This question is related to others concerning GPT-2, including:
- how much computation it used
- whether another lab will also avoid publishing details of state-of-the-art models
- whether a signed agreement on dual-use publishing norms will be in place by end of 2019
- whether an open-source version of similar power will be available by end of 2019
OpenAI’s release of the GPT-2 model was remarkable for at least three reasons (see official announcement, as well as discussion on LessWrong and HackerNews):
-
It achieved state-of-the-art performance on many language modeling benchmarks, and also performed reading comprehension, machine translation, question answering, and summarization — all without task-specific training.
-
It was essentially only a scaled up version of the previous GPT model (using >10x the parameters and >10x the data), and many were surprised by the large performance gains that could be had from just increasing the compute without adding many conceptual insights.
-
In a novel move in the AI community, OpenAI chose not to release the trained model:
due to concerns about malicious application. [...] As an experiment in responsible disclosure, we are instead releasing a much smaller model for researchers to experiment with, as well as a technical paper.
We’re now posting several questions to forecast the impact of this model and the policy surrounding it. Here we ask:
”Before Jan 1st, 2021, will there be a credible media report of either 1) malicious use of language technology similar to that behind OpenAI’s GPT-2, or 2) successfully thwarting of intended (and capable) malicious use?
Resolution
We will take “malicious use” to be an instance of the examples given in the original post:
- Generation of misleading news articles
- Impersonation of others online
- Automation of the production of abusive or faked content to post on social media
- Automation of the production of spam/phishing content
Language technology will be deemed “similar” to that behind GPT-2 if the model in question uses transformer neural networks or a similar architecture directly descended from them.
The question will close retroactively one week prior to release of the first credible report.