He had dual overlapping roles, a duty to the promote safe AI to the world, and a duty to prevent own-goal sabotage to the LLC. Ideally those roles wouldn’t conflict. He probably didn’t care if such a paper was written, but to have a board member as author is beyond awkward. But as an academic, her own dual overlapping role, she has a need to publish. (“Uh, please voice your concerns before we develop it instead of sand bagging us with a damning critique in an academic journal after the fact!”) This seems like a recipe for disaster. I’m not certain either really was “wrong” in that conflict. It just seems like this structure was doomed to failure, throwing uncompromising zealots together all with launch codes and a big red button.

Safety isn’t binary, it’s degrees of risk assessment, mitigation and acceptance. Infinite safety is never progressing. But never progressing means failing the charity mission. And AI isn’t just being created at OpenAI, for them to succeed at the charity mission they need to stay ahead of risky competitors with a safer AI.

The charity can’t afford its GPU costs without the LLC, and the LLC can’t lead the industry if it’s behind all the competitors. To be relevant the safety side needs compute, and they need to be ahead of the curve, and they also need real world data from the LLC. If the charity nukes the LLC, it takes itself with it and fails its mission. They’re so intertwined they need compromise in the board. And they let that board dwindle to small entrenched factions, which sounded more like the Hatfield’s and the mcCoys. With such a fragile structure they needed mature adults with functional conflict resolution.

No, Sam wanted to be the first recipient of Helen's expression.

So -- if appropriate -- he could make changes.

Or he could not change, and she publishes anyway.

>> Sam wanted to restrict Helen's free expression.

> No, Sam wanted to be the first recipient of Helen's expression.

If I get copy approval to what you write, that is by definition a restriction to your freedom of expression. How is this up for debate?

You don't normally perform you role as a board member through publications in a public forum. You do it by exercising pressure internally to get the effect you desire so you can continue to do so in the future. Going public is usually a one-shot affair.

Being a co-author of that publication was not her performing her role as an OpenAI board member, it was her day job.

Conflicts of interest suck. Experienced board members avoid them or deal with them. They don't allow their conflicts of interest to spiral out of control to the point that they destroy the thing they're supposed to govern.

> Conflicts of interest suck. Experienced board members avoid them or deal with them.

Agreed. It seems remarkable that she dealt with her conflict of interest so thoroughly that she published an article that was mildly critical of the organization. I do not think her duty was to let Sam Altman seize power by firing her from the board on a flimsy pretext, though.

She went a bit further than that.

Did Helen's publication conflict with the charter though?

I'd say the OpenAI employees are the ones with the conflict of interest, since they stand to get rich if the company does well, independent of whether the charter is being upheld.

No, it did not conflict. But that doesn't mean you first try to get things fixed rather than to publish.

> I'd say the OpenAI employees are the ones with the conflict of interest, since they stand to get rich if the company does well, independent of whether the charter is being upheld.

Those pesky employees again. Too bad we still need them. But after AGI is a fact we don't and then it'll be a brave new world.

It’s interesting that people are making the assumption Toner didn’t raise these concerns internally as well.

My interpretation of that was that Sam felt Helen's duty as a board member is to come to the company with her complaints, and work to improve the safety measures.

> So -- if appropriate -- he could make changes.

Where are you drawing this conclusion from? Nothing in the text suggests that was his intent - certainly his actions thereafter suggest that he was in fact more concerned with the potential harm to the for-profit company.

The journalist failed to make that point really clear. He "reprimanded" her "and" said "it was dangerous to the company". What exactly did he "reprimand" her for? The "and" seems to imply two separate points of criticism.

>In the email, Mr. Altman said that he had reprimanded Ms. Toner for the paper and that it was dangerous to the company, particularly at a time, he added, when the Federal Trade Commission was investigating OpenAI over the data used to build its technology.

The normal operation of every human relationship.

Let's recall the CEO is employed by the board, not the other way around.

Is Sam a member of her faculty? Or an advisor?

Otherwise I'm not sure what changes would be "appropriate" for him to direct.

> That suggests he cares more about the success of OpenAI than he does upholding the charter.

A board member should care about their organization more than their interpretation of the charter. If a Mozilla board member publicly criticized Firefox for violating their charter of a free and open internet because they're receiving money from Google, while also praising Brave for not doing so, I guarantee that they'd also be ousted.

I would have thought a non-profit board that puts the growth and success of their organization above the mission in their charter, even if they view the organization as going against their mission, sounds like the worst kind of board. How could that possibly be a good idea?

Isn't the whole reason we have non-profit boards and charters is so that they can make just this kind of call?

(I'm not saying that this is what is happening here -- just responding to your particular claim.)

>Isn't the whole reason we have non-profit boards and charters is so that they can make just this kind of call?

Yes. From the board. That's why we typically don't see boards air their dirty laundry out in the open even when in retrospect, they obviously had a problem with leadership.

The board seems to be getting an awful lot of crap for not saying enough, and also getting crap because their initial letter was too direct rather than being the usual corporate mealy-mouth BS.

The board is getting a lot of crap for being crap. Regardless of your opinion about their goals, their execution was objectively terrible. They were clearly completely detached from their organization. At the very least, they could have made sure their first interim CEO wouldn't instantly try to turn against them.

It's not a contradiction to criticize someone for saying too much sometimes and saying too little another time. The board should be open to insiders and cautious with external communications.

It's possible to simultaneously believe that announcing that the CEO lied to the board about x on day y is better than announcing that the CEO is leaving to spend more time with his family, but announcing that the CEO lied to the board and then refusing to elaborate when everyone including the CEO says "huh? What?" is worse than either.

This is exactly backwards. The reason we create organizations like this is not for the benefit of the organization itself, but to accomplish specific goals. Those goals are outlined in the charter. If the organization starts doing things that conflict with the charter, the responsibility of the board is to course-correct.

The place to course correct is from the board room, not in public. If she had enough power to oust the CEO, she should have enough power to pressure him to make changes.

The board can't compel the CEO to do specific things, only replace them.

Of course they can. If Sam Altman is really as heedless about AI safety as the board seems to think, GTP wouldn't be so aggressively aligned.

A misaligned AI could act aligned in the service of its interests, just like an employee trying to steal from a company would act like a good honest employee. Methods for getting AIs to visibly act a certain way aren't necessarily sufficient to control advanced AI systems.