www.lesswrong.com/posts/5aKRshJzhojqfbRyo/unless-its-governance-changes-anthropi...
1 correction found
New York's RAISE Act would apply only to models trained with over $100 million in compute.
The RAISE Act’s scope is not limited to models trained with over $100M in compute cost; it also includes an additional FLOP threshold and separately covers certain distilled models with a $5M compute-cost threshold.
Full reasoning
The post claims the NY RAISE Act would apply only to models trained with over $100 million in compute.
However, the bill text defines a “frontier model” as:
- A model trained using greater than 10^26 computational operations and with compute cost exceeding $100 million (so $100M compute cost is not the sole criterion; there is also a FLOPs threshold), or
- A model produced by knowledge distillation of such a frontier model, where the distilled model’s compute cost exceeds $5 million (so the Act’s definition also includes some models below $100M compute cost).
Because the Act explicitly includes (1) a FLOPs threshold in addition to the $100M compute-cost threshold, and (2) distilled models with a $5M compute-cost threshold, it is incorrect to say it applies only to models trained with over $100M in compute.
1 source
- NY State Senate — Senate Bill S6953B (RAISE Act) — Bill Text (PDF)
“‘Frontier model’ means either of the following: (a) an artificial intelligence model trained using greater than 10^26 computational operations … the compute cost of which exceeds one hundred million dollars; or (b) an artificial intelligence model produced by applying knowledge distillation … provided that the compute cost for such model … exceeds five million dollars.”