Скачать или смотреть On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning LLM: Is AI actually thinking 🤔

Is AI actually thinking—or just really good at guessing? 🤔
https://www.emergent-behaviors.com/th...

In this explainer video overview (#MadeWithNotebookLM) we’re unpacking "The Recipe for AI Reasoning." We head inside the CMU AI Laboratory to see how researchers are moving past the "messy internet" to build a controlled world for synthetic reasoning. 🧪✨

From the "Goldilocks Zone" of difficulty to the "1% Seed" rule, we break down the scientific blueprint for building smarter, more logical models.
What’s inside the recipe? 👩‍🍳

📈 The Sweet Spot: Why training on "too easy" or "too hard" tasks fails.

🏗️ Mid-Training: The secret bridge between general knowledge and expert skill.

🤖 Anti-Cheating: How "Process Rewards" stop AI from hacking the system.

📍 Chapters

00:00 — 🥣 The Recipe for AI Reasoning

00:13 — 🧠 Logic vs. Pattern Matching: The Big Debate

00:33 — 🔄 Refiners vs. Extenders

00:58 — 🔬 Inside the CMU AI Lab

01:43 — 🧩 What are Synthetic Reasoning Tasks?

02:06 — 📏 Depth vs. Breadth in AI

02:25 — 🐻 Clue 1: The Goldilocks Zone

03:20 — 🌱 Clue 2: Planting the 1% Seed

04:17 — 🌉 Clue 3: The Underrated Mid-Training Step

05:35 — 📝 Clue 4: Show Your Work (No Reward Hacking!)

06:26 — ⚙️ The Interplay: Pre-training, Mid-training, & RL

06:42 — 🏆 The Final Verdict: 3 Essential Ingredients

CMU: On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
https://arxiv.org/pdf/2512.07783
Charlie Zhang, Carnegie Mellon University, Language Technologies Institute, [email protected]
Graham Neubig, Carnegie Mellon University, Language Technologies Institute, [email protected]
Xiang Yue†, Carnegie Mellon University, Language Technologies Institute, [email protected]


#AI #MachineLearning #LLM #ArtificialIntelligence #DataScience #CMU #TechExplained #NeuralNetworks #FutureOfTech #ComputerScience #deeplearning