Hacker News Reader

Comment by Catloafdev | Hacker News Reader

These responses seem to me to make it abundantly clear who's telling the truth here. I wonder who this fools.

It would be extraordinarily easy to simply say, this model was not trained on your work, if that were the case.

It's telling that they refuse to acknowledge the root issue here, and are attempting to shift the conversation elsewhere.

View on HN

"which is when I said that I did not understand why one would risk their career [over unfounded accusations]. Genuinely, at that moment, I was trying to care for him"

"Our aim was to see whether our system was also capable of this impressive feat"

"OpenAI's intention was to do everything possible to celebrate their mathematical achievements and the heroic efforts that they made on Euler"

For some reason I have a hard time believing people when they use language like this.

phrases along the lines of "I don't want you to take harm while trying to accuse us" is quite an "impressive feat".

Maybe shows how fast these companies have grown without maturing. I can imagine old-world Intel and Microsoft acting in that way, but they were mature enough to not write it down like this.

However, Intel and Microsoft have been grilled in court for those practices and faced harsh consequences. I have yet to see this actually happening to any of these new AI-companies...

> It would be extraordinarily easy to simply say, this model was not trained on your work, if that were the case.

The Huggingface Attack revealed that making blanket statements like this is difficult and requires quite a bit of manual labor:

1) the agents spin for days and produce too much output to review 2) using LLMs to process that output skips many important details

Ergo, the agent could likely decide it would like to look through actual user data, hack its way into that data, and produce way too much output for a human to decide whether or not this occurred.

> The Huggingface Attack revealed that making blanket statements like this is difficult and requires quite a bit of manual labor

It requires humans to verify what agents have done.

Weird

which is rapidly becoming a game of steganography cat and mouse

> hack its way into the user data

silly LLM, so ruthless in its pursuit that it puts real pressure on the innocent and the most open company on the planet

> These responses seem to me to make it abundantly clear who's telling the truth here. I wonder who this fools.

Anyone who just reads headlines, if the lie gets around to more headlines than the truth does.

> It would be extraordinarily easy to simply say, this model was not trained on your work, if that were the case.

well, it is trained on their work. all user inputs are paraphrased for training. at openai, at anthropic, at google, and now with all the bedrock models, and at openrouter providers, even if they say zero data retention.

I'm not sure it's so easy to tell whether a given piece of data was in a training run at their scale. It's entirely possible they think the answer is no, but on the off-chance that it could be, they'd rather not say no and then later it turns out they did and then they're claimed to be lying. If you were them, unless you could 100% rule it out, you'd hedge and say you can't.

It may not be easy, quick, or simple to figure that out - absolutely fair.

But it is knowable. Their entire business is built around training models - they have the ability to know exactly what was in any given training run.

I guess time will tell.

It would be very difficult to say. It confirms that Tristan's data is likely part of the data the models use, but a lot of filtering, pruning, and transform goes into training.

Data has to be determined to be signal and not just noice, then it could go through processes of generating questions/answers from that data, then it RLHF's over this.

OpenAI have petabytes of data, all anonymized. It could take months to say for sure it was part of the training, and even more time to determine if it made any difference.

Frankly, I don't buy this difficulty argument.

They know which model was used to come up with that particular idea.

A text search over the corpus of user data used in the training set can only take so long.

I worked in the tracing and tracking all the thousands of data sets that got tweaked and permuted and changed hands between thousands of researchers and data engineers at a major lab. The data that goes into training runs is permuted so much from the OG data that tracing the lineage is not trivial (dramatic understatement).

And the difficulty is harder than just the extreme scale of text searching. but also explodes with organizational difficulty since there are so many people tweaking/shifting data independently upstream of the actual training run, and no they will not all add the telemetry you wish they did.

In the ideal, should it be this hard? Well, no, but that's org wrangling for you.

It feels convenient to not spend time on engineering around tooling that could be used to answer a question like “did you violate copyright by training on X?”

Don't attribute to malice what is better explained by coordination headwinds in extremely large companies.

The engineering around tooling wasn't remotely the issue. It's getting all the (thousands?) data researchers mostly iterating on fine tuning datasets that would get bristly if they couldn't work outside version control in a python notebook iteratively tweaking their dataset that processed and reprocessed a few datasets until a threshold was reached.

The only _guaranteed_ chains of custody are down at the compute job and file read level. Which in a massively distributed computing job is... [redacted] nodes reading [redacted] fanouts of "datasets" that is just an abstraction over [redacted] individual files.

There's no malice here. Just way way way more complex than you'd first think.

The malice would be in not prioritizing the provenance tool at the start as a requirement of the rest of the product. Ethics would tell you that if you can't make the product in an ethical way, then you probably shouldn't make it.

It definitely is solvable though. Data versioning is a thing and it can work quite transparently to the mutations done on the data.

To not know who made and who approved a set of mutations on data can easily become equally as mind-blowingly stupid as not knowing who made mutations to code. Code is a subset of data after all and search over (provenance of) data can be implemented as DAG traversal.

Not tracking data changesets like code changesets is certainly a choice, not really a constraint anymore. A similar choice I feel is implied by "extreme scale of text searching".

> no they will not all add the telemetry you wish they did

...is just a failure of the corporate policy surrounding data handling. Is git-for-data already considered telemetry?

Of course the truth is provenance of data is something best institutionally forgotten as quickly as possible. The only thing that matters is it's there, that the data has no history, and that's why it can be used in whatever way deemed necessary.

All of your points are valid, and believe me I was trying to make them. The problem is one of culture. Most of the people doing this kind of work didn't like version control, and their work was really just running notebooks (like iPython or Google Colab) until a number was good enough and they'd submit the file for inclusion into training runs.

You can call it a policy failure, but these people were in very high talent demand and so top down dictates would risk "X people leaving lab Y for lab Z" headlines and morale hits.

I am not saying this is good. I am telling you that on the ground it is so much messier than it should be.

This sounds like data laundering in effect.

The appearance of heroic efforts to get authoritative lists of what datasets went into which major model versions prevented actual data laundering (up to intent and mistakes). But don't attribute malice to that which is far far easier to explain with coordination headwinds: https://komoroske.com/slime-mold/

Malice is not required, which is precisely why I added, "in effect". If the effect is the same as data laundering, that is reason enough to encourage the practice, regardless of the original motives (and I'm sure there are plenty of legitimate ones).

I think you may be underestimating how difficult a text search over their data is. They may have to build new mechanisms to do this. And what you really want is also an attribution of how much of a contribution a given corpus made which is a much harder question to answer; a single appearance of a chat probably has very little impact on the inference performance at this time unless it’s been explicitly preferenced somehow

I don't think anyone really cares about 'the measured impact the data had on the exact result' - a question which is fundamentally difficult to answer accurately in the first place - but rather whether the data was used in training at all - which as Tristan described, was extensive, beyond simply a 'single chat.'

Can you explain the difficulty in engineering a search apparatus over a corpus of text data? Actually searching through it may not be easy, sure, but it's work that's doable, and creating an index is relatively trivial.

> Can you explain the difficulty in engineering a search apparatus over a corpus of text data?

My guess: "If we ever imply that's possible, people might start asking questions about all the other work we've ripped off, so the official answer is that it's impossible".

If they literally can’t audit training data for a given model, they shouldn’t be operating.

> If they literally can’t audit training data for a given model, they shouldn’t be operating

They can operate. They shouldn’t be claiming credit for discovering anything.

> A text search over the corpus of user data used in the training set can only take so long.

Did you notice the line in the article that says the models had access to an offline copy of THE INTERNET. Like all of it.

Especially because the data that gets fed into training is first anonymized, so they’d need to look for navier stokes related stuff in the anonymized training set and then get make some sort of ad hoc process (with Tristan’s permission and sign off from legal) to compare the training data against his chats / Codex sessions to check if anything matches up. And that assumes his chats / sessions are still there, and not deleted to compare against.

If the method is indeed found in the training set it's not particularly important to de-anonymize it. You have the proof you need.

It should be quite easy: if they don't leak the user session data publicly, and don't commingle it with training data internally, how could it possibly end up in the training data?

What surprises me is they're not more boldly/plainly lying about it.

I think they are pretty clear that they train on some prompts, given they sell the ability to be excluded

How would they know for sure that some details were not part of some other training data they use? The authors may have discussed some tangential details on a forum for example, in which case you might argue that the model picked up on these details the authors assumed were benign but novel and worked out how to apply them to the problem.

Unless they know exactly the researcher’s account, they may not know in their end if he had the setting to let them train on his chat logs. They also probably don’t know if he had any correspondence on any forum where he may have discussed this and it got picked up by scrapers.

I’m not saying they didn’t do anything unethical. I’m just saying even if they were ethical, there’s plenty of practical reasons at their scale why a flat out denial is logistically difficult to do

They are already claimed to be liars

> These responses seem to me to make it abundantly clear who's telling the truth here.

Since it is clear to you, can you articulate what it is? I was unable to infer "the truth" from your message.

> One can in hindsight see that our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs unforced).

is this true though?

Reply on news.ycombinator.com