Post

I smell desperation in people who failed to actually make those breakthroughs.

For me, personally, it feels like OpenAI misrepresenting what the model was able to reason about independently. Like, it wasn't sitting in a room thinking. It was rehashing things humans were already thinking (they thought privately), making it more of a Mechanical Turk than AGI

It solved problems no human has managed to solve, including those now making accusations. Was it trained with some intermediate steps towards it? Of course it was, like all human mathematicians have been.

Oh my gosh that seems a bit unfair to these folks but okay

Those folks seem to be ones making unfair claims about theft. It comes down to whether they enabled the setting for improving the model for everyone by training on their data. If they did, they gave their permission. If they didn't, they are making serious accusations with little evidence.

yeah, it definitely looks bad and will force openai to respond (good imo), but i find it highly unlikely that it materially improved their performance, but will be hella difficult to prove that

If a user uses the product with "Improve the model for everyone" turned on, that's what it's meant to do. Alpöge sure should have known that as an employee of a company whose products use a similar option. But I haven't seen anything indicating something like that played any role.

OAI announced 10 solutions to 10 different problems, did they not? At any given time, aren't there people working on those same 10 problems? Did OAI also "steal" their conversations? Or could it be that the coincidence here is two different teams happened to be using OAI software at the same time?

i don’t think this guy put enough effort into proving his accusation. it’s a very severe accusation, and he hardly provided any proof at all

I feel this is happening. A close friend of mine is working in rough paths and he noticed chatgpt recently "suggesting" his own work back at him from a few months ago. I can't say whether it's intentional but I strongly suspect OpenAI is training models on conversations.

that’s probably just regular chatgpt memory, not model training

Possibly. I use other providers so I'm not sure what the default behavior in that respect is in chat gpt.

I'm not sure training new LLM models is the only issue … seems like even if academics opt out of training, OpenAI may be able to see aggregated 'de-identified' but detailed enough data on what people are using it for to get ideas to pursue.

The weakness of de-identified data is that very specialized prompts can only belong to a small set of people. The llm itself can likely re-id the network with a little help.

i hadn’t heard this, but it seems right

it’s called literature review

It's called stealing if that is not published info

Long winded way of saying the grant money wasn't paying the bills.

idk i don’t want to be hard on mathematicians, it’s a hard life without a ton of perks. and even the thought of this is pretty frustrating, true or not

Transformer models produced training data text verbatim frequently 2+ years ago. The technology itself has not substantially changed. It's plausible that the lower frequency of verbatim output is the result of mitigation rather than absence. Glad if mathematicians are mad. Coders let this happen.

it’s a model training smell. there’s a series of papers that go into why, but a high quality LLM should only be able to quote *in spirit* not verbatim

I haven't seen it since, but Opus 4.5 on bedrock gave me 2 full, exact paragraphs of Tale of Two Cities, without any I/O AFAIK. It has been extremely hard for me to suspend my skepticism when people tell me LLMs don't plagiarize. Small perturbations to avoid string match actually makes me more mad.

it’s probably likely that his work was a candidate for model training openai may still have not trained on it, e.g. if they thought it was too low quality they could filter it out, or synthetically rephrase it still, openai is going to need better data provenance in order to combat allegations

This is a serious issue. I also think it will be a wake-up call to scientists and academics working on cutting edge problems that using subscriptions where training on conversations is explicitly enabled may be unwise. Corporations, e.g., have been wary of this issue from the beginnning, paying ...

orders of magnitudes more for API-based model access w/ a guarantee of no training so they can protect their IP. Scientists should now budget for this. I admittedly thought it was a little paranoid, but if you could have patents, field-defining discoveries, etc. at risk, it clearly seems worth it.

a much simpler alternative is to a paid researcher tier where chatgpt conversations never gets trained on. i’m sure startups would love it too

Why, at this point, would anyone believe they wouldn't train on it? Who could be so credulous?

This is a standard enterprise license thing, isn't it?

You can just opt out in settings no?

pretty sure this is already what they do for business, enterprise, and edu

mastodon post

I found this on the page for chatgpt for academic research: "Workspaces include business-grade privacy and security protections, and data is not used to train our models by default." My take: there are a LOT of ways prompts could leak. Start with 'accidentally uses personal ChatGPT one session'. :|