This question is related to others concerning OpenAI LP, including:

• What their valuation will be by end of 2019
• Whether they will take concrete steps toward profitability by end of 2019

as well as

• By mid-2020, what will be the maximum compute (measured in petaflop/s-days), used in training by a published AI system?

Following the announcement of OpenAI LP, their CTO Greg Brockman participated in the following Twitter exchange:

𝔊𝔴𝔢𝔯𝔫:‏ This is a very unusual structure and raises a lot of questions about conflicts of interest. Could you guys explain a little more about the legal basis for this and what exactly you feel unable to do as a normal nonprofit that requires this exotic hybrid structure?

Greg Brockman‏: We will need to raise billions of dollars. If we do not, the Nonprofit would fail at its mission. We then needed a structure that let us custom-write (very unusual!) rules like the following: - Fiduciary duty to the charter - Cap returns - Full control to OpenAI Nonprofit

𝔊𝔴𝔢𝔯𝔫:‏ That buys a lot of GPUs/TPUs indeed for things like OA5 or GPT-2 or far larger still, but it's hard to see what that has to do with AI safety. There's nothing in AI safety which benefits from throwing 10,000 GPUs at it, is there?

Greg Brockman‏: Per our Charter (openai.com/charter/ ) our primary means of accomplishing the mission is to build safe AGI directly. As you say it's going to take a lot of compute!

(Interestingly, our safety research looks similar in profile to other ML research and is also scaling fast.)

This question tries to get at this disagreement, by asking:

By July 1st 2020, will there be an AI safety experiment described in a published paper, pre-print or blog post, that used at least 1800 pfs-days of compute (calculated using the OpenAI methodology)?

Resolution

A “published AI system” is a system that is the topic of a published research paper or blogpost. In order to be admissible, the paper/blog post must give sufficient information to estimate training compute, within some error threshold.

An AI safety experiment shall be defined as one that is credibly claimed to have as its primary motivation to differentially make progress on AI safety over AI capabilities. For example, OpenAI mentioning in an ordinary paper that their mission is safety-related does not count, the particular experiment that is conducted must be traceable back to this mission.

To anchor on something, examples of previous safety experiments by OpenAI and DeepMind:

• Learning from Human Preferences (Amodei et al., 2017)
• Reinforcement Learning with a Corrupted Reward Channel (Everitt et al., 2017)
• AI Safety Gridworlds (Leike et al., 2017)

And examples of things which are not safety experiments:

• AlphaStar (the small paragraph in the announcement mentioning possible safety relevance can not credibly have been claimed as the key motivation for running the experiment)
• GPT-2
• OpenAI Five
• ...and most other things