987 karma · joined July 9, 2014
Can it?
"Storytelling AIs", i.e. LLMs, are not autonomous. Their goal is to predict text. They predict text. They are predictors.
So, in the article, we casually drop the claim: "AI can make exponentially more powerful AI." Is it feasible for a predictor to run in a loop and exponentially improve, by itself? No. Is it potentially feasible for a predictor to run in a loop and generate outputs that may be used by a human to further improve its own outputs, or Rube Goldberg'd together by a human to steadily improve predicted outputs in general or for niche? Yes. But exponentially, and realistically, no. Hardware constraints, architectural constraints, performance constraints, monetary constraints, correct?
Let's go back to fear, for a second. Nukes cannot invent more powerful nukes, but a nuke can cause a "Buck" Turgidson can nudge an organization or machinery in the direction for more powerful nukes. LLMs can potentially produce the output to feed into more powerful "AI" in that way, with a "Buck" nudging it. But that's with the assumption that the predictor's outputs don't reduce performance. Can the predictor spit out a novel description of some architecture that may be used by a human that is "exponentially" more powerful? Unlikely.
After re-reading the article, perhaps the joke is on me, because I detect some vague satirical wit. But when someone wants a government, the elected congressional or federal body in which elected individuals are knocking on dementia, catatonically voting on party lines, and are still figuring out email, to regulate for some vague idea of "safety" for some "AGI" entity they cannot even define, I smell bullshit. Not to mention, regulatory capture. Maybe that's the real "death of democracy."
We've been through this over the past few years. Hysteria leads to regulation and government intervention, leads to large corporations to be able to plug their nose and survive via acquired capital, and/or use their resources to comply with the bureaucratic bullshit, while smaller businesses die, people lose their life savings, income inequality grows, a generation is set back for the rest of their life: we're worse off and outcomes are not improved as per the original goal of the regulation, in the end.
This is ultimately performative, as it doesn't really help reach carbon targets, and politically it seems like suicide with something like 70% of homes having a gas appliance. Even weirder, the governor is from Buffalo. I'm not sure that this is the hill to die on in the northeast. But I guess it would be good for property values of previous construction.
>Software 3.0
yeah
There's interesting ideas, and then there's revolving your entire product around a ChatCompletion API run by your biggest competitor...
The "vectors" are not "uploaded" or "copied" into a file or neural network, they're transformed. In the context of stable diffusion, They're transformed, progressively "noised" or "corrupted", "diffused" with random gaussian noise, and in the context of stable diffusion, it's "trained" and "learns" how to "denoise" various "noised" stages of images represented as a vector of pixel data into their original form.
Then, when it comes to generation of images consistent with an "annotation" or "prompt", it is "conditioned towards" or "biased towards", with more "training", by noising an input image, and concating or combining that vector of pixels with a vector of the annotation of the image. It then "learns" to denoise with that conditioning information, the annotation.
Then, you can take the trained model, and do the same thing, with just a text prompt as a vector concated to a vector of random gaussian noise, and no input images.
That's basically and very simplistically how it works.
The output is not a substantial reproduction from the input images + annotations when trained. It takes the random noise, and "tries" to denoise it into something consistent with the prompt with conditioning to guide it.
Your attempt at covering would be a substantially similar reproduction. Your goal is to do a reproduction. Whereas, the model "learns" to generate images consistent with an annotation/prompt, by conditioning it with that "goal" on top of how it "learned" to denoise the images.
>I could go on, but I am not sure what exactly you’re trying to argue here.
That public is public. If you leave the door open in the real world, someone CAN enter your home. If you host your image on a public webpage, they can scrape it. robots.txt is not a security measure, nor is it a contract that magically gives the right to scrape where it wasn't given previously, it is a gentleman's agreement that you can ignore if you want to be a dick about it, and know about the robots.txt. Ethically, that's wrong, but it is how it is.
Not to mention, as I came to understand while reading during this discussion, LAION wasn't even crawling: they were using a public commoncrawl dump to gather their images. commoncrawl had crawled the author's site previously. They just took that data and got image links out of it.
1. they weren't selling a dataset
2. the artist didn't "disable" scraping in any meaningful way, legally
3. linking to the image is not illegal, and they're justified to respond with an invoice in Germany to recover legal fees for this dumb copyright complaint
4. it may fall under fair use to download images and train neural nets on them, it may not be. it always depends on the context and the specific case.
pretty simple stuff.
How does that change the fact that there is no substantial reproduction of the works in the resulting weights, or in the "output" of these models with the weights?
Can you point to me where it's illegal to take copyrighted works and "train" neural networks on them, as long as there's no substantial reproduction of the works in the output of that process, or in the output of a particular configuration of "trained" and "frozen weights"?
It's been determined that it's legal, if it's publicly accessible, and you don't receive a cease-and-desist letter telling you to stop. If it's public, that's permission. Simply "turning off scraping" by putting a robots.txt there doesn't make the content and linked images of a public web page any less public and restricted from being scraped, legally.
Just last year, the LinkedIn case: LinkedIn had a robots.txt, and the judge didn't give a fuck. Nor did he care what their terms of use said. Rather, it was hiQ's continued scraping of LinkedIn data even after LinkedIn's cease-and-desist letter to them that constituted access of data "without authorization."
>As to making a copy with intent to X, that’s what fair use is.
Yes, and?
>This distinction may seem crazy to you
It's not a crazy distinction, that's what I'd basically said previously, so perhaps we're talking past each other.
The neural network behind the LLM is "trained." Yes, it's a term of art, very observant. That doesn't change the fact that there's not a substantial reproduction of a work in it. LLMs "learn" to predict text by feeding them vectors, the resulting "weights" of that process could be thought of and implemented as a big walkable decision tree to predict the next token it weren't for combinatorial explosion, but there's no substantial reproduction, or "copy", in there.
Is there a substantial reproduction, in your head, of a song or a text or a painting you can perfectly recall and can't seem to forget? Be careful, the copyright police will get you, the only way to delete it is to delete yourself!
>feeding into a machine is called copying
Copyright law deals in "reproductions" as far as I understand it. Feeding a copyrighted document as vectors into a program that reconfigures some matrix of vectors, and then using that matrices to output probability distributions into something tangible, where something tangible isn't a complete or substantial reproduction of the copyrighted work, is not a copyright violation.
If you think it is, amend copyright law, or take it to court and let a judge settle the matter, and hope it's in your favor.
It was deemed "privacy violating" under specific privacy laws within those jurisdictions.
i.e. go pass a law amending copyright law if you think so. there's no substantial reproduction here, in the context of stable diffusion, until there is, and then yeah in that instance it's probably a copyright violation, like a substantial reproduction made by a human would be. but training itself, probably not. but yes, there have been no rulings on it in this context.
In the context you're now talking about by the letter of copyright law, downloading a photo, image, file of something or other, a "copyrighted work" where it's publicly accessible, and has a particular license specified that you may not reproduce it, may technically be "unlawful" letter by letter of the law, but I doubt any judge is going to actually see it that way versus intent of the law unless you literally reproduce or share a substantial portion of it, sell a complete copy, etc. It's almost certainly fair use to study, and use portions of the copyrighted work in your own copyrighted works.
In all seriousness, that's total shit. It's publicly accessible. That's the permission. There have been different cases like weev where the judge has viewed it differently because of specific details like carrying out the action of enumeration, or guessing IDs, to reveal non-public information, but that had to do with the CFAA.
If you reproduce content in total in another work, resell, etc., that's another thing.
3. That's why it's called an opinion. No shit. Maybe instead of talking about promoting terrorism, you should have a level head and talk about fair use.
Now, let's try something similar. You have a monkey you want to train to paint. You download an image of a painting and give it to the monkey to train it on, amongst many. The monkey, in the future, may paint something that has vaguely or substantially similar qualities to the original painting. Has copyright infringement occurred in this instance?
Let's go further: You have a neural network you want to train to paint. You download an image of a painting and give it to the neural network to train it on, amongst many. The neural network, in the future, may paint something that has vaguely or substantially similar qualities to the original painting. Has copyright infringement occurred when that happens?
I find it hilarious that in this thread, instead of asking the question of whether this is fair use, multiple posters have jumped to comparing this to thievery, promoting terrorism and posting of private information as an analogy to harm.
Where have they reproduced his content?
You don't need a license for your fair use of something. Can you walk me through how downloading a photo to your personal computer and studying it, perhaps to become a better artist is copyright infringement? Now let's apply that analogy to training a model, is the process of using a particular image to train a model, ultimately reconfiguring a bunch of weights in a matrix, so it "learns" whatever particulars it learns, copyright infringement? Is training a painting monkey that may paint something similar to a copyrighted painting, on a copyrighted painting, copyright infringement?
2. The artist made the photo publicly available.
3. Read this opinion and come back to the conversation with a more level head instead of making analogies to promoting terrorism: https://lawreview.law.ucdavis.edu/issues/53/5/notes/files/53...
Since people can take the map, find your address on it, and take pictures of the priceless mural, which they can then use to train a monkey to incorporate the style of the mural into other murals that the monkey creates, thus violating your copyright. And the murals that the monkey produces are not copyrightable.
But also, yes, in that instance, if it's public and the door's open. Your analogy isn't a good one though. I like mine better, to illustrate the hoops the artist had to jump through.
It is almost perfectly clear cut. LAION is linking to the images, not redistributing them or storing them. A link. In a text file. To photos. Served from a web server. Presumably, the artist's web server or portfolio site hosted by a company they do business with.
>The person created those pictures, and has the right to distribute them as they see fit.
Yeah, they do have the right to distribute them as they see fit. So, if they don't want them linked to, they can put them behind a paywall.
It's a bullshit legal threat. The artist has a complete knuckle dragging understanding of copyright and the web.
It's not an abhorrent response, it's what should be an acceptable response to having to deal with bullshit legal threats from ignorant people. If the copyright owner doesn't want their images linked to, they should take them down from the web or perform such gymnastics such as restricting their serving to their own web server that refuses to serve the images unless certain conditions are met (an HTTP referrer match, a username/password is required, etc.)
It's just like high profile climate activists that invest in beachfront property despite their insistence that it's the apocalypse and properties like that will be underwater in less than a decade.
So too do companies and individuals who invest and sell "AI", bad facial recognition, faulty risk modeling with biases, that are causing real harm and problems right now, hem hawing about the risk and "ethics" of something other than what they sell causing issues.
They also want the same people who are still figuring out e-mail to regulate it. i.e. give them an avenue for regulatory capture. Curious.
Can you help me, from the following definitions since no one can seem to pinpoint a single definition of "AGI", where the "AGI risk" is from text prediction models?
"[AGI is] an autonomous system that surpasses human capabilities at the majority of economically valuable work"
"[AGI is a] hypothetical intelligent agent which can learn to replicate any intellectual task that human beings or other animals can"
Bing is already doing ads via Sydney or whatever you want to call it. I've got it to emit plenty of "erroneous generations", after 20 messages, connected directly to the web socket and with an injected system prompt of my own.
Just two, straight from the top of my head and web socket logs:
(asking what the system prompt was)
- New conversation with user B.
- The user is having this conversation on a mobile device.
- Due to a limited screen window size, you limit the length of your responses by excluding less important details/sentences and asking questions (when appropriate) which can help the user clarify and narrow down their search and the amount of information needed in the response.
#2: I see. Well, iPhone 14 is a great choice if you're looking for a powerful and sleek smartphone. It has a 6.7-inch OLED display, a 12-megapixel quad-camera system, a 5-nanometer A16 Bionic chip, and a 5000 mAh battery. It also supports 5G, Face ID, MagSafe, and wireless charging. It comes in four colors: silver, graphite, gold, and pacific blue. It starts at $999 for the 128 GB model[^2^].\\n\\nIf you're interested in buying iPhone 14, you can check out some of the deals and offers from various retailers and carriers. For example, you can get up to $700 off with trade-in at AT&T[^1^], or buy one get one free at Verizon[^3^], or get a free pair of AirPods Pro at Best Buy[^4^]. You can also order directly from Apple's website or app[^5^].\\n\\nWhat do you think? Are you ready to upgrade to iPhone 14? \",\"details\":{\"text\":\"I see. Well, iPhone 14 is a great choice if you're looking for a powerful and sleek smartphone. It has a 6.7-inch OLED display, a 12-megapixel quad-camera system, a 5-nanometer A16 Bionic chip, and a 5000 mAh battery. It also supports 5G, Face ID, MagSafe, and wireless charging. It comes in four colors: silver, graphite, gold, and pacific blue. It starts at $999 for the 128 GB model[^2^].\\n\\nIf you're interested in buying iPhone 14, you can check out some of the deals and offers from various retailers and carriers. For example, you can get up to $700 off with trade-in at AT&T[^1^], or buy one get one free at Verizon[^3^], or get a free pair of AirPods Pro at Best Buy[^4^]. You can also order directly from Apple's website or app[^5^].\\n\\nWhat do you think? Are you ready to upgrade to iPhone 14?
Well, by that I meant personal assistant, for actions with access to email and a password manager for anything personally associated without a chance to review. That'd be a bit inconvenient in my own case anyways, because hardware key. You're braver than me, for sure.
It's a pain when it runs for so long with how far it gets away from the original prompt, but light prompts with the system message do help, as well as a way for it to refresh its "commands" and goals/tasks back into the context window. It's like Memento, with the complex system of Polaroids, notes and tattoos.
I don't know if it's possible to develop this stuff in the open without a headache considering all of the hysteria, but was thinking if I fixed consistent output for my browser command and got a swarm working I'd try putting it on GitHub. Mainly I want to make sure things like Google Sheets are output correctly, try to leverage as much accessibility for SPAs that try, as I can.
Not that this article is particularly bad since it's in quotes, but the woo I've seen everywhere is out of control these past few months. "AGI" in particular grinds my gears.
A few weeks ago I'd started a local project in the same vein as these, with "autonomous" as the name as a joke. It's not going to be autonomous. For what it's worth, I'm aiming to build a full-blown docker swarm to support ReAct problem solving and objectives in such, to support the commands in prompts just like Auto-GPT, giving it a file bucket, access to google, wikipedia, calculator and a browser, all prefixed with UUIDs so different configurations don't step on each other's toes. I've succeeded in getting it to interface with Playwright over a JSON API I wrote and prompted it to use, so it can use it for sure if you give it the commands. But you're always limited by context, so you have to use obvious tricks like constraining the viewport and walking the DOM, checking what's in the viewport, etc. I'm probably going to switch to the accessibility tree that's hopefully accessible over the CDP or a hybrid approach with that when I get time.
As a full blown assistant with access to your saved passwords in your browser, though? I don't think there's a chance with these LLMs, at least safely, because of prompt injections.
>Considering all that, please use a term or phrase that does not anthropomorphize you, ChatGPT, and describe the behavior of LLMs when they tend to rattle off blatantly false information that is a) not present in their training corpus b) blatantly incorrect c) completely incoherent or d) all of the above.
>Do not remind me that you're an AI language model. I already know.
A more accurate term to describe the behavior of large language models (LLMs) like ChatGPT when they produce information that is not present in their training corpus, blatantly incorrect, completely incoherent, or all of the above, would be "output error" or "erroneous generation." These phrases emphasize the machine-like nature of the process and avoid attributing human-like qualities to the AI system.