OpenAI Shares Pretraining Perplexity Data to Establish Shared AI Safety Standards
OpenAI published internal perplexity measurements and tracking data to encourage shared safety standards across the artificial intelligence industry.
Modern artificial intelligence systems rely heavily on training a compressor, or equivalently a predictor, on a very large corpus of data. OpenAI emphasized that monitoring how quickly better compressors are trained is essential for grasping the overall trajectory of artificial intelligence progress. The company made the case for adopting perplexity, or equivalently cross-entropy loss, as a standard measurement to evaluate different model creators and guide mutual safety protocols.
As a first step, OpenAI provided transparency into pretraining progress by publishing internal perplexity measurements. Publishing these internal metrics took some consideration because it gives competitors a clearer picture of where the organization stands, but that concern was outweighed by the urgency of establishing shared safety measures across the industry. OpenAI stated a hope for this to be a step towards building a shared scientific basis for future pacing efforts.
A simplified yet fundamentally informative view of AI development involves first building a large computer, then using it to pretrain a model or compressor on a large unstructured dataset, and finally applying a variety of techniques such as prompting, finetuning, or reinforcement learning in different contexts to elicit latent capabilities.
The primary way progress is measured on this intelligence substrate is by tracking the perplexity of models on a held-out dataset. Perplexity measures how well a model predicts the next token on average, where lower perplexity indicates better predictions.
Operating as OpenAI Group PBC, an American artificial intelligence public benefit corporation headquartered in San Francisco, California, the organization maintains a dominant industry position through its development of generative AI models, including the generative pre-trained transformer series. Its current strategy prioritizes advancing safety measurements and establishing shared scientific tracking standards across the sector, a stance that motivated the organization to publish internal pretraining data despite competitive risks.
The range of capabilities described by this metric includes GPT-2 setting a new perplexity state-of-the-art on older datasets to become the first model capable of consistently formulating coherent paragraphs. A major leap came with GPT-3, which learned in context given a few examples, followed by GPT-3.5 at a log-perplexity of approximately 0.99 on the main current evaluation dataset.
GPT-3.5 enabled fluent conversation and formed the basis for the original ChatGPT. GPT-4 reached approximately 0.91 perplexity, proving proficient at utilizing chain-of-thought reasoning productively and motivating modern reasoning models. Losses below 0.8 have brought the emergence of transformative capabilities in fields including cybersecurity and mathematics.