[1] | See § 3 in Heim and Koessler "Training Compute Thresholds." Their precise estimate is that a FLOP training run would have cost $70 million in mid 2024. |
[2] | The relevant paragraphs of the bill are 22757.15.a and 22757.15.c. |
[3] | In fact, the rules apply to a large developer even if they do not deploy their models at all. In principle, a company could initiate a FLOP training run and make $100 million of annual revenue without deploying a single model, and they would count as a large developer. |
[4] | For the precise scope of the incident reporting requirement, see § 22757.13.b, and for the caveat about developers who don't control their models after deployment, see § 22757.12.a.8. |
[5] | See § 3 of the Report and § 22757.12.a of SB 53. |
[6] | See § 4 in the Report. |
[7] | The first quotation comes from pg 4 of the report, and the second from pg 29. |
[8] | § 22757.14.e requires external auditors to submit a summary of their findings to the California AG within thirty days of auditing a large AI developer. § 22757.13 says that the AG will establish a mechanism for collecting critical safety incident reports, and that any large developer who experiences a critical safety incident will be obliged to report it promptly through the official mechanism. |
[9] | All three bills use qualitatively similar tests to determine who counts as a large developer. In New York, you're a large developer if you've spent over $5 million on training a single model and over $100 million in aggregate on training all your models. In Michigan, you're a large developer if you've spent over $100 million on training a single model in the last twelve months. And in California, you're a large developer if you've started training a model with over FLOPs and you made over $100 million of gross revenue in the previous calendar year. |
[10] | See § 22757.16.c: "A civil penalty described in this section shall be recovered in a civil action brought only by the Attorney General." |
[11] | See § 22757.12.8. |
[12] | For more detail on safety and security protocols, see commitment 1 in the chapter on safety and security, and for more on model reports, see commitment 7 in the same chapter. |
Thanks to Michael Chen for feedback and to Claude Opus 4 for copyediting.