Powered flight was an important goal, but that wouldn’t have justified throwing all the world’s resources at making Da Vinci’s flying machine work. Some ideas are just dead ends.
Transformer based generative models do not have any demonstrable path to becoming AGI, and we’re already hitting a hard ceiling of diminishing returns on the very limited set of things that they actually can do. Developing better versions of these models requires exponentially larger amounts of data, at exponentially scaling compute costs (yes, exponentially… To the point where current estimates are that there literally isn’t enough training data in the world to get past another generation or two of development on these things).
Whether or not AGI is possible, it has become extremely apparent that this approach is not going to be the one that gets us there. So what is the benefit of continuing to pile more and more resources into it?
Powered flight was an important goal, but that wouldn’t have justified throwing all the world’s resources at making Da Vinci’s flying machine work. Some ideas are just dead ends.
Transformer based generative models do not have any demonstrable path to becoming AGI, and we’re already hitting a hard ceiling of diminishing returns on the very limited set of things that they actually can do. Developing better versions of these models requires exponentially larger amounts of data, at exponentially scaling compute costs (yes, exponentially… To the point where current estimates are that there literally isn’t enough training data in the world to get past another generation or two of development on these things).
Whether or not AGI is possible, it has become extremely apparent that this approach is not going to be the one that gets us there. So what is the benefit of continuing to pile more and more resources into it?