How AI and Human Cognition is Fundamentally Similar | Deep Analysis

Considering the evident confidence radiating from OpenAI in reaching AGI with o1-type models (see    • Why o1 is a BIG deal  , the contemplation of how similar these models are to human cognition has become more important than ever. If we could show that they are fundamentally alike in functioning and training, there would be no apparent reason to believe that they couldn't surpass humans in cognitive work.

To illustrate this, I relate concepts from cognitive science to concepts from machine learning to show their fundamental similarities. I'll use the concept of system 1 and 2 thinking to structure this explanation:
System 1 (intuitive): GPT-4, Claude, etc.
System 2 (reasoners): The new test-time compute models like o1.

I'll end the video with a brief statement of the recent scaling law limits. Ilya Sutskever commented on these limits: “Scaling the right thing matters more now than ever.” As I'll show, the scaling law limits are by no means a dead-end. Instead, it signifies that we have reached System 1's full potential while we are just at the beginning of scaling System 2.

Timestamps:
00:00 - Introduction
00:35 - System 1: Hebbian vs. Unsupervised Learning
01:26 - System 1: Generalization
02:50 - System 1: ID & OOD Data
04:01 - System 1: Cortical Columns vs. Self-supervised
05:33 - System 1: Imagination vs. Generation
07:11 - System 2: CoT + ToT
08:51 - System 2: Limbic System vs. RL
09:38 - System 2: RL-HF
10:07 - System 2: o1 / evaluating the CoT
11:20 - System 2 is just System 1 iterated
11:53 - Scaling Law Limits
13:05 - Summarizing Table & The Importance of Theory


Music by: Bensound.com/royalty-free-music
License code: VJIWLQZMMKYLXHWX