OpenAI GPT-2 Release and Implications

OpenAI has developed GPT-2, a large-scale unsupervised language model that generates coherent paragraphs of text and achieves state-of-the-art performance on several language modeling benchmarks. The model demonstrates the ability to perform rudimentary reading comprehension, machine translation, question answering, and summarization without any task-specific training.

Technical Architecture and Training

GPT-2 is a direct scale-up of the original GPT model, featuring over 10 times the parameters and trained on over 10 times the amount of data.

  • Model Scale: The full model contains 1.5 billion parameters.
  • Training Objective: The model was trained simply to predict the next word given all previous words in a text.
  • Training Data: GPT-2 was trained on a dataset of 8 million web pages (approximately 40GB of Internet text) curated via outbound links from Reddit that received at least 3 karma to ensure diversity and quality.
  • Architecture: It is based on the transformer architecture.

Zero-Shot Capabilities and Performance

GPT-2 achieves state-of-the-art scores on a variety of domain-specific language modeling tasks in a "zero-shot" setting, meaning it was not trained on any data specific to those tasks and was evaluated only as a final test.

Language Modeling Benchmarks

GPT-2 outperformed models trained on domain-specific datasets such as Wikipedia, news, or books on the following benchmarks:

Dataset Metric GPT-2 Result Previous Record Human
Winograd Schema Challenge Accuracy (+) 70.70% 63.7% 92%+
LAMBADA Accuracy (+) 63.24% 59.23% 95%+
Children’s Book Test (Common Nouns) Accuracy (+) 93.30% 85.7% 96%
Children’s Book Test (Named Entities) Accuracy (+) 89.05% 82.3% 92%
Penn Tree Bank Perplexity (–) 35.76 46.54 unknown
WikiText-2 Perplexity (–) 18.34 39.14 unknown
enwik8 Bits per character (–) 0.93 0.99 unknown
text8 Bits per character (–) 0.98 1.08 unknown
WikiText-103 Perplexity (–) 18.34 18.3 unknown

General Language Tasks

While not reaching state-of-the-art for specialized systems, GPT-2 can perform tasks like reading comprehension, summarization, and translation through specific prompting of the trained model without any fine-tuning.

Text Generation and Failure Modes

GPT-2 can generate realistic and coherent continuations of text based on an arbitrary input prompt. However, the model exhibits several failure modes:

  • Repetitive Text: The model may produce repetitive sequences.
  • World Modeling Failures: The model occasionally generates logically impossible scenarios (e.g., writing about fires occurring underwater).
  • Unnatural Topic Switching: The model may switch topics abruptly.

OpenAI notes that performance varies by topic; the model is more successful with topics highly represented in the training data (e.g., Brexit, Lord of the Rings) and performs poorly on highly technical or esoteric content.

Policy Implications and Misuse Risks

OpenAI identifies several potential beneficial applications and malicious uses for large language models:

Potential Benefits

  • AI writing assistants
  • Capable dialogue agents
  • Unsupervised translation between languages
  • Better speech recognition systems

Potential Malicious Uses

  • Generation of misleading news articles
  • Online impersonation
  • Automated Content Production: Automating the production of abusive, faked, or spam/phishing content for social media.

OpenAI warns that these technologies reduce the cost of generating fake content and waging disinformation campaigns, suggesting that the public must become more skeptical of online text.

Release Strategy: Staged Disclosure

Due to concerns regarding the potential for deceptive or abusive language generation at scale, OpenAI implemented an experiment in responsible disclosure through a staged release:

  1. Initial Release: Only a much smaller version of GPT-2 (117M parameters) and sampling code were released initially.
  2. Interim Update (May 2019): A larger 345M parameter version was released.
  3. Partner Sharing: The 762M and 1.5B versions were shared exclusively with partners in the AI and security communities to improve societal preparedness.

OpenAI did not release the full dataset, training code, or the 1.5B model weights to the general public. This strategy was intended to give the community time to assess the properties of the models and evaluate the impact of each release stage.

Sources