7 ms·

because corporations are using providers with ZDR in the contract. If OAI or any of the cloud providers violate this they're getting sued to oblivion.

dofm 2mo ago
The problem is that there is an enormous, nearly unignorable incentive to work around it. So they will. As the customer base becomes more and more corporate (which it will), they end up with disproportionately more customers whose experiences cannot be used to train the model to make it better for those customers. Either way, corporate customers cannot leach off the training from consumers handing over their personal data forever; there aren't enough specialists in that training set to improve the models with no loss of corporate trust. Betrayal of their trust is inevitable.
- WarmWash 2mo ago
  Conspiracies are for the chronically online
  - dofm 2mo ago
    This is not a conspiracy theory. It's futurology, maybe, but pretty basic stuff at that. At some point, where does the training advantage for specialist LLMs come from, if not progressively encroaching on customer data for the benefit of equivalent customers?
  - rightbyte 2mo ago
    It is rather vague whether you count Sam Altman et al. as "chronically online".
coffeefirst 2mo ago
These are the same people who performed the largest scale breach of copyright in history on the theory that they could get away with it. I’m not making any accusations, but we should not underestimate their tolerance for legal and financial risk. It may be a little paranoid to insist on self hosting based on that, but I’m not so sure that it’s crazy.
- dofm 2mo ago
  Trade secrecy is all anyone has left. It's not paranoid at all. You would hope that most serious companies have a tier of corporate knowledge protection that is somewhere between Coca-Cola/KFC herbs recipe secrecy and Stringer Bell's note-taking exhortation: "is you giving the LLM notes on our unique advantage?"
- WarmWash 2mo ago
  It has been ruled that training on copyright is not a breach of copyright unless you subverted payment for it. Which they did do, but scale is relatively miniscule to the full dataset.

dofm 2mo ago

The problem is that there is an enormous, nearly unignorable incentive to work around it. So they will. As the customer base becomes more and more corporate (which it will), they end up with disproportionately more customers whose experiences cannot be used to train the model to make it better for those customers. Either way, corporate customers cannot leach off the training from consumers handing over their personal data forever; there aren't enough specialists in that training set to improve the models with no loss of corporate trust. Betrayal of their trust is inevitable.

WarmWash 2mo ago
Conspiracies are for the chronically online
- dofm 2mo ago
  This is not a conspiracy theory. It's futurology, maybe, but pretty basic stuff at that. At some point, where does the training advantage for specialist LLMs come from, if not progressively encroaching on customer data for the benefit of equivalent customers?
- rightbyte 2mo ago
  It is rather vague whether you count Sam Altman et al. as "chronically online".

coffeefirst 2mo ago

These are the same people who performed the largest scale breach of copyright in history on the theory that they could get away with it. I’m not making any accusations, but we should not underestimate their tolerance for legal and financial risk. It may be a little paranoid to insist on self hosting based on that, but I’m not so sure that it’s crazy.

dofm 2mo ago
Trade secrecy is all anyone has left. It's not paranoid at all. You would hope that most serious companies have a tier of corporate knowledge protection that is somewhere between Coca-Cola/KFC herbs recipe secrecy and Stringer Bell's note-taking exhortation: "is you giving the LLM notes on our unique advantage?"
WarmWash 2mo ago
It has been ruled that training on copyright is not a breach of copyright unless you subverted payment for it. Which they did do, but scale is relatively miniscule to the full dataset.