OpenAI today appear the final model in its staged absolution for GPT-2, the spooky text architect the AI community’s been talking about all year.

GPT-2 uses apparatus acquirements to accomplish novel text based on a bound input. Basically, you can type a few sentences about annihilation you like and the AI will spit out some ‘related’ text. Unlike most ‘text generators’ it doesn’t output pre-written strings. GPT-2 makes up text that didn’t ahead exist– at least according to OpenAI‘s analysis paper.

The non-profit made account in February when it appear that it would not absolution the full-sized models for GPT-2 to the accepted public all at once. Instead, the aggregation opted to absolution it in four parts over eight months.

An OpenAI blog post from February explains:

Due to our apropos about awful applications of the technology, we are not absolution the accomplished model. As an agreement in amenable disclosure, we are instead absolution a much abate model for advisers to agreement with, as well as a abstruse paper.

The full model contains 1.5 billion parameters. The more ambit a model is accomplished with, the ‘smarter’ it appears to be – just like humans, convenance makes perfect.

Initially OpenAI appear a model with 124 actor ambit after followed by releases with 355 and 774 million. Each abundance showed a cogent advance in adequacy over antecedent iterations. We arrested out the 774M model and were blown away. You can try it yourself at this link where developer Adam King has translated the model into a UI.

Along with the new model 1.5B model weights, OpenAI also appear its GPT-2 apprehension models in an effort to preemptively combat misuse. Unfortunately, according to OpenAI, the detector isn’t as good as the generator. In a blog post today the aggregation said:

We conducted centralized apprehension analysis and developed a apprehension model that has apprehension rates of ~95% for audition 1.5B GPT-2-generated, Specifically, we based a arrangement classifier on RoBERTaBASE (125 actor parameters) and RoBERTaLARGE (355 actor parameters) and fine-tuned it to allocate the outputs from the 1.5B GPT-2 model versus WebText, the dataset we used to train the GPT-2 model.

We accept this is not high enough accurateness for standalone apprehension and needs to be paired with metadata-based approaches, human judgment, and public apprenticeship to be more effective. We are absolution this model to aid the study of analysis into the apprehension of constructed text, although this does let adversaries with access better evade detection.

We’ll get into the adversarial (and positive) use cases for GPT-2’s full absolution once we’ve had the chance to agreement with the complete model. In the meantime, you can download the model here on Github, check out the model card here, and read OpenAI‘s blog post here.

Read next: Facebook spelled 'password' wrong in new Instagram renders