I remember being told about them when they started as a possibly good place to work. I looked at them and said no. That non-profit stuff was bogus from the start. I expected legal trouble. Well I was wrong. They avoided that and people became rich. I'm glad I said no regardless. I have a life.
OpenAI pretended to be "open" and "research-friendly" before they found out they've found something incredibly valuable. Then, with the help of Microsoft, it all became about ROI.
Our system won't work, and more importantly we won't make any money, if we can't steal other people's work but we can't admit that because we'd be sued into oblivion.
This largely reflects the hopeless inefficiency of these learning models. They have to hoover up enormous quantities of data because nobody has yet figured out how to learn this stuff efficiently.
No, we can't tell you what we trained it on, because we know we weren't allowed to take it. And besides, that would make it easier for you to know that the emergent capabilities we are claiming are really just data leakage. Look, shiny!
Is this the worst example of misdirection and misleading "openness" from modern companies? It certainly seems to be one of the worst for some years, other than perhaps FTX and Theranos.
I've been wondering a lot lately about what products I use are training these models. So many of their licenses allow for my data/usage to be used for "internal purposes". For example are or when will all our Google Meets be used to train Gemini and would Google even have to notify us if that is the case ? Would training Gemini be equivalent to recording a meeting without consent or do they (likely) see this as different ?
Publicly available is a totally legit answer. Google search result provides me an answer to the query without a need to click on the source (if it is legit enough of a source). How is this different (except there could be some hallucination thrown in - making click and verify even more important)
Implausible deniability. “It wasn’t welded down, so it can’t be theft.”
I remember being told about them when they started as a possibly good place to work. I looked at them and said no. That non-profit stuff was bogus from the start. I expected legal trouble. Well I was wrong. They avoided that and people became rich. I'm glad I said no regardless. I have a life.
OpenAI pretended to be "open" and "research-friendly" before they found out they've found something incredibly valuable. Then, with the help of Microsoft, it all became about ROI.
I get it. But they should stop pretending...
Reading this, and the linked convo between Gates and Altman, I can't help rephrasing Conrad's exclamation, "The hubris! The hubris!"
tl;dr:
Our system won't work, and more importantly we won't make any money, if we can't steal other people's work but we can't admit that because we'd be sued into oblivion.
This largely reflects the hopeless inefficiency of these learning models. They have to hoover up enormous quantities of data because nobody has yet figured out how to learn this stuff efficiently.
"Other than that, Mrs. Lincoln, how was the play?"
Would be interesting to see BitTorrent traffic into OpenAI's IP addresses.
No, we can't tell you what we trained it on, because we know we weren't allowed to take it. And besides, that would make it easier for you to know that the emergent capabilities we are claiming are really just data leakage. Look, shiny!
Is this the worst example of misdirection and misleading "openness" from modern companies? It certainly seems to be one of the worst for some years, other than perhaps FTX and Theranos.
I've been wondering a lot lately about what products I use are training these models. So many of their licenses allow for my data/usage to be used for "internal purposes". For example are or when will all our Google Meets be used to train Gemini and would Google even have to notify us if that is the case ? Would training Gemini be equivalent to recording a meeting without consent or do they (likely) see this as different ?
You inspired this poll https://x.com/garymarcus/status/1768623567892291780?s=61
Follow the money, grasshopper. Because it's a learning curve until it's an invoice, then it's a mistake.
This emperor, on the other hand, is almost *all* clothes.
I thought you had a positive one coming next @Gary Marcus, or was that the best one can say at this point? 😁
the positive one got preempted by the OAI news!
No news - no money. Have to crank the handle.
Looking forward to the bigger piece.
Publicly available is a totally legit answer. Google search result provides me an answer to the query without a need to click on the source (if it is legit enough of a source). How is this different (except there could be some hallucination thrown in - making click and verify even more important)
When OpenAI announced they would hold back on GPT3 to protect the world that’s all anyone needed to know. They have the Musk playbook.