“People don’t like hearing this but this is the reason why startups don’t hire big tech folks.” made the same learnings. As a small company you want people that get sh*t done. Folks from Google & co usually aren’t that.

SFT Memorizes, RL Generalizes. New Paper from @GoogleDeepMind shows that Reinforcement Learning generalizes at cross-domain, while SFT primarily memorizes. rule-based tasks, while SFT memorizes the training rule. 👀 Experiments 1️⃣ Model & Tasks: Llama-3.2-Vision-11B; GeneralPoints (text/visual arithmetic game); V-IRL (real-world robot navigation) 2️⃣ Setup: SFT-only vs RL-only vs hybrid (SFT→RL) pipelines + RL variants: 1/3/5/10 verification iterations (”Reject Sampling”) 3️⃣ Metrics: In-distribution (ID) vs out-of-distribution (OOD) performance 4️⃣ Ablations: Applied RL directly to base Llama-3.2 without SFT initialization; Tested extreme SFT overfitting scenarios; Compared computational costs versus performance gains Insights/Learning 💡 Outcome-based rewards are key for effective RL training 🎯 SFT is necessary for RL training when the backbone model does not follow instructions 🔢 Multiple verification/Reject Sampling help improve generalization up to ~6% 🧮 Used Outcome-based/rule-based reward focusing on correctness 🧠 RL generalizes in rule-based tasks (text & visual), learning transferable principles. 📈 SFT leads to memorization and struggles with out-of-distribution scenarios.

Prediction: One of the defining political conflicts of the next decade will be AI vs the extremely powerful teachers unions