3 ms·

What I personally think is that maybe these AI are being really inefficient with their training and still capable of producing all of this wealth. As we discov

What I personally think is that maybe these AI are being really inefficient with their training and still capable of producing all of this wealth. As we discover newer and better architectures, which I believe we will, then the machines are going to become better and better at doing this what they are doing. So in other words, I think we already have enough data for most things. Certainly for art and code, writing, etc.

didibus 4y ago
I think one issue is that the models learn by induction, not by deduction. That means they learn the rules from the examples, but cannot reason about rules deductively. Arguably, it means even if you were able to assemble the smallest amount of data that contained the entire array of scenarios, from which you'd learn the proper rules, it's unclear how small that would be, even assuming the most optimal learning model. As I see it, we currently have good machine inductive learning, but are really far from any machine deductive learning.
- deltasevennine 4y ago
  Introducing causality into the model could very well just be another layer. In fact it is quite clear that stable diffusion produces a super set that is slightly bigger than solutions with causal reasoning taken into account. One more additional filter on top of that for causal deduction is a very realistic next step. In a nutshell stable diffusion is like a calculator that gives me 3 answers. One of those answers is right, the other 2 are wrong. The next step in calculator creation is putting a filter over the output to get the right answer from the 3 possibilities. >As I see it, we currently have good machine inductive learning, but are really far from any machine deductive learning. Yeah, we already covered 50% of the AI equation in the past decade. The other deductive part is actively being researched and a revolution in your lifetime is HIGHLY probable given how fast the inductive part came to fruition.

didibus 4y ago

I think one issue is that the models learn by induction, not by deduction. That means they learn the rules from the examples, but cannot reason about rules deductively. Arguably, it means even if you were able to assemble the smallest amount of data that contained the entire array of scenarios, from which you'd learn the proper rules, it's unclear how small that would be, even assuming the most optimal learning model. As I see it, we currently have good machine inductive learning, but are really far from any machine deductive learning.

deltasevennine 4y ago
Introducing causality into the model could very well just be another layer. In fact it is quite clear that stable diffusion produces a super set that is slightly bigger than solutions with causal reasoning taken into account. One more additional filter on top of that for causal deduction is a very realistic next step. In a nutshell stable diffusion is like a calculator that gives me 3 answers. One of those answers is right, the other 2 are wrong. The next step in calculator creation is putting a filter over the output to get the right answer from the 3 possibilities. >As I see it, we currently have good machine inductive learning, but are really far from any machine deductive learning. Yeah, we already covered 50% of the AI equation in the past decade. The other deductive part is actively being researched and a revolution in your lifetime is HIGHLY probable given how fast the inductive part came to fruition.