What should the EA/AI safety community change, in response to Sam Altman's revealed priorities?
By SiebeRozendal @ 2024-03-08T12:35 (+30)
Given that: Altman is increasingly diverging from good governance/showing his true colors, as demonstrated in his power play against board members & his current chip ambitions.
Should there be any strategic changes?
Jason @ 2024-03-08T16:26 (+28)
[Also informed by FTX]
Epistemic status: some of these assertions are lightly held, and it's possible they may not apply to specific situations that well. Analyzing someone's motives can be challenging, and I have zero inside knowledge of dynamics at OpenAI.
- Money, power, fame, etc. are powerful motivating -- and often corrupting -- factors in human motivation.
- A model that depends on a single person (or a small group of interconnected people) being able to resist the allure of money, power, fame, etc. because of their virtue, value alignment, etc. rests on awfully shaky ground. I get the sense that EAs are, as a whole, too ready to assume that other EAs have low susceptibility to corruption from these sorts of influences.
- (Based on a low n) Assessment of a person's ability to resist money, power, fame, etc. may be significantly more difficult when that person's track record is fairly limited. In the non-EA world, it would be seen as crazy to give someone who is ~30 years old a role that could have a major impact on humanity's future. In those worlds, people are -- at least in theory -- only given roles of extraordinary trust after decades of proving themselves capable of having resisted corrupting influences to some degree. Although personality/identity change continues throughout the life course, my (simplified) understanding is that significant changes in the 20s are not uncommon.[1]
- At the risk of quoting the Bible in the EA Forum: "Do not be misled: 'Bad company corrupts good character.'" Or as most parents of teenagers know: if your kid starts closely hanging out with people who are bad news, odds are that your kid's behavior will start changing rather than your kid having a salutary effect on the others. Microsoft may not be the evil empire anymore, but it is still a profit-seeking publicly-held corporation, and we all know what types of incentives such corporations have. I obviously can't speak for each OpenAI employee who assisted Altman in overthrowing the board, but it seems likely that those employees' personal interests would align with working with an organization that is "winning" the AI race and becomes exceptionally profitable/powerful/famous ASAP.
What to do about these truths is a more difficult question. There are probably some low-hanging fruit: when making a decision, model what would happen if a core person or interconnected core group goes rouge. Incorporate the risk of empowering people who may fall prey to the allure of money, fame, power, etc. into one's assessments. But strongly mitigating against the allure of money, power, fame, etc. when trying to get stuff done -- especially when trying to get stuff done quickly -- is rather difficult.
- ^
This is not a 41-year old asserting that we shouldn't trust anyone under thirty! I am, however, asserting that we have much less data about their resilience to the allure of money, power, fame, etc. than we have for a 50-year old who has occupied positions over a period of time with an increasing exposure to that allure. Risk models should reflect that increased uncertainty.
trevor1 @ 2024-03-10T05:22 (+5)
I would like to add that certain types of people might be predisposed towards power seeking (and succeeding at power seeking), rather than just being corrupted by power, status, money, or fame.
Social Dark Matter offers some interesting takes on this; it's more nuanced than it appears e.g. neurotic people might even be more reputation-obsessed but also potentially more likely than the median human to internalize moral values (or, in the case of EA, commit to internalizing moral values in a lasting way). This is purely speculative food for thought to illustrate the complexity of this situation (empirically researching the psychology of different kinds of powerful people is difficult due to nonresponse bias making samples disproportionately stacked towards people who aren't as powerful as they look).
Ben_West @ 2024-03-12T20:46 (+4)
Are there statements or predictions which you believe have been falsified by recent events because of these four points?
The things you are saying all sound to me like things people would have agreed with in 2015. The oldest EA-ish post I can find about OAI specifically concludes that the author is distrustful of Sam (though not for the reasons you state). This 2017 post says:
I’ve talked with a lot of people about this in the AI risk community, and they’ve often attempted to steelman the case for OpenAI, but I haven’t found anyone willing to claim, as their own opinion, that OpenAI as conceived was a good idea.
Prominent safety-oriented people have left OAI (Paul Christiano, Dario Amodei, etc.), which maybe could be interpreted as them having been mistaken about what to expect at OAI, but I'm not sure to what extent the four points you listed above were causal in their decisions.
It feels to more to me like people were aware that their priorities differed from OAI's, but thought it was still net positive to work there. And I'm not sure that any of them actually disagree with that take now? I'd be interested to hear comments from those more knowledgeable than me.
Jason @ 2024-03-13T14:50 (+2)
Thanks. I would be interested on others' thoughts on that point as well.
To clarify, I didn't read the post's question as "What did EA do wrong with OpenAI?" but instead as "What should we learn from Altman 'showing his true colors'?" So it wasn't intended as a postmortem of OpenAI per se. I strongly suspect that most EAs would have agreed with my general points in 2015 . . . but how much weight did they give them versus other considerations? I wasn't there, so will have to rely on others to say.
Looking at the posts you shared, both seem to take Altman and OpenAI mostly at their world that the mission was . . . open (as opposed to proprietary) AI. That doesn't seem to be the mission anymore, and I'd argue OpenAI is more dangerous as a result.[1] So it seems there may have been inadequate consideration of the principles I listed in analyzing the risk of that particular failure mode. Given that OpenAI was primarily funded elsewhere, it's not clear how much EA could have done about this risk, other than perhaps being more wary about encouraging people to work for/with OpenAI.
- ^
Whatever the downsides of publicly releasing AI information, an organization focused on that mission isn't likely to rake in the massive amounts of $$$ needed to be where OpenAI is now on capabilities. At some point, OpenAI pivoted to acting much more like any for-profit company would be expected to act, as opposed to an organization acting in the public interest.
SiebeRozendal @ 2024-03-08T17:01 (+4)
I agree with your points in general, but I'm confused at the context. Are you implying here that EA empowered and trusted Sam Altman?:
I get the sense that EAs are, as a whole, too ready to assume that other EAs have low susceptibility to corruption from these sorts of influences.
Jason @ 2024-03-08T17:20 (+4)
That [edit: sentence] wasn't specific to Altman; as the top of post notes, these propositions were influenced by both the FTX debacle and Altman. The lack of accountability and oversight at certain EA organizations also contributed to that [edit: sentence], which was intended to assert that understating corruption risk from insiders may be a blind spot for EA more generally.
That being said, I think there are some data points which could suggest that EAs and EA-related entities should have been more skeptical of Altman and OpenAI than they were. As this post notes, there was a large grant to OpenAI in 2017, and as of 2021 their non-safety related jobs were still being promoted on the 80K job board even after some worrisome signs of Altman's turn in focus. I've also seen several AI-focused posters opine that EA's involvement with OpenAI has been one of the movement's significant failings, which could suggest an error in trusting Altman's motives.
SiebeRozendal @ 2024-03-08T19:18 (+5)
Okay, got it!
The grant was also to buy a board seat, which makes it very different from a normal grant.
The 80K job promotion is indeed odd.
I think there was plenty of skepticism towards OpenAI, but maybe less so at the top
Ian Turner @ 2024-03-08T17:10 (+1)
That is explicitly true, no? Open Philanthropy was an early OpenAI donor.
Maynk02 @ 2024-03-09T21:27 (+10)
I might be untapped with the latest update here, but did they release the reason behind Altman's firing? I don't think it was ever answered by him in the subsequent interviews. Gradually questions died down or probably dropped from the questioner's list due to a clause, maybe. Now that he is back at the table,[1]I think it has become more urgent to get the original motivations out.
SiebeRozendal @ 2024-03-10T07:43 (+5)
No they didn't, and it looks like we aren't going to see the investigation, unless somebody leaks it. But it looks to me that it had something to do with his pattern of manipulative behavior, and allegedly he lied to other board members that McCauley wanted Toner fired (this was stated in the NY Times article on Murati, I think), which sounds like the proximate cause to me.
But if such behavior came up during the investigation, I'm confused how the investigators could NOT conclude there was good reason for his firing (maybe they're not so independent?) or why the board didn't say something like "Mr. Altman was attempting to get a board member fired by providing false information" (too risky for libel?). Maybe he lied to Sutskever or Brockman, and they didn't want to corroborate it? Questions, questions..
Nick K. @ 2024-03-10T08:04 (+7)
From gwern's summary over on lesswrong it sounds like the actual report only stated that the firing was "not mandated", which could be interpreted as "not justified" or "not required". Is it clear from the legal context that the former is implied?
Jason @ 2024-03-10T14:31 (+6)
I'm pretty sure WilmerHale chose that word carefully. I do not see an implication that it was not justified. Moreover, as the statement says, boards have very broad discretion over this kind of stuff. So "not mandated" says very little indeed!
SiebeRozendal @ 2024-03-10T11:20 (+3)
Ah yes, that's a great summary I hadn't read yet. Link: https://www.lesswrong.com/posts/KXHMCH7wCxrvKsJyn/openai-facts-from-a-weekend?commentId=eFuasCwaKJr2YiScY
And it looks like likely that phrase actually meant "not required"!
Maynk02 @ 2024-03-10T12:06 (+1)
I am not sure if leaks are a reliable source in these cases. For one, these instances don't have material evidence. Somebody (or a bunch of somebodies) can only try to come forward to take action. But I am afraid that's what they tried to do. It was like the first necessary crisis (the sooner, the better) for later events to unfold. I am unsure about their nature. Partially based on the new board's current update on choosing the new members.
No they didn't, and it looks like we aren't going to see the investigation, unless somebody leaks it.
SiebeRozendal @ 2024-03-10T16:18 (+2)
Sorry I don't follow - not sure whom you refer to by "they" and "their"
Maynk02 @ 2024-03-10T17:57 (+1)
Somebody (or a bunch of somebodies) can only try to come forward to take action. But I am afraid that's what they tried to do.
Here, "they" refers to folks from OpenAI who tried to come forward and do something about Sam's manipulative behavior or lies or whatever was happening. Anyone who may potentially provide the leaks or shed some light.
It was like the first necessary crisis (the sooner, the better) for later events to unfold. I am unsure about their nature.
Here, I am unsure about the nature of the events.
I hope it is clear now.