https://arxiv.org/abs/2111.08171

So basically with a little prompt engineering Codex can solve college level probability and linear algebra problems. Codex won't directly or immediately cause tech unemployment. that's not the point anyway. the implication of it is that the neural network scaling hypothesis is true and that NNs, when of sufficiently large size and trained on a wide range of experiences, start meta-learning. no need to study the brain in detail to figure out what the neocortex is doing and how to replicate it in silicon - just train NNs on large datasets and the neural networks will meta-learn powerful algorithms that generalize.

timelines to AGI are increasingly short.

(http://www.autoadmit.com/thread.php?thread_id=4973764&forum_id=2...#43518169)

not to mention transformers are also work for image generation, reinforcement learning, image recognition, etc. if we were having to create highly specific models to solve certain tasks, that would be evidence in favor of the view that AGI will require substantial engineering effort. we are seeing the exact opposite. even a simple highly regularized MLP can accomplish what a transformer can when trained on enough data.

(http://www.autoadmit.com/thread.php?thread_id=4973764&forum_id=2...#43518325)

What is the missing piece that's been filled in that answers questions like those raised by Pearl http://ftp.cs.ucla.edu/pub/stat_ser/r475.pdf

Jordan https://hdsr.mitpress.mit.edu/pub/wot7mkc1/release/9

or Hanson https://spectrum.ieee.org/economics-of-the-singularity

(http://www.autoadmit.com/thread.php?thread_id=4973764&forum_id=2...#43518398)