What a quick and intelligent move by Satya! I wonder what might have been offered to Sam to keep him as a Microsoft employee. Whatever it is, perhaps he now has all the financial resources to build a chip and compete with NVIDIA, or even develop AGI (which is not good imo).
The irony here is the mission of OpenAI is to keep AI free and open to everybody and not concentrated in the large corporations like Google & Microsoft and now Sam Altman finds himself at Microsoft.
This seems the most likely (and markets mostly agree). He was trying to oust board member(s) over an academic piece that called out OAI trailing Anthropic in some safety aspects (hardly a controversial view).
@Cameron has posted the most likely thing to surface to date. While there are probably NDAâs there was a surprising lack of back channel / off the record comments. Kara Swisher seemed to speak to almost everyone involved and didnât get this info.
Remember we are talking about a private company. A company not listed on the stock exchanges.The board owed/owes none of us anything. Its structure ensures that. It is wholly irrelevant what Kara Swisher or anyone of that ilk thought/thinks.
Most of us who work in the arena had an inkling of what was going on (and had known for weeks or even months). Of course some of us did not know all the details but we were well aware of several problematic issues. We also are well aware that the âmastermindâ of the technology never was the CEO. Admittedly he is highly respected - but not as the tech master.
I doubt that was the cause (the above market hasnât updated much on it) also the article doesnât really make sense - is Q* supposed to be a new training algorithm or a new model ? They write about it in both ways in terms of training efficiency and capabilities.
Q-Star sounds like a reinforcement learning (RL) algorithm to me (Q-learning + A-Star ?). RL is used in the last step of training ChatGPT (and other chatbots) to train the system to maximize user satisfaction in a sense (RLHF). Some believe that RL is strictly necessary for a model to be able to extrapolate vs only interpolate. The current RLHF methods used in practice are very very basic, afaik, so I can imagine that if they managed to make something more complex work well, that could be huge for A(G)I.
But equally, it could be an AI Doomer over-interpretation of some basic RL thing they were doing, who knows.