GPT-2 was considered too dangerous to release in full when it launched in 2019. GPT-3 shattered those concerns by demonstrating that scale brought capability far beyond simple text generation.
Parameter Count
GPT-2: 1.5 billion parameters
GPT-3: 175 billion parameters — over 100x larger
What Scale Unlocked
GPT-3 demonstrated emergent abilities not present in GPT-2: few-shot learning, code generation, and coherent long-form reasoning. This validated the scaling hypothesis and set the direction for the entire field.