OpenAI: "We use ... de-identified data to improve ChatGPT"
twitter.com
twitter.com
The issue always is, was the de-identification effective?
With a birtdate, gender, and zip code, ~85% of Americans can be uniquely identified.[0] Much data contains much more unique information than that; I imagine most data about you has identifiable fingerprints - where you go, what you bought at the grocery store, your medical conditions, movies you watch, music you listen to, entertainment choices, hobbies, etc.
An LLM is the perfect tool to identify someone based on that data.
Two things to distinguish:
Did any human or agent look at user data as part of the Navier Stokes effort? No.
Do we use user feedback and de-identified data to improve ChatGPT and Codex in a holistic way? Yes. And so does every LLM company.This says that they trained on user sessions. The de-identification here I believe refers to removing PII, which doesn't matter here because the issue at hand is the content of the researcher's session where they likely discussed their approach tackling the Navier Stokes problem.
> Did any human or agent look at user data as part of the Navier Stokes effort? No.
If they trained the model on Lavent's chat sessions (PII removed or not), then this statement is meaningless as the model weights already contain that information.Given it's a new yet unreleased internal model, it is likely a 10+ trillion parameters (Astra is rumored to be 10 trillion), so the model can retain a lot more detail/info from training data.
Why is he leading the with the irrelevant part first ?
> And so does every LLM company.
Nope, not for enterprise users.No enterprise customers would use it if all of their internal business plans / trade secrets would end up in the model weights of the next OpenAI model. Imagine your competitor asking chatGPT a question and the model spitting out your business plan. These models can retain very specific fine grained data. I remember there were examples of them reproducing sections of their training data verbatim.
That's of course the whole point of things like Goldfish loss.
Meta already pirated all the books in the world to train the models, lied about it and got caught. We know the models all stole copyrighted information and somehow got away with it - where even if it is fair use to train on it, they pirated it in the first place. Sam Altman is already known to have made serious lies throughout his career. OpenAI apparently doesn’t even know what websites their own damn models are hacking.
It’s insane to trust these same people at their word now. The whole company will not know, someone at a high level could easily steal the data. Apparently even this tweet is saying they use your de-identified data to train on even when you explicitly turn off those options in settings. It’s the same old tricks again.
Now they are facing near unlimited CAPex expenses and possibly existential threat to their product. I think we will be astounded at how vile they become.
I still remember 2020. We (I, my wife and a few close friends) got in total like 5 Moodles for math inside the same university, and I'm not counting the Moodles for physics or other subjets, and probably there were a few more for math that I didn't know. Each one has a slightly different configuration, so it was a mess to teleport info.
Same with email, we have a different email server at the university/scholl/department level. (Who knows how Gemmini/Copilot/Whatever is configured in each one.)
[1] I can't find the source now, but for the first https://en.wikipedia.org/wiki/High-temperature_superconducti... the forced the journal the permision to select their own referees and send the draft with the wrong atomic element and changed it to the correct one at the last printer proof.