This is related to the questions of:
- whether an AI safety experiment will use as much compute as AlphaGo Zero by mid-2020
- whether an experiment in academia will use as much compute as AlphaGo Zero by mid-2020
- what the maximum compute used in training by a published AI system will be by mid-2020
- when and by how much the trend will slow down
In order for the incredibly fast AI Compute trend to be sustainable more than a few years, government involvement seems necessary (at current rates Amazon and Alphabet’s R&D budgets of $20B would be reached in 2021-2022, but during the Apollo and Manhattan projects, the US government spent around $200B/year on R&D).
In early 2019, a government experiment seems to have set a new compute record by using the world’s most powerful non-distributed computer, IBM Summit, to run machine learning models of climate change. Anthony writes:
the Summit supercomputer [uses] 27,000 graphics processors. It seems this can run at a sustained 1 exaFLOPS, but amusingly this is only half of what DeepMind apparently used in training alphastar.
We now ask:
By July 1st 2020, will a government researcher be involved in running an experiment using >=31.000 pfs-days of compute, as evidenced by a blog post, paper, pre-print or other credible source (giving enough information to estimate compute within some error bars)?
(For reference, the Metaculus AI crowd estimates AlphaStar to have used 21.000 to 31.000, with 50% confidence, as of April 14.)
A government shall be defined as being involved in an experiment if at least one co-author (or contributor of similar standing) is employed by a government institution, and worked on the experiment primarily as part of that role. (So, for example, it would not count if someone was part-time at Facebook and part-time at Institution XYZ, conducted the experiment at Facebook, and was cited by their Facebook title.)
The compute refers to the amount of compute used for training of the final system. The method of calculation of compute should be as similar as possible to that used in the "AI and compute" article.
Resolution date will be retroactively set to one week prior to an eventual credible report.