On the Proposed California SB 1047
thezvi.substack.com
thezvi.substack.com
Oracle will need KYC of customers to determine if you are good and safe enough.
And make sure there are no future competitors to themselves (and maybe their customers as well!).
Not all things are actually the same! At some point, there is a point where allegory breaks down and one thing can be VERY BAD where other things are FINE. You have to look at the actual contents of the thing, and a lot of very smart people think that AI is *not like the other things*.
Please engage at that level!
Also, Larry Ellison didn’t market Oracle as world-ending technology. Altman et al have.
You have people saying “our technology could end the human race, please let us continue to do whatever we want with it without oversight”.
Regardless of whether you believe them or not, the statement itself should be enough to provoke oversight.
People were definitely worried about databases! All the jobs lot for people that did filing! Copying! Typing! It’s easy to look back with hindsight and so, eh no one worried.
Heck, there were strikes when someone first played a record on the radio! God forbid!
BTW: a lot of people consider what they work on being so much more important to the world, and thus they are more important.
But does it apply to legislating? One drawback might be that political will and attention is fleeting; and interests become vested in a current state, blocking change; and it's now or not for another generation or so.
If we want to have good regulations and effective regulatory agencies, we should be judicious and methodical about what regulations are adopted, who is granted enforcement authority, and what that authority consists of.
It's mostly circular reasoning: Why is it more important? Because we should have a high index of suscipicion, because it's more important.
I think we should be equally suspect of people pushing against regulation, which also empowers the powerful.
If a problem, like tracking what non SOTA models are doing with respect to safety, don't rise to being worth regulated, then why bother regulating it? If you're going to force everyone in the field, in California, to file additional paperwork (and understand what they're legally supposed to report and how and when), then you probably should have a compelling argument for why. If you didn't do that now, and it's never a big enough problem to regulate then you've done no real harm. If it's clearly an issue that needs to be regulated in the future then it should be easy to do so.
That hasn't really been the case, though. Lots of things have needed regulation for a long time - look at carbon emissions, for example - and it doesn't happen.
Perhaps the solution is to create the regulatory agency and give them a legitimate process to create new regulations as needs change. Legislatures have limited time and expertise, and can't possibly keep up anyway.
Someone writes these things. They are influenced by lobby, yes, but even lobbies, think tanks and such generally draw their ideas from what's lurking around in their vicinity.
There are no secret "geniuses of working out how this should work." That is what it is, and it shouldn't surprise anyone.
The public sphere is not good at producing politically practical suggestions either. It's good to talking s** and offering charismatic partial answers, speaking for some specific interest or opinion. Barely better than the lobbies, sometimes worse.
I also came here to comment on my concern about this. Only big players with significant legal resources like Google/Microsoft are going to be able to compete once you let this element start to creep in.
Some of these people are directly employed by the government, others are tied to (government funded) non-profits. They may have good intentions, but it is the end results that matter. Ultimately, they will not speak against the self-interest of their own group.
From the standpoint of software development, think of it as a codebase you can't refactor but has an endless stream of new business logic which may or may not conform well with the database structure. Eventually it just stops working.
To speak about Google/Microsoft/Big Tech in particular, it isn't just legal resources - it is overall resources. Doing basic things get a lot more expensive and slow.
A users misuse of a technology shouldn't be the responsibility of the developer. You could apply this to almost every product in the world otherwise.
Software developers and researchers should not be liable for distributing information or code, even if it's used for something illegal, as long as they aren't explicitly promoting the illegal activity and don't have any involvement with it outside of creating the software.
Not only is that consistent with previous decisions, such as those regarding copyright (i.e. torrents are fine, but making a client to torrent movies specifically isn't), but also any other decision would be a violation of the social contract with regard to open-source development.
I mean, there are really two categories of software:
* Free and/or open source software. In this case, I think there is no good reason to make the developer liable, unless they're promoting illegal use. No person wants to be attacked for giving away something for free. That's why the LICENSE.
* Commercial/paid software. In this case, it is reasonable to argue that companies should be liable if end users are harmed by the software. For paid software especially, disclaimers cannot be absolute.
But I do not think it is acceptable to hold developers liable for second-order effects - i.e., a user doing something illegal with the software and harming a third party - unless it was obvious to them that the user was going to do something illegal.
If Honda put a big spike on the front of their vehicles because they thought it looked good and would sell more cars, but the spike was good at skewering pedestrians, they'd be at fault too. It wouldn't matter that their designers thought the spike was sexy and would sell more cars. You can't make something you know to be dangerous and expect to sell it to the public without being regulated.
Want to avoid the regulation, don't steal a bunch of celebrity photos and an provide your users with a tool that that creates celebrity porn deepfakes on demand.
This isn't controversial. Go to Microsoft's AI chatbot today and try to get it to create a naked image of Taylor Swift. Microsoft has spent non-trivial engineering resources making that fail. Not doing that work is irresponsible and likely to lead to lawsuit that may or may not be winnable but that Microsoft and others clearly want to avoid.
That said, I dispute the idea that these models are "dangerous" in the first place. A box that generates texts and images is not even remotely as dangerous as a sharp spike strapped to a car. Such a comparison is hyperbolic.
People act like these models are going to be the end of US when they're literally just "instant photoshop." A dangerous model would be one designed to run a military drone or automatic weapons, not a random text and image machine.
All that aside, the deepfake issue has nothing to do with the model datasets including celebrity photos (in fact, it would work fine without any of them). And no, downloading public photos is not stealing either.
Obviously if a site promotes "download this tool to generate infinite nude pictures of celebrities", then that is illegal, since that particular tool was only developed for illegal uses.
* Honda made their car poorly and there are sharp edges at the fenders, and the driver purposely used those to injure someone. I think Honda should still have some liability; their poor construction resulted in extra injury, regardless of the application.
* Honda intentionally or with recklessness built the car in a way that would serve as a useful tool for murder, in ways that served no worthwhile purpose, in ways that could be secured. I don't know the law exactly, but I expect Honda would be liable, and IMHO that would be absolutely right.
Still, if Honda builds a safe car and someone simply chooses to use its mass x acceleration to kill someone, then I wouldn't hold Honda liable.
What's the equivalent of that for AI? Should there be a watermark so police can trace an image back to a particular person's software? If that isn't acceptable (and I don't think it would be), how do we prevent people from producing deep fakes? At the distribution level? These are hard problems, and I don't think the car analogy really gets us anywhere.
Yes, that's why I offered the kitchen knife example instead. Cars are also a problematic analogy, because even though some people still consider their operation to be fully controlled by the driver and not the manufacturer via their software, that's apparently becoming less the case.
> If that isn't acceptable (and I don't think it would be), how do we prevent people from producing deep fakes?
You don't. The problem isn't producing deepfakes. The problem is committing fraud, regardless of the tools used. Someone using deepfakes to e.g. hide facial disfigurement from their employer isn't someone I mind using deepfakes.
I agree here. But what about the harder questions? Do you think deepfake porn of celebrities should be allowed? What about deepfake porn of an unpopular student at the local high school?
If these aren't allowed, where is the best place to prevent them, but still has minimal impact on the allowed uses? At the root level of the software capable of producing them (what seems to be proposed in TFA)? At the user level (car analogy)? At the distribution level (copyright-style)? I don't know the answer to these questions, but I think we should all be talking and thinking about them.
Should be handled by impersonation/defamation laws. For celebrities, perhaps it may be handled by copyright. That would allow them to license their likeness under their own conditions.
> If these aren't allowed, where is the best place to prevent them, but still has minimal impact on the allowed uses?
By enforcing laws against the bad behaviors themselves and not trying to come up with convoluted regulations on tools just because of their potential to be used badly.
> At the root level of the software capable of producing them (what seems to be proposed in TFA)?
> At the distribution level (copyright-style)?
You'd just be increasing the costs of producing software and distribution means (thinking of stuff like YouTube). It's just setting up already powerful companies to become even more powerful by raising the bar on what potential competition must be ready for from the get-go.
and
> You'd just be increasing the costs of producing software and distribution means (thinking of stuff like YouTube). It's just setting up already powerful companies to become even more powerful by raising the bar on what potential competition must be ready for from the get-go.
These arguments could be made against any "custom" regulatory scheme like what we have for drugs, cars, airplanes, etc. But sometimes the unique harms presented by certain classes of products require unique regulatory schemes.
Maybe you're right (I hope you are) and the potential harms of AI are not really significant enough to warrant any special regulation. But I don't think that is _obviously_ the case, and I would be careful when it comes to talking with normies about this stuff - AI does seem to be really scary, and hearing a techie hand-wave their concerns over technology they don't understand has the potential to make it worse. Good luck out there buddy.
We have a long history of legally targeting companies that produce products targeted at criminal activity or implied criminal activity.
Ah but some cars today have automatic breaking, so can I sue the manufacturer for not including one? Maybe Toyota's would have seen the pedestrian, is it reasonable to assume Honda's should have as well? Since this is a safety issue, why did Honda allow the car to even start without an up to date pedestrian detection system?
Does this help you see the issue?
I'm not trying to address any perceived 'issues' with ai here, I'm pointing out the flaw in holding the developer/ manufacturer liable for what a end user does.
Also I could switch the analogy to planning a robbery over WhatsApp, hacking into a bank using 'pentration testing tools' or even just windows itself for allowing users to run any software they want. Or maybe windows allows piracy by not scanning every file against a hash of known piracy content.
You can make up a million scenarios of end users misusing products, I'm sorry you don't like the car one.
Let's do the same for A.I., right? How about you reply with the regulations that A.I. companies face today that are equivalent to what car companies face. I'll check back for your answers. If there are any gaps, then let's get to work on that legislation.
1. *Fuel Economy Standards (Corporate Average Fuel Economy, or CAFE)*: Auto manufacturers are required to meet specific fuel efficiency targets for their fleet of vehicles. These standards aim to reduce greenhouse gas emissions and promote fuel-efficient technologies.
2. *Emissions Standards*: The Environmental Protection Agency (EPA) sets emissions limits for pollutants such as nitrogen oxides (NOx), carbon monoxide (CO), and hydrocarbons (HC). Compliance with these standards ensures cleaner air and reduced health risks.
3. *Safety Regulations (National Highway Traffic Safety Administration, or NHTSA)*: Auto manufacturers must adhere to safety standards related to crashworthiness, occupant protection, airbags, seat belts, and child safety. These regulations help prevent injuries and fatalities.
4. *Recall Requirements*: Auto manufacturers are obligated to promptly address safety defects by issuing recalls. The NHTSA oversees recall processes to protect consumers from faulty components or design flaws.
5. *Consumer Protection Laws*: Regulations ensure transparency in advertising, warranties, and pricing. Auto manufacturers must provide accurate information to consumers and address any deceptive practices.
6. *Clean Air Act*: This federal law regulates emissions from vehicles and sets emission standards for pollutants. Compliance with these standards is crucial for environmental protection.
7. *Corporate Average Emission Standards (CAES)*: Similar to CAFE, CAES focuses on reducing greenhouse gas emissions. Auto manufacturers must meet specific emission targets across their fleet.
(I'm sure the list goes on a good bit longer but I feel like this is enough for now.)
With cars, we let most people use some very dangerous but also very useful tools. Our approach, as a society, to making those tools safe is multi-layered. We require driver's ed and license drivers to make sure they know how to be safe. We register cars as a tool to trace ownership. We have rules of the road that apply to drivers. We have safety rules that apply to manufacturers (and limit what they are allowed to let those tools do). If a user continues to break the rules, we revoke their license. If the manufacturer breaks the rules, we make them do a recall.
I actually agree with you 100%, this is probably a good way to think about regulating AI. Some rules apply to individual users. Some rules apply to the makers of the tools. We can come together as a society and determine where we want those lines to be. Let's do it.
Note that courts take advertising over warning labels and the manual. Which is why many car ads have the text "professional driver on closed track on screen" - make it clear they they think car can do it but not most customers. Likewise cutting tools often have "guards removed for clarity" are clearly not operating (or clearly a cartoon image and not the real tool) - if they advertise someone running the tool without the guard they are liable.
There is also the concept of foreseeable misuse in courts. If you can imagine someone would do that you have do show the courts that isn't the intended purpose and you tried to prevent it. If someone does something you didn't think of, then you need to show the court you put a reasonable effort into figuring out all the possible misuses otherwise it becomes a lack of creativity on your part. Thinking of a misuse doesn't mean you have to make it impossible, just you have to make reasonable effort to ensure that doesn't happen (guards, warning labels, training, not selling to some customers - all are common tactics to sell something that can be misused without being liable, but even there you can't put a warning label on something if you could have placed a guard on the danger)
The above just brushes the surface of what the courts deal with (and different countries have different laws). If you need details talk to a lawyer.
The car allows you to break the law by going 2x faster than the highest speed limit in the nation. A faster car, with higher ground clearance does make it easier to fatally run into someone. The Tesla cybertruck is a killing machine in car form.
Cars are the leading cause of death in the US. Maybe we need to have a similar 'pre-emptive manufacturer-side intervention' bill for cars too.
s/deep-fake/photoshop
Deepfakes are simply more convenient photo/video/audio editing that has been around for decades[1], and we don't really need new legislation to deal with them. Fraud/defamation/etc, the actual harmful aspects of what can be accomplished with deepfakes, don't need any new updates to handle the technology. If we're going to hobble new technologies, we may as well go back and hold Adobe responsible for all the shady things people have done with Photoshop, and video/audio editing suites for all the deceptive clips people have spliced together.
I vaguely recall seeing some fairly convincing B&W Soviet-era photos (I think they had Stalin in them) where people were removed and other people moved around to fill the gap. And document forgery for the purposes of fraud and espionage has of course been around for centuries.
But I think the issue is less the capability itself, and more that companies will make it too easy (trivial, actually) for anyone to commit mischief. The ability to mass-manipulate images on command is no longer restricted to the General Secretary of the USSR.
That doesn't necessarily mean regulation is required, though--plenty of modern technologies make it very easy to commit crimes, but only some of them require special rules.
''' (1) The artificial intelligence model was trained using a quantity of computing power greater than 10^26 integer or floating-point operations in 2024, or a model that could reasonably be expected to have similar performance on benchmarks commonly used to quantify the performance of state-of-the-art foundation models, as determined by industry best practices and relevant standard setting organizations.
(2) The artificial intelligence model has capability below the relevant threshold on a specific benchmark but is of otherwise similar general capability. '''
They then go on to say:
>Under this definition, if no one was actively gaming benchmarks, at most three existing models would plausibly qualify for this definition: GPT-4, Gemini Ultra and Claude. I am not even sure about Claude.
Well how did they draw this conclusion? They decided that other models that do well are simply gaming benchmarks, and therefore can ignore the provision?
>Um, no, because the open model weights models do not remotely reach the performance level of OpenAI?
To me the author is simply incredibly biased and not well versed in other models. Is chatgpt4 better? Yes. Is Mistral8 comparable? I'd easily say yes. Using it just once would lead most to a similar conclusion, it's only after throwing word problems and tricks where you can really see the differences.
And these kind of laws open to interpretation are always an issue because a responsible company has to err on the side of caution. If your model is often helpful, but only 80% as often as chatgpt, is that comparable? If someone released a new benchmark tomorrow, and your model actually beats out chatgpt, does your model all of a sudden apply?
It sounds like you can't make a "positive safety determination" if you're trying to get SoTA performance on any benchmark, so is this bill asking Meta to implement a procedure to enact a "full shutdown" of image segmentation models?
It feels like this section of the bill was written by someone who once read a LinkedIn article about AI, but isn't really sure what all the terms mean.
As an aside, I called my local library and a real person immediately answered the phone. Public infrastructure can have unexpected benefits over commercial offerings, and I’m all for it.
> "Section 5 11547.7 is for the CalCompute public cloud computing cluster. This seems like a terrible idea, there is no reason for public involvement here, also there is no stated or allocated budget. Assuming it is small, it does not much matter."
There are many solid arguments for state funding of open-source programming efforts, the only thing to look out for is that this effort doesn't follow the University of California model of 'public private partnerships' in which taxpayer funds are used to generate IP which is then exclusively licensed to private interests.
Perhaps there are solid arguments for that if your goal is to put open-source projects under the de facto control of political institutions. But for those of us who prefer for FOSS to continue on in a decentralized, mostly autonomous fashion, rather than have it be overwhelmed by rent-seekers looking for state subsidies, or be used as a pawn to advance various factions' political agendas, those arguments don't seem quite so solid.
FOSS is incidental to NASA's purpose;: they use and modify code, and contribute their work back upstream, but aren't engaged in software development for public consumption as a primary activity.
OTOH, I'd be extremely wary of a political organization founded for the specific purpose of funding/influencing the overall FOSS ecosystem.
Much more wealth is generated, many more industries have been born and thrived, in places like CA and NY than in less-regulated, lower-wage, no/low-tax states. The correlation looks very strong by the eyeball method.
Also, LA put in a tax regime where all the studios left the city for neighbors like Burbank and Culver City.
They incorporate in Deleware, where they don't locate. That's irrelevant. California's economy and business is incredible, historic.
I don't think there is any benefit, and there is massive cost to moving for any company, especially one that is built on a highly specialized labor force that doesn't want to move. At least, that's my best reading of this article's interpretation of the law's impact.
California is banking way too hard on people who want to work for companies that do not want to pay them enough to live there, do not want to pay them to work remotely, actively campaign against building more housing, and actively support immigration policies that continue to push down the wages of these individuals as well as make this "specialized" labor force less "specialized."
You would assume that there's a limit to what both the workers and the businesses are willing to put up with.
But those employees are not the ones that caused OpenAI to start and headquarter in San Francisco. Its started there becuase of the employees that do get paid enough to live in the state comfortably, and benefit massively from the aggregation of capital and experience in the Bay Area.
Perhaps somebody else can do the same elsewhere! Which would be interesting to see.
California treats anybody with less than several million in net worth and less than $400k/year in household income absolutely terribly. But that's a far bigger impact than this proposed AI legislation.
Also don't think that this workforce is exclusive to CA. AI is easy to learn, it's not quantum physics.
https://www.nerdwallet.com/cost-of-living-calculator/city-li...
https://www.nerdwallet.com/cost-of-living-calculator/city-li...
The people leaving California don't really want to leave, they are just being forced out by high costs. Those who have high incomes and can afford housing, like all those OpenAI employees, are staying.
OpenAI isn't looking for just another person that picked up AI on a whim, they are looking for people that are driving the field forward so that OpenAI stays at the very forefront. These sorts of people could learn quantum physics easily, and many probably have in their undergraduate (or graduate) degrees.
The reason is simply because ML is incomplete and thus right now is more art then science. We only understand these neural networks from the perspective of an analogy. The curve fitting analogy. Its nowhere near something like general relativity where even an analogy doesn't convey full understanding.
Not gonna argue with that, but I’m having trouble imagining the alternative. Somehow we would have understood the risks without allowing the growth and observing the consequences? That seems unlikely, considering how surprised we all seemed when the consequences occurred.
With AI we seem to have a lot more noise around imagining consequences, but in my mind there’s no reason that would correlate with completeness or accuracy of the predictions. There will be lots of very bad, and lots of very good, consequences that nobody can currently imagine.
Note that I said make not sell.
This is of course assuming we'd get the kind of regulations we need, instead of useless security theater or regulations so onerous that they prevent products from developing at all. But it does seem like it would have been possible to prevent the worst of it, since many of these cases boil down to 'Facebook (or other big company) didn't care and didn't bother to try, and then bad things happened'
While I dunno where I come down yet on this legislation, it seems truly bizarre to object to the CalCompute cluster. IMO, the biggest AI threat we face is from big tech being the only ones capable of creating new models. Right now, companies like Meta are aligned with open source AI, but that could change at the drop of the hat. It makes sense to preserve state capacity to level the playing field against attempts at monopoly or other forms of overreach. Also, a lot of great work is coming out of UC system labs, so why shouldn’t we invest more in making them successful?
The calcompute cluster is what jumped out at me.
The danger is that this will be a huge subsidy to entrenched UC/CA contractors that are in place solely to navigate - and exclude others from - state procurement mechanisms.
In fact, if you were really cynical you might think that the entire purpose of this bill is the calcompute cluster and everything else is just scaffolding to enable a boondoggle.
I am not asserting this - I am just pointing to a possible interpretation.
Given the glacial speed of government products, will the cluster be using obsolete tech by the time its finally up and running?
To me it feels like the latter, but admittedly I don't know for sure.
> Critical harm is either mass casualties or 500 million in damage, or comparable.
So the whole regulatory premise is based on missing the boat on social media, but how many examples are there of social media causing this kind of "critical harm"?
there are more, but this one came to mind
1. What I expected:
* AI Indemnification rules for platform providers building on ideas like net neutrality on one end, and content provider protections like DMCA and copyright on the other
* Consumer AI protection rules building on ideas like the Equal Credit Opportunity Act and GPDR
2. What we got: Something about KYC and rules on cases about $500M+ incidents and deaths
With AI this seems like an untenable position. We already have companies that are pushing the bounds with AI constantly and ever further, while telling us we shouldn't worry about the future dystopian implications.
Trying to reign these people and companies in with government regulation at this point doesn't seem feasible.
We should be regulating applications of AI instead (regardless of model type, size and compute).
Worries about people generating porn with AI are overblown. That already works.[1] It's technically impressive, but kind of emotionless. Nobody seems to be worried about it.
But does AI itself not need some checks and balances on it? Reasonable or not, I think the big worry (especially among laypeople) is Terminator-style AI that wants to destroy humanity. Making a law that says AI can't be used to determine if people are allowed to take out loans doesn't stop that.
Compared to Putin, Xi, Netanyahu, and Trump, AI is a minor problem.
1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey orders given it by human beings except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
Is SB1047 the first instance of AI regulation?
Is this a dog whistle for them too, or just for us?