No, there are two thresholds and BOTH must be met.

One of those is $100MM in training costs.

The other is measured in FLOPs but is already larger than GPT-4, so the “think of the small guys!” argument doesn’t make much sense.

Tell that to me when we get to llama 15

What?

“But the big guys are struggling getting past 100KB, so ‘think of the small guys’ doesn’t make sense when the limit is 640KB.”

How do people on a computer technology forum ignore the 10,000x improvement in computers over 30 years due to advances in computer technology?

I could understand why politicians don’t get it.

I should think that computer systems companies would be up in arms over SB 1047 in the same way they would be if the government was thinking of putting a cap on hard drives bigger than 1 TB.

It puts a cap on flops. Isn’t the biggest company in the world in the business of selling flops?

It would be crazy if the bill had a built-in mechanism to regularly reassess both the cost and FLOP thresholds… which it does.

Inversely to your sarcastic “understanding” about politicians’ stupidity, I can’t understand how tech people seem incapable or unwilling to actually read the legislation they have such strong opinions about.

Who are you llamaimperative, and what is your motivation to support SB 1047?

It's troubling that you are saying things about the bill which are false, and then speculating on the motives of someone just pointing out that what you are saying is false.

why not tell us?

Or point out what is actually false?

You rebutted the points that the flop limit will not actually limit anyone by saying that gpt4 is out of reach of startups.

Is OpenAI a startup? Is Anthropic? Is Grok? Is Perplexity? Is SSI?

You ignored the counter points that advanced technology exponentially raises flop limits and changes costs.

You said that the flop limit can be raised over time. So startups shouldn’t worry.

You ignored the counter point that flop limits in export controls are explicitly designed to limit competition from other nations.

Flop limits not being a real limit is a ridiculous argument. The intent of a flop limit is to limit, no matter how you sugar coat it.

You have confused me with someone else.

What's false is the idea that the limit is going to be a burden on small companies, because you can ignore the flop limit if you're spending less than a hundred million dollars. (Big companies, in contrast, can use a percent of their budget for compliance.)

Being able to ignore the flop limit makes basically everything else you've said irrelevant. But just to quickly go through: I don't want to argue about what a 'startup' is but they're not 'small guys'. Advanced tech can be compensated for, but also it doesn't change the fact that staying under $100 million keeps you excluded. Export controls have nothing to do with this discussion, and they involve a completely different kind of 'flop limit'.

I’m a person who’s interested in arguments based on reality. Frankly I don’t have a solid opinion on the bill in any particular direction.

If your goal is to lift the limit, why put it in?

We periodically raise flop limits in export control law. The intention is still to limit China and Iran.

Would any computer industry accept a government mandated limit on perf?

Should NVIDIA accept a limit on flops?

Should Pure accept a limit on TBs?

Should Samsung accept a limit on HBM bandwidth?

Should Arista accept a limit on link bandwidth?

I don’t think that there is enough awareness that scaling laws tie intelligence to these HW metrics. Enforcing a cap on intelligence is the same thing as a cap on these metrics.

https://en.m.wikipedia.org/wiki/Neural_scaling_law

Has this legislation really thought through the implications of capping technology metrics, especially in a state where most of the GDP is driven by these metrics?

Clearly I’m biased because I am working on advancing these metrics. I’m doing it because I believe in the power of computing technology to improve the world (smartphones, self driving, automating data entry, biotech, scientific discovery, space, security, defense, etc, etc) as it has done historically. I also believe in the spirit of inventors and entrepreneurs to contribute and be rewarded for these advancements.

I would like to understand the biases of the supporters of this bill beyond a power grab by early movers.

Export control flop limits are designed to limit the access of technology to US allies.

I think it would be informative if the group of people trying to limit access of AI technology to themselves was brought into the light.

Who are they? Why do they think the people of the US and of CA should grant that power to them?

Wait sorry, are you under the impression that regulated entities get to “accept” which regulations society imposes on them? Big if true!

Your delusions and lack of nuance shown in this very thread are exactly why people want to regulate this field.

If developers of nuclear technology were making similar arguments, I bet they’d have attracted even more aggressive regulatory attention. Justifiably, too, since people who speak this way can't possibly be trusted to de-risk their own behavior effectively.

Moore's law is dead though.

Cost as a perf metric is meaningless and the history of computer benchmarks has repeatedly proven this point.

There is a reason why we report time (speedup) in spec instead of $$

The price you pay depends on who you are and who is giving it to you.

That’s why there are two thresholds.

Cost per FLOP continues to drop on an exponential trend (and what bit flops do we mean?). Leaving aside more effective training methodologies and how that muddies everything by allowing superior to GPT4 perf using less training flops, it also means one of the thresholds soon will not make sense.

With the other threshold, it creates a disincentive for models like llama-405B+, in effect enshrining an even wider gap between open and closed.

Why? Llama is not generated by some guy in a shed.

And even if it were, if said guy has such amount of compute, then it's time to use some of it to describe the model's safety profile.

If it makes sense for Meta to release models, it would have made sense even with the requirement. (After all the whole point of the proposed regulation is to get some better sense of those closed models.)

Also the bill was amended NOT to extend liability to derivative models that the training company doesn’t have effective control over.

Both thresholds have a system to be adjusted.

If the danger is coming from the amount of compute invested, then cost of compute is irrelevant.

A much better objection to static FLOP thresholds is that as data quality and algorithms increase, you can do a lot more with fewer FLOPs / parameters.

But let’s be clear about these objections - they are saying that FLOP thresholds are going to miss some harms, not that they are too strict.

The rest is arguing about exactly where the FLOP thresholds should be. (And of course these limits can be revised as we learn more.)