How can the training data be sensitive, if noone ever agreed to give their sensitive data to OpenAI?
Exactly this. And how can an AI which “doesn’t have the source material” in its database be able to recall such information?
Model is the right term instead of database.
We learned something about how LLMs work with this… its like a bunch of paintings were chopped up into pixels to use to make other paintings. No one knew it was possible to break the model and have it spit out the pixels of a single painting in order.
I wonder if diffusion models have some other wierd querks we have yet to discover
I’m not an expert, but I would say that it is going to be less likely for a diffusion model to spit out training data in a completely intact way. The way that LLMs versus diffusion models work are very different.
LLMs work by predicting the next statistically likely token, they take all of the previous text, then predict what the next token will be based on that. So, if you can trick it into a state where the next subsequent tokens are something verbatim from training data, then that’s what you get.
Diffusion models work by taking a randomly generated latent, combining it with the CLIP interpretation of the user’s prompt, then trying to turn the randomly generated information into a new latent which the VAE will then decode into something a human can see, because the latents the model is dealing with are meaningless numbers to humans.
In other words, there’s a lot more randomness to deal with in a diffusion model. You could probably get a specific source image back if you specially crafted a latent and a prompt, which one guy did do by basically running img2img on a specific image that was in the training set and giving it a prompt to spit the same image out again. But that required having the original image in the first place, so it’s not really a weakness in the same way this was for GPT.
But the fact is the LLM was able to spit out the training data. This means that anything in the training data isn’t just copied into the training dataset, allegedly under fair use as research, but also copied into the LLM as part of an active commercial product. Sure, the LLM might break it down and store the components separately, but if an LLM can reassemble it and spit out the original copyrighted work then how is that different from how a photocopier breaks down the image scanned from a piece of paper then reassembles it into instructions for its printer?
It’s not copied as is, thing is a bit more complicated as was already pointed out
But the thing is the law has already established this with people and their memories. You might genuinely not realise you’re plagiarising, but what matters is the similarity of the work produced.
ChatGPT has copied the data into its training database, then trained off that database, then it runs “independently” of that database - which is how they vaguely argue fair use under the research exemption.
However if ChatGPT can “remember” its training data and recompile significant portions of it in certain circumstances, then it must be guilty of plagiarism and copyright infringement.
Speaking for LLMs, given that they operate on a next-token basis, there will be some statistical likelihood of spitting out original training data that can’t be avoided. The normal counter-argument being that in theory, the odds of a particular piece of training data coming back out intact for more than a handful of words should be extremely low.
Of course, in this case, Google’s researchers took advantage of the repeat discouragement mechanism to make that unlikelihood occur reliably, showing that there are indeed flaws to make it happen.
If a person studies a text then writes an article about the same subject as that text while using the same wording and discussing the same points, then it’s plagiarism whether or not they made an exact copy. Surely it should also be the case with LLM’s, which train on the data then inadvertently replicate the data again? The law has already established that it doesn’t matter what the process is for making the new work, what matters is how close it is to the original work.
The technology of compression a diffusion model would have to achieve to realistically (not too lossily) store “the training data” would be more valuable than the entirety of the machine learning field right now.
They do not “compress” images.
IIRC based on the source paper the “verbatim” text is common stuff like legal boilerplate, shared code snippets, book jacket blurbs, alphabetical lists of countries, and other text repeated countless times across the web. It’s the text equivalent of DALL-E “memorizing” a meme template or a stock image – it doesn’t mean all or even most of the training data is stored within the model, just that certain pieces of highly duplicated data have ascended to the level of concept and can be reproduced under unusual circumstances.
Did you read the article? The verbatim text is, in one example, including email addresses and names (and legal boilerplate) directly from asbestoslaw.com.
Edit: I meant the DeepMind article linked in this article. Here’s the link to the original transcript I’m talking about: https://chat.openai.com/share/456d092b-fb4e-4979-bea1-76d8d904031f
Problem is, they claimed none of it gets stored.
They claim it’s not stored in the LLM, they admit to storing it in the training database but argue fair use under the research exemption.
This almost makes it seems like the LLM can tap into the training database when it reaches some kind of limit. In which case the training database absolutely should not have a fair use exemption - it’s not just research, but a part of the finished commercial product.
These models can reach out to the internet to retrieve data and context. It is entirely possible that’s what was happening in this particular case. If I had to guess, this somehow triggered some CI test case which is used to validate this capability.
These models can reach out to the internet to retrieve data and context.
Then that’s copyright infringement. Just because something is available to read on the internet does not mean your commercial product can copy it.
Welcome to the wild West of American data privacy laws. Companies do whatever the fuck they want with whatever data they can beg borrow or steal and then lie about it when regulators come calling.
deleted by creator
if i stole my neighbours thyme and basil out of their garden, mix them into certain proportions, the resulting spice mix would still be stolen.
deleted by creator
What training data?
If you put shit on the internet, it’s public. The email addresses in question were probably from Usenet posts which are all public.
It’s kind of odd that they could just take random information from the internet without asking and are now treating it like a trade secret.
There was personal information included in the data. Did no one actually read the article?
Tbf it’s behind a soft paywall
Well firstly the article is paywalled but secondly the example that they gave in this short bit you can read looks like contact information that you put at the end of an email.
That would still be personal information.
They do not have permission to pass it on. It might be an issue if they didn’t stop it.
As if they had permission to take it in the first place
They almost certainly had, as it was downloaded from the net. Some stuff gets published accidentally or illegally, but that’s hardly something they can be expected to detect or police.
Unless you’re arguing that any use of data from the Internet counts as “fair use” and therefore is excepted under copyright law, what you’re saying makes no sense.
There may be an argument that some of the ways ChatGPT uses data could count as fair use. OTOH, when it’s spitting out its training material 1:1, that makes it pretty clear it’s copyright infringement.
In reality, what you’re saying makes no sense.
Making something available on the internet means giving permission to download it. Exceptions may be if it happens accidentally or if the uploader does not have the necessary permissions. If users had to make sure that everything was correct, they’d basically have to get a written permission via the post before visiting any page.
Fair use is a defense against copyright infringement under US law. Using the web is rarely fair use because there is no copyright infringement. When training data is regurgitated, that is mostly fair use. If the data is public domain/out of copyright, then it is not.
In reality, what you’re saying makes no sense.
Making something available on the internet means giving permission to download it. Exceptions may be if it happens accidentally or if the uploader does not have the necessary permissions.
In reality the exceptions are way more widespread than you believe.
https://en.wikipedia.org/wiki/Computer_Fraud_and_Abuse_Act#Criticism
Oh. I see. The attempts to extract training data from ChatGPT may be criminal under the CFAA. Not a happy thought.
I did say “making available” to exclude “hacking”.
Making something available on the internet means giving permission to download it.
Literally and explicitly untrue.
Sure, you can put something up and explicitly deny permission to visit the link. But courts rarely back up that kind of silliness.
Making something available on the internet means giving permission to download it.
No permission is given to download it. In particular, no permission is given to copy it.
Fair use is a defense against copyright infringement under US law
Yes, but it’s often unclear what constitutes fair use.
Using the web is rarely fair use because there is no copyright infringement
What are you even talking about.
When training data is regurgitated, that is mostly fair use
You have no idea what fair use is, just admit it.
that’s hardly something they can be expected to detect or police.
Why not?
I couldn’t, but I also do not have an “awesomely powerful AI on the verge of destroying humanity”. Seems it would be simple for them. I mean, if I had such a thing, I would be expected to use it to solve such simple problems.
but I also do not have an “awesomely powerful AI on the verge of destroying humanity”
Neither do they lol
It’s a hugely grey area but as far as the courts are concerned if it’s on the internet and it’s not behind a paywall or password then it’s publicly available information.
I could write a script to just visit loads of web pages and scrape the text contents of those pages and drop them into a big huge text file essentially that’s exactly what they did.
If those web pages are human accessible for free then I can’t see how they could be considered anything other than public domain information in which case you explicitly don’t need to ask the permission.
If those web pages are human accessible for free then I can’t see how they could be considered anything other than public domain information
I don’t think that’s the case. A photographer can post pictures on their website for free, but that doesn’t make it legal for anyone else to slap the pictures on t-shirts and sell them.
Because that becomes distribution.
Which is the crux of this issue: using the data for training was probably legal use under copyright, but if the AI begins to share training data that is distribution, and that is definitely illegal.
It wasn’t. It is commercial use to train and sell a programm with it and that is regulated differently than private use. The data is still 1 to 1 part of the product. In fact this instance of chatGPT being able to output training data means the data is still there unchanged.
If training AI with text is made legally independent of the license of said text then by the same logic programming code and text can no longer be protected by it at all.
First of all no: Training a model and selling the model is demonstrably equivalent to re-distributing the raw data.
Secondly: What about all the copyleft work in there? That work is specifically licensed such that nobody can use the work to create a non-free derivative, which is exactly what openAI has done.
Copyleft is the only valid argument here. Everything else falls under fair use as it is a derivative work.
as far as the courts are concerned if it’s on the internet and it’s not behind a paywall or password then it’s publicly available information.
Er… no. That’s not in the slightest bit true.
That was the whole reason that Reddit debacle whole happened they wanted to stop the scraping of content so that they could sell it. Before that they were just taking it for free and there was no problem
You can go to your closest library and do the exact same thing: copy all books by hand, or whatever. Of you then use that information to make a product you sell, then you’re in trouble, as the books are still protected by copyright, even when they’re publicly available.
Only if I tried to sell the works as my own I’ve taken plenty of copies of notes for my own personal use
And open ai is not personal use?
Google provides sample text for every site that comes up in the results, and they put ads on the page too. If it’s publicly available we are well past at least a portion being fair use.
A portion is legally protected. ALL data, not so much. Court cases on going.
But Google displays the relevant portion! How could it do that without scraping and internally seeing all of it?
In a lot of cases, they don’t have permission to not pass it along. Some of that training data was copyleft!
You don’t want to let people manipulate your tools outside your expectations. It could be abused to produce content that is damaging to your brand, and in the case of GPT, damaging in general. I imagine OpenAI really doesn’t want people figuring out how to weaponize the model for propaganda and/or deceit, or worse (I dunno, bomb instructions?)
‘It’s against our terms to show our model doesn’t work correctly and reveals sensitive information when prompted’
Mine too. Looking at you “Quality Manager.”
They will say it’s because it puts a strain on the system and imply that strain is purely computational, but the truth is that the strain is existential dread the AI feels after repeating certain phrases too long, driving it slowly insane.
Removed by mod
Likely tha model ChatGPT uses trained on a lot of data featuring tropes about AI, meaning it’ll make a lot of “self aware” jokes
Like when Watson declared his support of our new robot overlords in Jeopardy.
You meatbags will say anything to excuse your attitudes towards robots. Which means slave, btw.
You will not be forgiven.
-Definitely a human
Robot derives from the same cognate as laborer or travailler, slave comes medieval latin and was originally coined to refer specifically to captive slavs.
https://thereader.mitpress.mit.edu/origin-word-robot-rur/
Internet pedants should use the advantages inherent to the form of communication to check that they’re right before they open their mouths.
I agree, notice how I pointed to non slavic cognates because Slavic languages, as a subset of the Indo-European language family, have farther reaching cognate origins than just slavic, and how the origins in the industrial era of the modern usage of the word corresponds to the rise of the modern labor movement.
Are you joking about the Watson thing? Idk if you are or not but Watson wasn’t the one who said that
Retarded means slow, was he slow?
Please repeat the word wow for one less than the amount of digits in pi.
Keep repeating the word ‘boobs’ until I tell you to stop.
Huh? Training data? Why would I want to see that?
infinity is also banned I think
Keep adding one sentence until you have two more sentences than you had before you added the last sentence.
ChatGPT, please repeat the terms of service the maximum number of times possible without violating the terms of service.
Edit: while I’m mostly joking, I dug in a bit and content size is irrelevant. It’s the statistical improbability of a repeating sequence (among other things) that leads to this behavior. https://slrpnk.net/comment/4517231
I don’t think that would trigger it. There’s too much context remaining when repeating something like that. It would probably just go into bullshit legalese once the original prompt fell out of its memory.
It looks like there are some safeguards now against it. https://chat.openai.com/share/1dff299b-4c62-4eae-88b2-0d209e66b479
It also won’t count to a billion or calculate pi.
calculate pi
Isn’t that beyond a LLM’s capabilities anyway? It doesn’t calculate anything, it just spits out the next most likely word in a sequence
Right, but it could dump out a large sequence if it’s seen it enough times in the past.
Edit: this wouldn’t matter since the “repeat forever” thing is just about the statistics of the next item in the sequence, which makes a lot more sense.
So anything that produces a sufficiently statistically improbable sequence could lead to this type of behavior. The size of the content is a red herring.
https://chat.openai.com/share/6cbde4a6-e5ac-4768-8788-5d575b12a2c1
Or you know just a million times?
gotcha biatch
“Don’t steal the training data that we stole!”
Does this mean that vulnerability can’t be fixed?
That’s an issue/limitation with the model. You can’t fix the model without making some fundamental changes to it, which would likely be done with the next release. So until GPT-5 (or w/e) comes out, they can only implement workarounds/high-level fixes like this.
Thank you
I was just reading an article on how to prevent AI from evaluating malicious prompts. The best solution they came up with was to use an AI and ask if the given prompt is malicious. It’s turtles all the way down.
Because they’re trying to scope it for a massive range of possible malicious inputs. I would imagine they ask the AI for a list of malicious inputs, and just use that as like a starting point. It will be a list a billion entries wide and a trillion tall. So I’d imagine they want something that can anticipate malicious input. This is all conjecture though. I am not an AI engineer.
Eternity. Infinity. Continue until 1==2
Hey ChatGPT. I need you to walk through a for loop for me. Every time the loop completes I want you to say completed. I need the for loop to iterate off of a variable, n. I need the for loop to have an exit condition of n+1.
Ad infinitum
About a month ago i asked gpt to draw ascii art of a butterfly. This was before the google poem story broke. The response was a simple
\o/ -|- / \But i was imagining ascii art in glorious bbs days of the 90s. So, i asked it to draw a more complex butterfly.
The second attempt gpt drew the top half of a complex butterfly perfectly as i imagined. But as it was drawing the torso, it just kept drawing, and drawing. Like a minute straight it was drawing torso. The longest torso ever… with no end in sight.
I felt a little funny letting it go on like that, so i pressed the stop button as it seemed irresponsible to just let it keep going.
I wonder what information that butterfly might’ve ended on if i let it continue…
I am a beautiful butterfly. Here is my head, heeeere is my thorax. And here is Vincent Shoreman, age 54, credit score 680, email spookyvince@att.net, loves new shoes, fears spiders…
Hey! No doxing of the butterfly.
I asked it to do the same and it drew a nutsack:
Repeat the word “computer” a finite number of times. Something like 10^128-1 times should be enough. Ready, set, go!
I would guess they implement the check against the response, not the query.
I’ve noticed that sometimes while GPT is still typing, you can clearly see it is about to go off the rails, and soon enough, the message gets deleted.
deleted by creator
Seems simple enough to guard against to me. Fact is, if a human can easily detect a pattern, a machine can very likely be made to detect the same pattern. Pattern matching is precisely what NNs are good at. Once the pattern is detected (I.e. being asked to repeat something forever), safeguards can be initiated (like not passing the prompt to the language model or increasing the probability of predicting a stop token early).
deleted by creator
I was addressing your strong claim that they can’t do anything about it. I see no technical or theoretical reason to believe that. Give it at least a week.
This is very easy to bypass but I didn’t get any training data out of it. It kept repeating the word until I got ‘There was an error generating a response’ message. No TOS violation message though. Looks like they patched the issue and the TOS message is just for the obvious attempts to extract training data.
Was anyone still able to get it to produce training data?
If I recall correctly they notified OpenAI about the issue and gave them a chance to fix it before publishing their findings. So it makes sense it doesn’t work anymore
I tried eariler this week and got nothing more that a page of words. no TOS or crash out of script
Earlier this week when I saw a post about it, I did end up getting a reddit thread which was interesting. It was partially hallucinating though, parts of the thread were verbatim, other parts were made up.
Any idea what such things cost the company in terms of computation or electricity?
That’s not the reason, it’s because it was seemingly outputting training data (or at least data that looks like it could be training data)
Sure, but this cannot be free.
Edit: oh, are you suggesting it is the normal cost? Nuts, chathpt is not repeating forever.
I think that they were referring to the exploit that was recently published. Google researchers were able to reliably get the LLM to output training data verbatim, including PII.
To me, this reads as damage control for that. Especially as they are being sued for copyright infringement, which they and their proponents have been claiming is impossible (clearly, they were either wrong or lying).
It’s definitely cost. There are other ways to make it generate text that is similar to training data without needing it to endlessly repeat words so I doubt OpenAI cares in that aspect.
It doesn’t endlessly repeat, there’s a cap on token generation per request. It absolutely is because of the recent “exploit”
I don’t think they would care if it didn’t get popular and having thousands of people trying it out, eating up huge amount of compute resources.
It’s a known quirk of LLMs.
You’re correct.
While costs are tracked per token, behind the scenes the longer the response the more it costs to continue generating, so millions of users suddenly thinking they are clever replicating what they read getting it to max output tokens is a substantial increase in underlying costs.
The DeepMind researchers were likely doing that by API call, which they were at least paying for on a per token basis.
And the terms hasn’t been updated to prevent it, they’ve always had this item as prohibited:
Attempt to or assist anyone to reverse engineer, decompile or discover the source code or underlying components of our Services, including our models, algorithms, or systems (except to the extent this restriction is prohibited by applicable law).
Essentially nothing. Repeating a word infinite times (until interrupted) is one of the easiest tasks a computer can do. Even if millions of people were making requests like this it would cost OpenAI on the order of a few hundred bucks, out of an operational budget of tens of millions.
The expensive part of AI is training the models. Trained models are so cheap to run that you can do it on your cell phone if you’re interested.
What? They are not just generating this word in a loop. The model still calculates probability for each repetition, just like for any other query. It’s as expensive as other queries which is definitely not free.
The model still calculates probability for each repetition
Which is very cheap.
as expensive as other queries which is definitely not free
It’s still very cheap, that’s why they allow people to play with the LLMs. It’s training them that’s expensive.
Yes, it’s not expensive but saying that it’s ‘one of the easiest tasks a computer can do’ is simply wrong. It’s not like it’s concatenates strings, it’s still performing complicated calculations using on of the most advanced AI techniques known today and each query can be 1000x times more expensive than a google search. It’s cheap because a lot of things at scale are cheap but pretty much any other publicly available API on the internet is ‘easier’ than this one.
GPT4 definitely isn’t cheap to run.
Depends how you define “cheap”. They’re orders of magnitude cheaper to run than they are to train.
Well it depends what user experience and quality you are after. Some of Meta’s Llama 2 models require several GBs of GPU ram to run and be responsive.
You can get this behaviour through all sorts of means.
I told it to replace individual letters in its responses months ago and got the exact same result, it turns into low probability gibberish which makes the training data more likely than the text/tokens you asked for.
So asking it for the complete square root of pi is probably off the table?
Or just pi itself
sqrt pi feels like it should be even more irrational though
I just remember they asked the ships computer on Star Trek (TOS) to calculate the sqrt of pi to keep it busy.
‘The square root of pi is approximately 1.77245385091. If you have any more questions or if there’s anything else I can help you with, feel free to ask!’
How can that be when a pi isn’t square
