The 12,000 Token Secret of Reasoning AI (Inference Scaling)

Описание: What is test-time compute, and how do reasoning models like OpenAI's o3 and DeepSeek R1 actually think? In this video, we visualize the exact mathematics behind Large Reasoning Models (LRMs) and inference scaling using Manim animations.

For years, Large Language Models (LLMs) relied on System 1 thinking,guessing the next word instantly. Today, models use System 2 thinking, burning thousands of "hidden tokens" at test-time to build massive internal search trees before they answer. We break down the math of how Chain-of-Thought evolves into Monte Carlo Tree Search (MCTS), how Process Reward Models (PRMs) prune incorrect logical branches, and why the Snell et al. scaling laws prove that thinking longer is mathematically superior to building bigger models.

If you want a visual, step-by-step breakdown of the algorithms powering the next generation of AI, this is for you.

⏱️ Chapters:
0:00 The Paradigm Shift (System 1 vs. System 2)
3:55 The Limits of Next-Token Prediction
8:09 The Search Tree (The "Aha" Moment)
12:26 Process Reward Models
17:42 Monte Carlo Tree Search
22:17 Test-Time Compute Scaling Laws
25:20 The Cost of Thinking

▶ WATCH NEXT: "How AI Actually 'Thinks'"; our earlier introduction to test-time compute and MCTS reasoning, if you want the shorter on-ramp before this deep dive.
🔔 Subscribe for the next deep dive into the mathematics running underneath modern AI.

#TestTimeCompute #ReasoningModels #ArtificialIntelligence #MachineLearning #DeepSeekR1