Some thoughts on the Sutton interview
- Source
- dwarkesh.com
- Author
- Dwarkesh Patel
- Date
Context
Learning from human examples and learning from experience could be complementary steps in building AI. That is Dwarkesh Patel's central argument in his reflection on the Richard Sutton interview. could give a model enough initial knowledge to attempt useful actions, after which feedback from their outcomes could improve it further. Patel accepts persistent learning as an important unresolved problem but disputes the that language models must therefore be abandoned. He presents possible ways to add such learning as research ideas, leaving their effectiveness open.
- pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
- inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
“Imitation learning is just short horizon RL. The episode is a token long.”
Dwarkesh Patel
“As planes are to birds, supervised learning might be to human cultural learning.”
Dwarkesh Patel
“It’s a bit like saying to someone pasteurizing milk, “Hey stop boiling that milk - we eventually want to serve it cold!””
Dwarkesh Patel
“Evolution does meta-RL to make an RL agent. That agent can selectively do imitation learning.”
Dwarkesh Patel
“If the LLMs do get to AGI first, the successor systems they build will almost certainly be based on Richard’s vision.”
Dwarkesh Patel
Checking sign-in…
Loading comments…