The push to make everything learnable is raising the cost of AI beyond what individuals or even large university departments can afford.

Judged as of March 2019. True if contemporaneous evidence shows that the compute and data costs of state-of-the-art learning-centred AI research (for example, training costs of leading deep learning or reinforcement learning models such as AlphaGo Zero, OpenAI Five or large language models of 2018–19) had risen beyond what individual researchers or typical large university departments could fund, so that frontier results were coming mainly from well-resourced companies. Mostly true if costs had clearly risen sharply and were out of reach for individuals, but the evidence shows large university departments could still afford some frontier-level work, or the link to the push for learnability is only partly supported. Mixed if the evidence is evenly split: costs were rising for some kinds of frontier work while comparable results remained affordable to academia. Mostly false if costs had risen only modestly and most frontier learning results were still achievable within university budgets, so the claim overstates a real trend. False if the evidence shows that AI costs at the time were not beyond the reach of individuals or university departments.

Set before the research began.

1. E1 OpenAI Research | Conclusion OpenAI · 16 May 2018

   “We’re releasing an analysis showing that since 2012, the amount of compute used in the largest AI training runs has been increasing exponentially with a 3.4-month doubling time”

2. E2 OpenAI Five OpenAI · 25 June 2018

   “running on 256 GPUs and 128,000 CPU cores—a larger-scale version of the system we built to play the much-simpler solo variant”

3. E3 The Staggering Cost of Training SOTA AI Models | Synced Synced · 27 June 2019

   “Some of the models are so big that even in MILA (Montreal Institute of Learning Algorithm) we can’t run them because we don’t have the infrastructure for that. Only a few companies can run these very big models they’re talking about.”

4. E4 The Staggering Cost of Training SOTA AI Models | Synced Synced · 27 June 2019

   “Google suggests researchers with tight budgets could pretrain a smaller BERT-Base model on a single preemptible Cloud TPU v2, which takes about two weeks with a cost of about US$500.”

E1 reports that since 2012, the compute used in the largest AI training runs had grown exponentially, doubling every 3.4 months. E2 describes OpenAI Five running on 256 GPUs and 128,000 CPU cores. E3, from June 2019, quotes a researcher saying even MILA (Montreal Institute of Learning Algorithms) lacked the infrastructure to run some very large models and that only a few companies could run them. E4, however, reports Google's suggestion that researchers with tight budgets could pretrain a smaller BERT-Base model for about US$500, so some work stayed affordable, and the evidence ties the cost rise to the push for learnability only indirectly.