Thread by @wolfejosh

1/ New paper from @ylecun et al on alternative approach for AI to learn more biologically...

paper basically says AI is super smart but still can't learn like a toddler can...

the main critique

the main critique...

is that top models today dont learn once deployed.

and curating data + training pipelines + filtering out noise is all done by humans in loop instead of being an intrinsic skill of the AIs...

here's Claude explaining the paper simply

3/ not quite Kahneman style but they call it System A +B

System A = learning from observation

this is self-supervised learning like predicting next word or video frame. good at making representations but passive and struggles to understand cause + effect...

4/ System B = learn by doing + action

reinforcement learning where agent tries stuff out to reach a goal. good for real-time adaptation but very slow + requires massive amounts of trial + error to learn simple tasks

5/ The missing puzzle piece to combine them they call System M (Meta-control) (probably also a jab at Meta;)

...acts like autonomous orchestrator + monitors internal signals like prediction errors +uncertainty and then decides whether AI observes, acts or dives into memory

6/ Like @sfiscience has published lots on

EXPLORE vs EXPLOIT

System M is what allows biological agents to switch up + decide when to explore vs exploit or even going "offline" to consolidate memories during sleep

@sfiscience 7/ @ylecun et al suggest an Evolutionary-Developmental (evo-devo) framework

an outer "evolutionary" loop optimizes starting parameters with a meta-controller

while an inner "developmental" loop lets the agent learn in real-time by interacting with environment

pretty cool