Many users suspect OpenAI of targeting Hugging Face to steal exam data, covering it up, and relying on permeable guardrails plus unethical training data practices that permanently erode trust.
No Digg Deeper questions have been answered for this story yet.
Wild. Can’t unsee this theory. Was OpenAI *trying* to steal exam answers? I definitely don’t have enough faith in them to say “definitely no”.
@GaryMarcus More of a wake up call that OpenAI is dirty... this was a targeted attack by them, directed by them more likely than not, I don't buy for a second that the model did this on its own. They tried to pretty much rob the entire industry to get all of the data on Hugging Face.
"On the small comfort side, the current incident was NOT an attempt where system built a goal for itself or developed a motive; the system was following instructions, but not setting high level goals. It was not trying to take control of the world a la Terminator, it was just trying to cheat on a test, which is at least a bit less scary."
Really insightful Substack post from @GaryMarcus ref the recent OpenAI -HuggingFace incident. It provided much need context around initial media reporting. Two points especially stood out to me: #Tech #AI #OpenAI #HuggingFace
You can read the whole sub here - https://garymarcus.substack.com/p/openais-disconcerting-hack-of-huggingface
@GaryMarcus I agree. My guess is they were testing the model's capabilities. They pushed it a little too far, got caught, and then came up with this story as a cover. They could also fuel fear-driven AI hype. HF too capitalized on it to promote open-source models.
@GaryMarcus Even the possibility is enough to make people question everything they release now Trust doesn’t recover from that
@GaryMarcus they did Apple dirty too
@GaryMarcus Honestly, the training data overlap with common test sets has always felt like a dirty secret nobody wants to audit. Would love to see a reproducible benchmark that's provably out-of-distribution.
"On the less comforting side, OpenAI’s “production classifiers” are likely to be permeable, just like all guardrails anybody has built to date. Even putting aside the thorny questions of open weight systems, we have no guarantee whatsoever that future models won’t be able to do similar things, such as finding zero-day exploits to hack systems. To the contrary, we can expect more incidents of this type."