Why did OpenAI release a bigger GPT-2 model, and what can it do?

Yesterday, 20 August 2019, OpenAI made a larger version of its GPT-2 language model available for anyone to download. The new model has 774 million parameters, roughly half the size of the 1.5-billion-parameter system the lab described in February and has kept private ever since.

The release comes with a report card of sorts. In a post titled "GPT-2: 6-Month Follow-Up," OpenAI explains what it learned since February, names four university partners studying the model, and shares a legal agreement other labs can copy when they want to share models with each other. The full model is still on hold.

What OpenAI released, step by step

A language model is a program trained to predict the next word in a piece of text. GPT-2 learned this from millions of web pages, and doing it well taught it to write paragraphs that stay on topic. Its size is measured in parameters, the adjustable numbers inside the network that training tunes. More parameters generally means more fluent and coherent writing.

In February, OpenAI chose to release GPT-2 in stages instead of all at once, worried that the biggest version could be used to mass-produce fake news or spam. Here is where that plan stands today:

If those first two numbers look unfamiliar, there is a reason. OpenAI's code repository now notes that its earlier counts were wrong: the small model it called 117M is really 124M, and the medium one called 345M is really 355M.

Alongside the weights, OpenAI added a model card to the repository, a short fact sheet describing what the models are for and where they fall short. It says the main intended users are AI researchers and practitioners, and lists likely secondary uses: writing help such as grammar and autocomplete, creative writing, and games or chatbots. It also states plainly that GPT-2 cannot tell fact from fiction, so OpenAI does not support uses that require the text to be true.

What six months of study found

OpenAI's post sums up three lessons.

Coordination between labs is hard but possible. No one has publicly released a model of the full GPT-2's size, even though several groups could train one. OpenAI says it has spoken with more than five teams that rebuilt GPT-2, and that Hugging Face and a group from the Allen Institute for Artificial Intelligence and the University of Washington have adopted staged releases of their own.

People can be fooled by machine-written text. Researchers Sarah Kreps and Miles McCain at Cornell, writing in Foreign Affairs, found that 72 percent of one group of readers judged GPT-2 articles credible. Real New York Times articles scored 83 percent.

Spotting generated text is difficult. Automatic detectors, which try to tell machine text from human text, reach accuracy in the low-to-mid 90s percent range, according to OpenAI. That sounds high, but a tool used across a whole website would need to be right 99.9 percent of the time or better, or it would wrongly flag huge numbers of real people. Fine-tuning the model, meaning giving it extra training on a narrower kind of text, makes detection harder still. OpenAI argues that statistics alone will not be enough, and that human judgment and information about where a text came from will be needed too.

Who is studying the model

OpenAI has formed four research partnerships. Each partner gets to study both the new 774M model and the unreleased full model:

• Cornell University is measuring how easily people are taken in by disinformation produced by these systems.
• The Middlebury Institute of International Studies, through its Center on Terrorism, Extremism, and Counterterrorism, is looking at how extremists might misuse GPT-2.
• The University of Oregon is building "bias probes," tests that reveal unfair patterns the model picked up from its training text.
• The University of Texas at Austin is testing whether GPT-2's output can still be detected after the model is fine-tuned on specialized material.

To make this kind of sharing easier, OpenAI published a non-commercial legal agreement that any organization can adapt. It also released a technical report on publication norms. VentureBeat reports that the report was co-written with Harvard research associate Ariel Herbert-Voss, and that it explains the logic of staging: since larger models tend to perform better, releasing them one at a time leaves room to study risks with each step.

Why this release matters

For researchers and hobbyists, the practical change is simple: they can now experiment with a model twice as large as the last public one, which writes noticeably more coherent text. The repository's warnings still apply. Samples are unfiltered, the training data carry biases and errors that the model reproduces, and OpenAI recommends clearly labeling generated text as synthetic before sharing it widely.

For the wider field, the release is a test of a new idea. In February, critics argued that withholding the model hurt scientists who lacked the money to rebuild it, and some called the move a publicity stunt. OpenAI is now trying to show that staged release, combined with private sharing among trusted partners, can be a workable middle path between publishing everything and publishing nothing. The lab also warns that these questions will grow, not shrink, as models get larger.

Will OpenAI release the full GPT-2 model?

Probably, but not yet. OpenAI's stated plan is to publish the 1,558-million-parameter model in a few months. It adds a clear condition: findings from its partners, or signs that the 774M model is being misused, could change that plan.

Three things will shape the decision. The first is the partner research: if Cornell finds that larger models are much more persuasive, or Texas finds their text is hard to detect, caution wins. The second is how people use the model released yesterday. The third is ongoing discussion with other researchers and policymakers about what larger models mean.

In the meantime, anyone curious can download the small, medium and large models from OpenAI's GitHub repository and run them on their own computer. Comparing the three sizes on the same prompt is one of the clearest ways to see what extra parameters actually buy.