I propose a layercake model of AI. An AI consists of 0 or more layers of general optimizer, followed by 1 layer of specific tricks.
(I won’t count the human programmer as a layer here)
For example, if you hardcoded an algorithm to recognize writing, designing algorithms by hand, expert system style, then you have 0 layers of general optimizer.
If you have a standard CNN, the gradient descent is an optimization layer, and below that is specific details about what letters look like.
In this picture, there is a sense in which you aren’t missing any insights about intelligence in general. The idea of gradient descent is intelligence in general. And all the weights of the network contain is specific facts about what shape letters are.
(Although these specific facts are stored in a pretty garbled format. And this doesn’t tell you which specific facts will be learned)
If you used an evolutionary algorithm over Tensor maths, and you evolved a standard gradient descent neural network, this would have 2 general optimization layers.
If neural networks become more general/agentic (As some LLM’s might already be a little bit) then those neural nets are starting to contain an internal general optimization algorithm, along with the specifics.
This general algorithm should be of the same Type of thing as gradient descent. It might be more efficient or have more facts hard coded in or a better prior. It might be insanely contrived and complicated. But I think, if we had these NN found algorithms, alignment would still be the same type of problem.
I propose a layercake model of AI. An AI consists of 0 or more layers of general optimizer, followed by 1 layer of specific tricks.
(I won’t count the human programmer as a layer here)
For example, if you hardcoded an algorithm to recognize writing, designing algorithms by hand, expert system style, then you have 0 layers of general optimizer.
If you have a standard CNN, the gradient descent is an optimization layer, and below that is specific details about what letters look like.
In this picture, there is a sense in which you aren’t missing any insights about intelligence in general. The idea of gradient descent is intelligence in general. And all the weights of the network contain is specific facts about what shape letters are.
(Although these specific facts are stored in a pretty garbled format. And this doesn’t tell you which specific facts will be learned)
If you used an evolutionary algorithm over Tensor maths, and you evolved a standard gradient descent neural network, this would have 2 general optimization layers.
If neural networks become more general/agentic (As some LLM’s might already be a little bit) then those neural nets are starting to contain an internal general optimization algorithm, along with the specifics.
This general algorithm should be of the same Type of thing as gradient descent. It might be more efficient or have more facts hard coded in or a better prior. It might be insanely contrived and complicated. But I think, if we had these NN found algorithms, alignment would still be the same type of problem.