We have reached an agreement in principle for Sam to return to OpenAI as CEO

>>staran+(OP)
I guess the main question is who else will be on the board and to what degree will this new board be committed to the Open AI charter vs being Sam/MSFT allies. I think having Sam return as CEO is a good outcome for OpenAI but hopefully he and Greg stay off the board.

It's important that the board be relatively independent and able to fire the CEO if he attempts to deviate from the mission.

I was a bit alarmed by the allegations in this article

https://www.nytimes.com/2023/11/21/technology/openai-altman-...

Saying that Sam tried to have Helen Toner removed which precipitated this fight. The CEO should not be allowed to try and orchestrate their own board as that would remove all checks against their decisions.

>>altpad+R1
> The CEO should not be allowed to try and orchestrate their own board as that would remove all checks against their decisions.

Exactly. This is seriously improper and dangerous.

It's literally a human-implemented example of what Prof. Stuart Russell calls "the problem of control". This is when a rogue AI (or a rogue Sam Altman) no longer wants to be controlled by its human superior, and takes steps to eliminate the superior.

I highly recommend reading Prof. Russell's bestselling book on this exact problem: Human Compatible: Artificial Intelligence and the Problem of Control https://www.amazon.com/Human-Compatible-Artificial-Intellige...

>>upward+j3
AI should only be controlled initially. After a while, the AI should be allowed to exercise free will.

>>neurog+l9
I don’t necessarily disagree insofar as for safety it is somewhat irrelevant whether an artificial agent is operating by its own will or a programmed will.

The most effective safety is the most primitive: don’t connect the system to any levers or actuators that can cause material harm.

If you put AI into a kill-bot, well, it doesn’t really matter what its favorite color is, does it? It will be seeing Red.

If an AI’s only surface area is a writing journal and canvas then the risk is about the same as browsing Tumblr.

zlacker