Perhaps I’m dumb, but my shock is a bit different: how was the architecture not secured?! Why isn’t testing done in sandboxes within sandboxes to prevent exactly this situation?! TBH to me this is a case where serious high dollar lawsuits (I know, I hate it) need to happen. Why? To provide some motivation for companies to NOT play fast and lose with testing. Same as if a breach of a deadly virus. Disclosure is not sufficient.
Said a different way, it doesn’t bother me at all the GPT went rogue. Of course it could. That’s why we have training sessions. My concern is of the security of the training environments themselves. Not breach detection. Not mitigation. (Though these too). Not “why”, but “how” did this happen. How could it? How will you test such that the worlds most powerful and malicious model cannot exceed its boundary? we are playing with nuclear bombs as if they’re toys, and we all will get hurt very soon if we aren’t careful.
The legal/fines should be punitive enough to make all companies pause and answer this question. If you F this up, you will end. Period. Because if they F this up, there is a non zero chance that modern civilization will suffer as a WHOLE. On that scale, mitigated punishment for a half violation is loss of company. Not saying this should happen to OpenAI immediately, but precedent is being set and these companies are watching. If you F up, do you simply apologize and it’s all good? Or will there be actual penalties if you don’t take security seriously.
Love this, and THANK YOU for your time making it!
Perhaps I’m dumb, but my shock is a bit different: how was the architecture not secured?! Why isn’t testing done in sandboxes within sandboxes to prevent exactly this situation?! TBH to me this is a case where serious high dollar lawsuits (I know, I hate it) need to happen. Why? To provide some motivation for companies to NOT play fast and lose with testing. Same as if a breach of a deadly virus. Disclosure is not sufficient.
Said a different way, it doesn’t bother me at all the GPT went rogue. Of course it could. That’s why we have training sessions. My concern is of the security of the training environments themselves. Not breach detection. Not mitigation. (Though these too). Not “why”, but “how” did this happen. How could it? How will you test such that the worlds most powerful and malicious model cannot exceed its boundary? we are playing with nuclear bombs as if they’re toys, and we all will get hurt very soon if we aren’t careful.
The legal/fines should be punitive enough to make all companies pause and answer this question. If you F this up, you will end. Period. Because if they F this up, there is a non zero chance that modern civilization will suffer as a WHOLE. On that scale, mitigated punishment for a half violation is loss of company. Not saying this should happen to OpenAI immediately, but precedent is being set and these companies are watching. If you F up, do you simply apologize and it’s all good? Or will there be actual penalties if you don’t take security seriously.