GPT-2 can perform tasks it was never explicitly trained for, simply by prompting with the right text format. This works because the model encountered similar patterns in its diverse training data.

GPT-3 scaled up another 100× (175B parameters, 300B tokens, ~$4.6M) and discovered in-context learning — the ability to learn new tasks from just a few examples in the prompt, with no gradient updates.

RLHF requires training a separate reward model and running PPO — complex and unstable. Two alternatives have emerged:

DPO (Rafailov et al., NeurIPS 2023): Direct Preference Optimization skips the reward model entirely. It directly optimizes the LLM to prefer winning responses over losing ones using a simple classification loss. Same quality as RLHF, ~3× less compute.

GRPO (Shao et al., 2024): Group Relative Policy Optimization from DeepSeekMath. Instead of human rankings, it generates multiple responses and uses a verifiable reward (e.g., "is the math answer correct?") to rank them. No humans needed for domains with checkable answers.

Chain-of-thought (CoT) prompting asks the model to show its reasoning step by step before giving a final answer. This dramatically improves performance on reasoning tasks.

Wei et al. (2022) claimed that abilities like arithmetic and reasoning emerge suddenly at certain scales — absent in small models, present in large ones.

Schaeffer et al. (2023, NeurIPS) challenged this: they showed that "emergence" disappears when you switch from nonlinear metrics (exact-match accuracy) to linear metrics (token-level accuracy). The abilities were developing gradually all along — the metric just couldn't detect partial progress.