How much do these non-disparagement agreements really matter though? What smoking guns could there possibly be? Let's consider some hypothetical scenarios:
1. OpenAI secretly puts LSD into the water (aka CIA-level conspiracy) => someone would definitely report this, no matter what they've signed. As Bruce Schneider taught us, if you're committing a sufficiently large amount of crime and more than a handful of people know the details, someone will almost certainly rat on you. Hence, "agreements" don't matter for any major felonies - people will go ahead and report those, at the very least because doing otherwise means you'll eventually go to prison.
2. OpenAI is being "creative" around IP or otherwise does semi-shady dealings behind the curtain. Definitely possible but this is pretty much common knowledge already. AFAIK it was semi-confirmed that Sora was trained on Youtube videos, for example.
3. OpenAI doesn't really care about AI Alignment. Well... is this a surprise? Does it matter if 1, 5, 10 or 50 OpenAI employees say so on Twitter? Who will be persuaded by 10 employees stating this but not 1 employee (plus the entire AI safety community) saying this?
4. If Congress holds a hearing on AI safety and subpoenas former or current employees, all their NDAs will be pretty much void during the hearing, as a contract cannot demand that the other party breaks the law.
I guess all of this does matter to a small, highly elite group of people who are:
1. Talented enough to have been hired by OpenAI. I'd venture almost all are in the top 0.1%, if not 0.01% of all engineers worldwide.
2. Want to publicly criticize OpenAI.
3. Haven't sold their vested stock yet.
As a fellow engineer I'm more than sympathetic to my colleagues and I think every worker should get paid what they've been promised. But... these are the elite of the elite, the kinds of engineers who will find excellent, highly paid jobs no matter what. I truly hope Daniel Kokotajlo gets his money back but I'm also 99% sure that he can go get a job paying at least $600k/year tomorrow if he's inclined to do so.
I agree with (1) that if OpenAI was doing sufficiently blatant crimes, or (e.g.) they know they have a malign AI that is in imminent dander of killing thousands of people, then the employees would have gone to law enforcement, regardless of any NDAs. So we can be fairly confident it's not that.
b) the owners of said AI knew this was likely to happen, i.e. a re negligent
then I think the government would be fully justified in coming down very hard on the people responsible
OpenAI employees surely know this would be a likely government response, whether they were bound by NDAs or not.
So it can't be that, because if it was, the ex employees would be complaining loudly to the FBI about how Sam Altman is going to kill thousands of people.
I'm curious in what sense they use the word "vested" if that doesn't mean they now own that equity and can do what they want with it- if it can't be sold now because OpenAI is not yet a traded company, at the very least I'd expect someone to say "you can't just take back my ownership of vested shares". Anyone who understands this aspect of law better have a comment?
I only ever worked for one startup and it was already publicly traded. I didn't stay long enough for my options to vest, as it was a 5-year period for someone in a junior position (First big boy job out of college doing gruntwork), but it was my understanding that once my stock options were vested, there was nothing they could do to revoke them.
Re 2: having ex employees state it publicly could well take it from "everyone suspects" rumour to "legally actionable".
Re 3: going from 1 to 50 might not matter that much but going from 0 to 1 really does - at least from OpenAI's ability to recruit talent - and they were very strongly trying to keep the number at 0
Unless you stand for VC in personal and professional realms of existence, you might not believe that Altman represents the best of humanity. “Don’t be evil”, right?
> As I covered last time, if you do a casting call for 400 voice actors who are between 25 and 45, and pick the one most naturally similar to your target, that is already quite a lot of selection. No, they likely did not explicitly tell Sky’s voice actress to imitate anyone, and it is plausible she did not do it on her own either. Perhaps this really is her straight up natural voice. That doesn’t mean they didn’t look for and find a deeply similar voice.
I think this is disingenuous. There are a few possible explanations, but you've already taken intentionality off the table the moment you say "they likely did not tell Sky's VA to imitate anyone", which is the only thing that matters. I think it sounds more like Rashida Jones, and it might be in fact that people only think it sounds a lot like Scarlett because they've been primed to at this point. It's not illegal, and it's definitely not morally wrong.
This is starting to feel like a very biased axe to grind, which would be fine if you and Kelsey admitted it's about AI safety and not voice actor rights! Come on, you're better than this.
I'm not too upset about Scarlett Johansson, who I assume can take care of herself, but I am nonplussed about the safety team being gutted. I guess I should cancel my subscription and tell them why, but is there anything else I can do?
- Does moving from the free GPT 4o to Gemini or something have any impact?
- Are any of the competitors any better on safety?
Formatting error, when referencing "Neel Nanda (referencing Hilton’s thread):" the second tweet is not included in the quotation, so looks like it comes from you
The partisan split in those polls is super-concerning to me. I haven't seen much partisan split in these sorts of polls before, but that really is quite a hefty one. And it's the opposite way round to the kinds of noises we're hearing from leaders in those parties. That can't be a stable situation. The people (quite a few!) who say that the long term safety team shenanigans make them like OpenAI *more* are really concerning.
The one thing not quite factored into those polls is that people are reporting how they have *changed* their views, but not what their views were to begin with. For a sufficiently low baseline, additional bad news could be no worse than was already expected.
"First, the statement is not legally binding, as I understand it, without execution of a new agreement. No consideration was given, and this is not so formal, and it is unclear whether the statement author has authority in the matter"
No consideration is necessary, because it is to the advantage of the other side (the former employees, in this case). This is pretty legally binding. A judge would look very unsympathetically at OpenAI if they later tried to say "oh, wait, we take it back."
I use a mix of the 3 major AIs, for different tasks. If I was doing industrial-scale stuff I would consider switching for non-practical reasons, but I'm not going to sweat the $20.
The other question is complicated, but 'they are not known to do many of the things described in various posts' is certainly part of it.
Multi Voiced AI narration of this post. Every unique quoted person gets their own voice to distinguish them:
https://askwhocastsai.substack.com/p/openai-fallout-by-zvi-mowshowitz
You mean you didn't screen 400 voice actors and pick the ones who most closely matched the voice of the person being quoted? Lame.
I for one am honored that I am being played by Nick Offerman.
How much do these non-disparagement agreements really matter though? What smoking guns could there possibly be? Let's consider some hypothetical scenarios:
1. OpenAI secretly puts LSD into the water (aka CIA-level conspiracy) => someone would definitely report this, no matter what they've signed. As Bruce Schneider taught us, if you're committing a sufficiently large amount of crime and more than a handful of people know the details, someone will almost certainly rat on you. Hence, "agreements" don't matter for any major felonies - people will go ahead and report those, at the very least because doing otherwise means you'll eventually go to prison.
2. OpenAI is being "creative" around IP or otherwise does semi-shady dealings behind the curtain. Definitely possible but this is pretty much common knowledge already. AFAIK it was semi-confirmed that Sora was trained on Youtube videos, for example.
3. OpenAI doesn't really care about AI Alignment. Well... is this a surprise? Does it matter if 1, 5, 10 or 50 OpenAI employees say so on Twitter? Who will be persuaded by 10 employees stating this but not 1 employee (plus the entire AI safety community) saying this?
4. If Congress holds a hearing on AI safety and subpoenas former or current employees, all their NDAs will be pretty much void during the hearing, as a contract cannot demand that the other party breaks the law.
I guess all of this does matter to a small, highly elite group of people who are:
1. Talented enough to have been hired by OpenAI. I'd venture almost all are in the top 0.1%, if not 0.01% of all engineers worldwide.
2. Want to publicly criticize OpenAI.
3. Haven't sold their vested stock yet.
As a fellow engineer I'm more than sympathetic to my colleagues and I think every worker should get paid what they've been promised. But... these are the elite of the elite, the kinds of engineers who will find excellent, highly paid jobs no matter what. I truly hope Daniel Kokotajlo gets his money back but I'm also 99% sure that he can go get a job paying at least $600k/year tomorrow if he's inclined to do so.
I'm a simple person. I see "Bruce Schneider" and I hit the like button.
I agree with (1) that if OpenAI was doing sufficiently blatant crimes, or (e.g.) they know they have a malign AI that is in imminent dander of killing thousands of people, then the employees would have gone to law enforcement, regardless of any NDAs. So we can be fairly confident it's not that.
If, hypothetically,
a) a malign AI killsthousands of people
b) the owners of said AI knew this was likely to happen, i.e. a re negligent
then I think the government would be fully justified in coming down very hard on the people responsible
OpenAI employees surely know this would be a likely government response, whether they were bound by NDAs or not.
So it can't be that, because if it was, the ex employees would be complaining loudly to the FBI about how Sam Altman is going to kill thousands of people.
I'm curious in what sense they use the word "vested" if that doesn't mean they now own that equity and can do what they want with it- if it can't be sold now because OpenAI is not yet a traded company, at the very least I'd expect someone to say "you can't just take back my ownership of vested shares". Anyone who understands this aspect of law better have a comment?
I only ever worked for one startup and it was already publicly traded. I didn't stay long enough for my options to vest, as it was a 5-year period for someone in a junior position (First big boy job out of college doing gruntwork), but it was my understanding that once my stock options were vested, there was nothing they could do to revoke them.
Re 2: having ex employees state it publicly could well take it from "everyone suspects" rumour to "legally actionable".
Re 3: going from 1 to 50 might not matter that much but going from 0 to 1 really does - at least from OpenAI's ability to recruit talent - and they were very strongly trying to keep the number at 0
Unless you stand for VC in personal and professional realms of existence, you might not believe that Altman represents the best of humanity. “Don’t be evil”, right?
> As I covered last time, if you do a casting call for 400 voice actors who are between 25 and 45, and pick the one most naturally similar to your target, that is already quite a lot of selection. No, they likely did not explicitly tell Sky’s voice actress to imitate anyone, and it is plausible she did not do it on her own either. Perhaps this really is her straight up natural voice. That doesn’t mean they didn’t look for and find a deeply similar voice.
I think this is disingenuous. There are a few possible explanations, but you've already taken intentionality off the table the moment you say "they likely did not tell Sky's VA to imitate anyone", which is the only thing that matters. I think it sounds more like Rashida Jones, and it might be in fact that people only think it sounds a lot like Scarlett because they've been primed to at this point. It's not illegal, and it's definitely not morally wrong.
This is starting to feel like a very biased axe to grind, which would be fine if you and Kelsey admitted it's about AI safety and not voice actor rights! Come on, you're better than this.
I'm not too upset about Scarlett Johansson, who I assume can take care of herself, but I am nonplussed about the safety team being gutted. I guess I should cancel my subscription and tell them why, but is there anything else I can do?
- Does moving from the free GPT 4o to Gemini or something have any impact?
- Are any of the competitors any better on safety?
Formatting error, when referencing "Neel Nanda (referencing Hilton’s thread):" the second tweet is not included in the quotation, so looks like it comes from you
The partisan split in those polls is super-concerning to me. I haven't seen much partisan split in these sorts of polls before, but that really is quite a hefty one. And it's the opposite way round to the kinds of noises we're hearing from leaders in those parties. That can't be a stable situation. The people (quite a few!) who say that the long term safety team shenanigans make them like OpenAI *more* are really concerning.
The one thing not quite factored into those polls is that people are reporting how they have *changed* their views, but not what their views were to begin with. For a sufficiently low baseline, additional bad news could be no worse than was already expected.
"First, the statement is not legally binding, as I understand it, without execution of a new agreement. No consideration was given, and this is not so formal, and it is unclear whether the statement author has authority in the matter"
No consideration is necessary, because it is to the advantage of the other side (the former employees, in this case). This is pretty legally binding. A judge would look very unsympathetically at OpenAI if they later tried to say "oh, wait, we take it back."
Interestingly, despite claiming that "Sky" has been taken down, it's still available in the ChatGPT OpenAI app.
Do you still use ChatGPT? Do you use a different AI instead? What makes the other AI labs better?
I use a mix of the 3 major AIs, for different tasks. If I was doing industrial-scale stuff I would consider switching for non-practical reasons, but I'm not going to sweat the $20.
The other question is complicated, but 'they are not known to do many of the things described in various posts' is certainly part of it.
> Ozzie Gooen: On OpenAl's messaging:
Forcing my computer to use a font where the capital i and the lowercase ell look different is paying off!