Papers · GPT-2
AI PAPER

Language Models are Unsupervised Multitask Learners

Showed that a 1.5B parameter language model trained on internet text could perform diverse NLP tasks without explicit fine-tuning, establishing the zero-shot learning paradigm.

Authors
Alec Radford, Jeffrey Wu, Rewon Child +3
Institution
OpenAI
Published
February 14, 2019
Citations
28,000
Category
Architecture
Impact
foundational
PAPER EXPLAINED

What This Paper Does

GPT-2 scaled the autoregressive language model approach to 1.5 billion parameters, trained on WebText, a dataset of 8 million web pages curated by following outbound links from Reddit posts with high karma. The model uses a decoder-only Transformer that simply predicts the next token. The key finding was that this model could perform tasks like translation, summarization, and question answering without any task-specific training, purely through the patterns learned during language modeling.

Why It Matters

GPT-2 was the first clear demonstration that scale transforms language models from narrow text generators into general-purpose reasoning systems. The paper showed that a sufficiently large language model, trained on sufficiently diverse text, implicitly learns to perform tasks that were previously thought to require explicit supervision. This zero-shot capability suggested that language modeling might be a path to artificial general intelligence, an idea that would drive billions in investment.

The paper also sparked the first major debate about AI safety and release practices. OpenAI initially withheld the full model, citing concerns about misuse for generating disinformation. This staged release strategy was controversial but established that frontier AI labs had responsibilities beyond pure research.

Key Contributions

The central insight was that language modeling on diverse internet text implicitly teaches task completion. When the training data contains examples of translation, summarization, and Q&A embedded within natural text, a powerful enough model learns to perform these tasks by continuing patterns. No explicit task labels are needed.

WebText demonstrated that dataset curation matters as much as dataset size. By selecting text that humans found worth sharing, the resulting model produced notably more coherent and topical text than models trained on random web crawls.

What Changed Because of It

GPT-2 established the scaling hypothesis within OpenAI and the broader field. If 1.5B parameters could produce zero-shot task completion, what would 10x or 100x more parameters achieve? This question directly motivated GPT-3 and the compute scaling race that followed.

The text generation quality was sufficient to raise genuine concerns about automated disinformation, academic dishonesty, and spam generation. These concerns would only intensify with each subsequent generation and remain central policy questions today.

Connection to Current Models

GPT-2’s architecture is a direct ancestor of every GPT model that followed. The decoder-only Transformer with learned positional embeddings, the autoregressive training objective, and the approach of scaling parameters and data have remained essentially consistent through GPT-3, ChatGPT, and GPT-4. The staged release approach has also become standard practice at frontier labs, though the specific thresholds for responsible release continue to evolve.