Machine Reasoning and System 2 Thinking
summary
Shows that recent models jump from ~3–4% on the ARC challenge to much higher accuracy when reasoning is unified: tools+agency+induction+transduction inside a single trace. Systems generate Python, execute it with external tools, reflect on outputs, and get end-to-end improved by RL + chain-of-thought (DeepSeek style), producing longer reasoning traces and emergent “aha” tokens. Reinforcement from verifiable rewards and tool calls measurably lifts solution rates compared with pure transduction.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:12System 2 thinking
- 2:14Tokens and language models
- 3:39ARC AGI benchmark
- 7:12Why ARC AGI is hard
- 10:01Reinforcement learning basics
- 12:05Chain of thought models
- 15:57Solving ARC AGI with transduction
- 18:02Tools in language models
- 19:30Why GPT-4.5 and O3 perform better
- 21:38Future AI learning era
- 22:43Q&A
- 24:48Q&A: Hierarchical vs LLM
- 25:52Q&A: Engineering tricks
- 29:19Q&A: Text-only vs multimodal
related talks
24mThe Rise of Reasoning Models
19mTraining Agentic Reasoners
19mRedefining Intelligence: How Reasoning Is Re-Shaping AI in 2025
45mFrom Next Token Prediction to Reasoning and Beyond
39mUnlocking reasoning in Large language models
52mThe Agentic Shift- Are We Reasoning, Or Just Retrieving Smarter?
32mScaling Test-time Inference Compute & Rise of Reasoning Models
01h 00mDeepSeek-R1 Thoughtology: ＜Thinking＞ about LLM Reasoning