believe the same, or similar problems exist in LLMs, like they do in the situation involving generative image models above. Sure with enough training, those issues are lessened in impact, but where is the line of what is ok and what isn’t? Ultimately the models themselves don’t contain any copyrighted content, but they (by design) combine related ideas and patterns found in the training data, in a way that will always approximate it, depending on the depth of training data. While “overfitting” might be considered a negative in the industry, it’s still a possibility, and until there is some sort of regulations establishing the fitness of commercially available LLMs, I can envision situations in which management would cut training short once it’s “good enough”, leaving overfitting issues in place.
Lastly, with respect, I’d like to push back on both the notion that I’d like to ban AI or LLMs, as well as the notion that I’m not educated enough on the subject to adequately debate regulations on it. Both are untrue. I’m very much in favor of developing the technology, and exploring all it’s applications. It’s revolutionary, and worthy of the research attention it’s getting. I work on a variety of models across the AI and LLM space professionally, and I’ve seen how versatile it is; That said, I have also seen how over publicized it is. We’re clearly (from my perspective) in a bubble that will eventually pop. We’re claiming products use AI to do this and that across nearly every industry, and while LLMs in particular are amazing, and can be used in a ton of applications, it’s certainly not all of them— and I’m particularly cautious of putting new models in charge of dangerous or risky processes where they shouldn’t be before we develop adequate metrics, regulation, and guardrails. To summarize my position, I’m very excited to work towards developing them further, but I want to publicly express the notion that it’s not a silver bullet, and we need to develop legal frameworks for protecting people now, rather than later.
The law is (in an ideal world), the reflection of our collective morality. It is supposed to dictate what is “right” and “wrong”. That said— I see too many folks believing that it works the other way too, that what is illegal must be wrong, and what is legal must be ok. This is (decisively) not the case.
In AI terms, I do believe some of the things that LLMs and the companies behind them are doing now may turn out to be illegal under certain interpretations of the law. But further, I think a lot of the things companies are doing to train these models are seen as “immoral” (me included), and that the law should be changed to reflect that.
Sure that may mean that “stuff these companies are doing now is legal”, but that doesn’t mean we don’t have the right to be upset about it. Tons of stuff large corporations have done was fully legal until public outcry forced the government to legislate against it. The first step in many laws being passed is the public demonstrating a vested interest in it. I believe the same is happening here.
I’m aware the model doesn’t literally contain the training data, but for many models and applications, the training data is by nature small enough, and the application is restrictive enough that it is trivial to get even snippets of almost verbatim training data back out.
One of the primary models I work on involves code generation, and in those applications we’ve actually observed verbatim code being output by the model from the training data, even if there’s a fair amount of training data it’s been trained on. This has spurred concerns about license violation on open source code that was trained on.
There’s also the concept of less verbatim, but more “copied” style. Sure making a movie in the style of Wes Anderson is legitimate artistic expression, but what about a graphic designer making a logo in the “style of McDonalds”? The law is intentionally pretty murky in this department, with even some colors being trademarked for certain categories in the states. There’s not a clear line here, and LLMs are well positioned to challenge what we have on the books already. IMO this is not an AI problem, it’s a legal one that AI just happens to exacerbate.
That’s not what’s happening though, they are using that data to train their AI models, which pretty irreparably embeds identifiable aspects of it into their model. The only way to remove that data from the model would be an incredibly costly retrain. It’s not literally embedded verbatim anywhere, but it’s almost as if you took an image of a book. The data is definitely different, but if you read it (i.e. make the right prompts, or enough of them), there’s the potential to get parts of the original data back.
The key to safe AI use is to treat the AI the same as the user. Let them automate tasks on behalf of the user (after confirmation) in their scope. That way no matter how much the model is manipulated, it can only ever perform the same tasks as the user.
I am an American and I use it religiously for the record. Especially for version numbers. Major.minor.year.month.day.hour.minute-commit. It sorts easy, is specific, intuitive, and makes it clear which version you’re using/working on.
Facebook isn’t even recognizable as a company any more, they’ve become entirely overcome with greed, not even bothering to pretend to care about their users. All they’ve done for the past 15 years is copy other platforms, steal as much data as possible to sell to the highest bidder, and retain users by any means necessary, even if that means collapsing governments and allowing people to die. We need to do everything we can to keep them the hell away from us, because they’ve shown they will stop at nothing to make money.
Oh lawd he shrimpin
I’ve heard rumors that they are leveraging LLM’s like ChatGPT to simulate normal user activity, and that’s disregarding the fact that the API cutoff only happened 2 days ago, and many may not have switched yet. I wasn’t able to verify this anywhere, but based on Lemmy numbers alone, I think we have enough to cast reasonable doubt on the platform’s DAU. I think we’ll have a much better idea of how many folks actually stuck around in a few weeks, or even a month.
Good luck finding thousands of new moderators that you don’t have to pay, or even hiring that many employees before hemorrhaging ad revenue.
Indexing and lookups on datasets as big as companies like Google and Amazon are running also take trillions of operations to complete, especially when you take into account the constant reindexing that needs to be done. In some cases, encoding data into a neural network is actually cheaper than storing the data itself. You can see this in practice with gaussian splatting point cloud capture, where they are training networks to guide points in the cloud at runtime, rather than storing the position of trillions of points over time.