What happens if the model is hosted in a data center that is not in California but in a different US state? But the developers interact with it remotely from the usual Bay Area campuses?
Arguably, this would be the default due to cheaper electricity being available elsewhere. There also will likely need to be careful control over model weights due to their value, so there would never be an instant where the model itself is on computers inside California, merely the source files that defined it.
Just to keep it kosher, the elite Bay area devs would set a COMPUTE_LIMIT=10^26.
The engineers who deal with the training hardware itself could make a patch to disable the compute limit, making the source not “California compliant” with the commit made outside of California, by a company not actually incorporated in California. This kinda workaround is standard elsewhere.
And obviously then if California decides the law applies anyway, the company will immediately take the case to federal court. Federal judges must agree that California law overrides federal law in this case.
California is not capable of extracting tax revenue from companies like Google in any meaningful way, so we shouldn’t expect them to be capable of taking stronger, less directly self-benefiting action. If they can’t get Google to pay them, they can’t get Google to stop AI.
What is California’s great track record in this space? They have caused “May cause cancer in California” to be printed many times. We shouldn’t expect them to save us.
What happens if the model is hosted in a data center that is not in California but in a different US state? But the developers interact with it remotely from the usual Bay Area campuses?
Arguably, this would be the default due to cheaper electricity being available elsewhere. There also will likely need to be careful control over model weights due to their value, so there would never be an instant where the model itself is on computers inside California, merely the source files that defined it.
Just to keep it kosher, the elite Bay area devs would set a COMPUTE_LIMIT=10^26. The engineers who deal with the training hardware itself could make a patch to disable the compute limit, making the source not “California compliant” with the commit made outside of California, by a company not actually incorporated in California. This kinda workaround is standard elsewhere.
And obviously then if California decides the law applies anyway, the company will immediately take the case to federal court. Federal judges must agree that California law overrides federal law in this case.
California is not capable of extracting tax revenue from companies like Google in any meaningful way, so we shouldn’t expect them to be capable of taking stronger, less directly self-benefiting action. If they can’t get Google to pay them, they can’t get Google to stop AI.
What is California’s great track record in this space? They have caused “May cause cancer in California” to be printed many times. We shouldn’t expect them to save us.
Well yes. Also, while the California government has passed many laws and made many efforts to reduce the use of fossil fuels, https://www.forbes.com/sites/rrapier/2023/12/15/us-producers-have-broken-the-annual-oil-production-record/?sh=180d45276cc6 to an extent all this does is send money elsewhere.