Question 1

What are the core limitations of transformers and backpropagation for building AGI?

Accepted Answer

_The train-then-freeze design of today's models is the wrong shape for a mind that keeps learning._

> Because of the underlying backpropagation algorithm, learning and inference are separate phases. You train a model, you freeze it, and then you use it. If you try to fill the network up with knowledge and keep going, at some point you get catastrophic forgetting: adding a little new knowledge causes it to forget a lot of old knowledge. These systems know a lot, but they don't reorganize themselves in real time the way parts of the human brain do.

Source: [Rethinking the AGI Race, with Benjamin Goertzel (eCornell Keynote)](https://ecornell.cornell.edu/keynotes/view/K052826/)

Question 2

What is predictive coding and why does it matter for AGI?

Accepted Answer

_Changing the learning rule, not just the architecture, may be what unlocks continual learning._

> There's something called predictive coding which more closely resembles how the brain works. On smaller neural nets, predictive coding networks are better at continual learning: they don't have as much catastrophic forgetting, each neuron can learn asynchronously and independently, and you can do inference and learning in the same phase. The need for hyperscaler server farms is less, because the learning is asynchronous.

Source: [Rethinking the AGI Race, with Benjamin Goertzel (eCornell Keynote)](https://ecornell.cornell.edu/keynotes/view/K052826/)

Q&A, continual learning

What are the core limitations of transformers and backpropagation for building AGI?

What is predictive coding and why does it matter for AGI?