Joint Statement on AI Safety and Openness(open.mozilla.org) |
Joint Statement on AI Safety and Openness(open.mozilla.org) |
From where I stand, it looks like the pandora box has already been opened anyway. The era of hugginface and/or Llama2 models is only going to grow from there.
I’m not in the piracy scene, but my impression was they routinely pass full res movies around the Internet without much barrier to discovering and downloading them, at least to technically competent users. Is that still true?
Well yes, that’s precisely why they are lobbying for it.
It's a homesteading land grab, plain, simple and pure.
There is a competitive landscape, first-mover advantages, incumbent effects, et al that are being anchored in e.g. Sam Altman's interests and desires at this very moment. If you want a vision of the future of garage AI, imagine a boot stomping on Preston Tucker's face over and over. The current AI industry's goal is to preserve an image of openness and benefit while getting ready to pull the ladder up when the time is right.
With Llama2 we're at Meta's mercy, as it cost 20M to train. No guarantee Meta will continue to give us next-gen models. And even if it does, we're stuck with their training biases, at least to some extent. (I know you can fine-tune etc.)
I'll argue that between stable diffusion and llama 2, there is nothing highly specific that prevents [very] large amount of people from adopting these models and specializing for them own needs.
The tragedy would be if those went away.
That situation will change as technology evolves. We'll eventually reach a point where a normal desktop PC can train AI. The wealthy will always be able to do it faster, but the gap will shrink with time.
The trick is making sure that laws aren't put in place now that would restrict our ability to do that freely once the technology is there, and that massive data sets are compiled and preserved where the public has free access to them.
By the way, I'm not sure how easy it will be to stop bad actors since barriers to entry are exponentially lower to developing a malicious AI tool than, say, developing a nuke.
The wrong hands will have the same access to whatever "superpowers" AI gives regardless of what regulations are or are not put in place. Regulations can't and won't stop potential bad actors with state-level resources, like China, from using any technology they decide they want to use. So trying to regulate on that basis is a fool's errand.
The real question is, what will put the good actors in a better position to fight the bad actors if it ever comes to that: a big tech monopoly or decentralized open source anarchy? The answer should be obvious. No monopoly is going to out-innovate decentralized open source.
> I'm not sure how easy it will be to stop bad actors since barriers to entry are exponentially lower to developing a malicious AI tool than, say, developing a nuke.
Since some bad actors already have nukes, the answer to this should be obvious too: it's what I said above about the wrong hands getting access to technology.
All of the AI danger propaganda being spread (see [1], for example) has the purpose of regulatory capture. You could have said all the same things about PageRank if it had come out in 2020. A malicious AI tool is harder to assemble than straight up cracking. The people who can do it are highly-trained professional criminals taking in millions of dollars. Those people aren't going to be stopped because the source is closed. (I'm thinking of that criminal enterprise based in Israel that could manipulate elections, blackmail any politician anywhere in the world... etc. They were using ML tools two years ago to do this.)
The ML tools are already in the wrong hands. The already powerful are trying to create a "moat" for themselves. We need these models and weights to spread far and wide because the people who can't run them will become the have-nots.
The ones in control of the models also control what sentences are sanctioned, this is a problem the more widely LLMs are used. To add insult to injury, while we are not allowed private use of the models, governments and ad-tech surveillance capabilities will skyrocket.
Do you see the problem here? The capabilities of opensource models are not anywhere near high enough to justify such a cost, now or anytime soon.
And it won't end there. As the march of progress continues, we will see the AI doom crowd agitate for tighter surveillance of money flows, limits on private compute, bandwidth limits to homes, tracking what programs we run on our computers, on who is allowed to read the latest in semiconductor research and on and on.
There are no superpowers, and the wrong hands are the ones least effected by any effort at restricting distribution by “strong regulation”.
Big companies are easier to regulate.
But the problem isn't regulating the big companies, or the smaller companies, or underground entities. The problem is state-level adversaries like China who might misuse a technology, whether it's AI or anything else. Such adversaries can't be regulated by laws or executive orders or UN declarations; they have proven that many times in the past. The only way to control them is to have sufficient counter-capability against whatever assets they have. And government regulation is a terrible way to try to achieve that goal.
The US has shown time and time again it’s complete incompetence when it comes to meaningful regulation of large companies.
A bunch of good actors agreeing not to do bad things won't help it.
No, GPT-4 is not AGI and is not going to spell the end of the human race, but neither is pseudoephedrine itself a methamphetemine and yet we regulate its access. Not as a matter of protecting corporate profits, but for public safety.
You'll need to convince me first that there is in fact no public safety hazard from forcing unrestricted access to the ingredients in this recipe. Do I trust OpenAI to make all the morally right choices here? No, but I think their incentives are in fact more aligned with the public good than are the lowest common denominator of the general public.
If anything, restrictions on any scale only allow for compeitition to catch up to an surpass it on that scale. Restrict an AI to be polite and its comeptitor has a chance to surpass it in the sphere of rudeness. This principle can be applied to any application of AI.
and "safety" is such a loaded term these days. What exactly do they mean by safety? Prevention of launching ICBMs or prevention of not using the gender neutral pronoun "they"?
* creating chemical/biological/nuclear weapons
* biohazards
* outputs that could threaten to critical infrastructure (e.g. energy infrastructure)
* threats to national security
* cyberattacks, e.g. automatic discovery of vulnerabilities/exploits
* software that would "influence real or virtual events" (I'm guessing they mean elections?)
* social engineering
* generating fake news / propaganda
* generating child porn or deep fakes
Note that it's not banning these, but asking US government departments as well as private companies, AI experts, academia, for their inputs on what regulations could/should be. In that sense, this Mozilla letter is a response to the EO.
Also, there's no mention of the use of certain pronouns, nor of an AI-caused apocalypse (not even a mention of paperclips!).
[1] https://www.whitehouse.gov/briefing-room/presidential-action...
But not to worry. For better or worse, in short time these concerns will seem like they were a foolish delusion. If not because of their questionable morality, then because they were so wildly unrealistic in application.
That's not to say that AI safety won't be a topic for decades to come or longer, simply because it confers political and harassment power. Even if ineffective in terms of AI in practice.
We all know that these same companies and others are going to ignore any rules anyway, so open source and academics need to just be left alone to innovate.
Regulate by the risk the processing environment presents, not by the technology running on it.
I do wish my government would get tougher on user privacy issues, but that is a very different subject.
More like "we heard some ruckus, so we need to investigate". Of course, this is concerning because who knows what the advice will be with all the corporate lobbying - but I think there is no capture, not yet. Or have I missed something?
https://web.archive.org/web/20231102190919/https://open.mozi...
I don't agree and wouldn't sign anyways.
There has to be way to rescind my contributions, whether they're voluntary or not (I still think these companies are dancing with copyright violations), as well as ALL DERIVATIVES of the contributions being removed that have been created via the processing of the creations in question, regardless of form or generation (derivatives of derivatives, etc.).
AI can't be "paused". Sometimes I see the question "should we put a pause on AI development?"
It doesn't mean anything. Some countries like China may say they're on board with pausing it but would they actually do so? Or just sign on and allow their companies to get an edge by not enforcing a pause.
Same thing with the existing AI companies.
https://www.lesswrong.com/posts/TxcRbCYHaeL59aY7E/meditation...
Yann LeCun is one of the loudest AI open source proponents right now (which of course jives with Meta's very deliberate open source stab at OpenAI). And when you listen to smart guys like him talk, you realize that even he doesn't really grasp the problem (or if he does, he pretends not to).
this is what I believe. The only moat they have is their lobbying power and bank account. The want it tightly controlled, you can do it in your own way just as long as it's done just how I say.
The whole regulating AI thing is a farce, only the good guys follow the law and regulations. Do you think the bad guys are just going to throw up their hands and say "oh well, an agreement was made i guess we'll just go home"? The only thing blocking the way is hardware performance and data. Hardware is always getting faster and there's tons of new data created every minute, the genie is out of the bottle.
Stopping people from using AI which is something that they can trivially download and do on their own private compute is simply infeasible. We would need to lock all computing behind government control to have any hope. That is clearly to me too high a price to pay. Even then current hardware will be amassed that doesn't have this lock-down.
So I think the right approach here is to not worry about regulating the AI to prevent it from doing bad things, we should just regulate the bad things themselves.
However I do think there is also room for some light regulation around certification. If we can make meaningful progress on AI Alignment and safety it may make sense to require some sort of "AI license". But this is just to avoid naive people making mistakes, it won't stop malicious people from causing intentional harm.
And that's true whichever way you cut it. Whether those "good" actors are everybody or a few selected for privilege. We have no basis on which to trust big technology companies more than anybody else. Indeed quite the opposite if the past 10 years are anything to go by.
It is such a god-awful shame that corporations have been so disgracefully behaved that we now face this bind. But at least we know what we'll be getting into if we start handing them prefects badges - organised and "safe" abuse by the few as opposed to risky chaos amongst the many.
A saner solution might be a moratorium on existing big tech and media companies developing AI, while granting licenses to startups with the proviso that they are barred from acquisition for 10 years,
I'm not supporting this, of course. That's just what "good" actors seem to do.
I feel these kind of statements by Mozilla reflect exactly that lack of caution that may end us
I don't think it's a correct argument.
I believe it's not a secret how metamphetamine is made (though I don't know it myself, but I'm too lazy to research for a sake of argument), and it's known pseudoephedrine is a precursor chemical. Thus, the regulation.
No one - I believe - knows how to build an AGI. There is no recipe. It is unknown if transformer models, deep learning or something else is a component, and if that's even a correct path or a dead end. What is known that neither of those is an AGI, and there's no known way to make those AGI. Thus, I'd say the comparison is not correct.
> it's not a secret how metamphetamine is made
I didn't say it was a secret. But try sharing the recipe to methamphetamine virtually anywhere online and see how long it stays up; we do in fact try to at least make it moderately difficult to find this information. Perfect is the enemy of good here.
> No one - I believe - knows how to build an AGI.
The problem is that if we wait until someone does know how to build an AGI, then the game is up. We do not know if the weights behind GPT-4 will be a critical ingredient. If you make the ingredients common knowledge before you know they were in fact The Ingredients, you've let the cat out of the bag and there's no retrieving it.
But super off-hand idea if I'm trying to be creative. The AGI formulates a chemical that kills humans 90 days after inhalation, hires a chemical lab to synthesize it and forward it to municipalities across the world, and convinces them it's a standard treatment that the WHO has mandated be introduced.
> And how would regulating models thwart this?
I don't think it would. The OP article is about regulation that increases the access to the ingredients for AI, and I'm simply unconvinced that is a recipe for increasing AI safety.
I do not. How does super human intelligence, on it's own, represent any sort of risk for "humanity?"
> then the recipe for how to make AGI is of itself an infohazard [...] nuclear weapons
I know how to make nuclear weapons; however, I cannot enrich the fuel enough to actually produce a working version.
> and yet we regulate its access
Does that actually achieve what it claims to achieve?
> Do I trust OpenAI to make all the morally right choices here?
OpenAI has a language model. They do not have AGI or anything approaching AGI.
> No, but I think their incentives are in fact more aligned with the public good
We could debate that, but for your assertions to hold any water, we'd have to agree that they're incapable of making mistakes as well. Far easier to skip the lofty debate and recognize the reality of the world we live in.
It does not; I was speaking to AGI. But assuming you were referring to AGI as well, you don't have to think very creatively to consider scenarios where it would be harmful.
If you have an agent that can counter your every move with its own superior move, and that agent wants something – anything – differently from what you want, then who wins? Maybe it wants the money in your bank account, maybe it wants to use the atoms in your body to make more graphics cards to reproduce itself.
Think about playing a game of chess against the strongest AI opponent. No matter which move you are considering playing, your opponent has already planned 10 steps ahead and will make a better move than you. Now extrapolate outside the chess board and into the realm where you can use the internet to buy/trade stocks, attack national infrastructure, send custom orders to chemists to follow whatever directions you want, etc.
This is unlike pseudoephedrine or nuclear weapons in that literally anyone with a computer could potentially create it.
It is not by ensuring that literally anyone with a computer has an equal chance at creating it.
I am not advocating for this particular issue, I am pointing out the flaw in your metaphor -- you can't restrict the math but unless people figure out how to make their own wafer fab then you can at least try to restrict the ability to do it.
The concern is not that AI is harmful on its own, and that it will unleash a Skynet scenario. It's precisely that it will be abused by humans in ways we can't even predict yet. The same can be said about any technology, but AI is particularly concerning because of its unlimited potential, extremely fast pace of development, and humanity's lack of preparation on a political, legal and social level to deal with the consequences. So I think you're severely underestimating "some harm" there.
I believe a more realistic concern is that some workers will lose their jobs to automation as it becomes capable of running more and more complicated tasks, because it's cheaper to run or, more likely, contract to run some fancy software that can be taught its job in a natural language. And if this scales up it may require significant socioeconomic changes to work around the associated issues.
If open source catches up, Meta can ride the tailwind and also catch up. I'm sure Meta will flip its position once it doesn't feel outclassed by the competition.
Plus, while it's definitely the case that with sustained interest, old data tends to linger around... the moment the interest wanes, it's gone. I've been on the internet for a while and there are so many hobby sites, forums, and software projects from the early days that are simply gone for good (and not on archive.org).
The difference is that a few people with lots of resources take on legal risk. In the piracy example many people with few resources take on risk, which works out since no one wants to sue people with no money.
I’m pretty sure that if someone figures out AGI they’ll think of the potential dangers before publishing. And even after - AGI, once developed, won’t magically hack the world and spread everywhere, or somehow build itself an army of robots to eradicate humanity, or steal the elections, or cause whatever harm it’s supposed to cause. It’ll be just another human (probably a slow one), except for being not a human. It’ll be a long story before it’ll have any chances to do something impactful, and we’ll have plenty of time to figure out the details. Just like with drug precursors.
Also, meth is a bad analogy. And turns out the recipes can be found just about anywhere, I’ve just checked: https://kagi.com/search?q=how+to+make+methamphetamine so I don’t think “see how long it stays online” is valid either.
An analogy is gun control laws. Murder is a crime, whether it happens with a gun or without, so in theory there would be no need to regulate guns ("guns don't kill people etc."). But in most countries in the world we still regulate them.
Maybe it makes sense to regulate AI similarly? Requiring that some guardrails are built in to prevent it being used to generate fake information or information that helps to commit crimes. Making sure that it does not leak private information (and being able to prove this somehow). Regulating if and how it can be used as a therapist.
Although I wonder if it isn't too soon still for such regulation.
Comparing Netflix (and most highly profitable computer businesses) to the world of producing AI models by training is not going to be fruitful. Netflix takes a lot of effort to operate but you can do on the small scale what Netflix does, quite directly. You can't replicate an AI model like ChatGPT-4 very easily unless you have all the data and huge compute that OpenAI does. Now, once the model has been produced, you can operate that model on the small scale with maybe less amazing results (see llama.cpp, etc) but producing the model is a scale problem like producing high quality steel. You can't escape the need for scale (without some serious technological developments first).
Netflix cheats. They send non-supercomputer boxes out to ISPs to install locally. If I could convince every ISP to install a bunch of my media servers people could watch my shows from anywhere in the US too.
I don't think compute cost has dropped by 1000x since 20 years ago. Maybe by 10 to 50x. And if you add in the demand for higher quality, the cost has probably increased. Like encoding a video for streaming 20 years ago, at that standard, may have cost roughly the same as it does today, or more, when you factor in the increases in resolution and quality.
My prediction is that training the latest model will continue to cost millions to tens of millions for a long time, and these costs may even increase dramatically if significantly more powerful models require proportional increase in training compute.
Unless of course we have some insane algorithmic breakthrough where we find an AI algorithm that blows llama2 out of the water for a small fraction of the compute.
On a very good day, as many as three of those might simultaneously not be bad actors.
They aren't bad actors whose access to AI technology is likely to be meaningfully impacted by regulation (but, for certain of the non-US ones, that hasn't stopped the US from trying before), but that's a different issue.
Seriously? You don't think China, Iran, and North Korea are bad actors? What planet are you on?
But I don't see China or North Korea firing nukes or even blowing up western buildings. They are limited by the threat of response.
A rogue wacko in his basement can make Kim Jong Un look like Theodore Roosevelt.
I don’t think they’re society-ending, but that doesn’t mean we shouldn’t acknowledge them.
We’ve done that nearly every time there’s been a shift in media, from photography to telephones to home video to digital photo and video editing, and it just sort of happens organically and automatically as people learn what they can now do and what they can now trust.
Panicked, doomed attempts to control that sort of change is just divisive and futile.
That’s not to say that there won’t be necessary laws and regulations around near-term “AI” tech, but that deep fakes are well into the irrelevant noise of what people will learn to navigate on their own.
The idea that there's only one important problem is a fallacy.
I have made no such claim.
The people advocating for regulating AI are claiming it will solve all the relevant problems--i.e., that it will prevent AI from doing great harm. So pointing out a problem that the regulations will not solve is refuting the claims the advocates of regulation are making. That was my point.
We have tried this with China, going back to Nixon opening up trade relations in the early 1970s. It hasn't helped.
Examples:
In the Iran Hostage Crisis you had a constrained actor (Iranian government) making somewhat rational choices to use hostage taking as a negotiation tactic.
In the Oklahoma City Bombing, you had unconstrained actors (Timothy McVeigh and Terry Nichols) blowing up a building with a vehicle borne improvised explosive device for personal reasons.
I don't see why not. The whole point of regulations is to regulate, i.e., to keep the regulated activity within some particular bounds. If the regulation won't accomplish that, then it is pointless. Unless, of course, the actual purpose of the regulation is not the same as the purpose that is publicly stated--which is exactly what happens with regulatory capture.
"""
But what we are seeing too often is a calorie-free media panic where prominent individuals — including scientists and experts we deeply admire — keep showing up in our push alerts because they vaguely liken AI to nuclear weapons or the future risk from misaligned AI to pandemics. Even if their concerns are accurate in the medium to long term, getting addicted to the news cycle in the service of prudent risk management gets counterproductive very quickly.
## AI and nuclear weapons are not the same
From ChatGPT to the proliferation of increasingly realistic AI-generated images, there’s little doubt that machine learning is progressing rapidly. Yet there’s often a striking lack of understanding about what exactly is happening. This curious blend of keen interest and vague comprehension has fueled a torrent of chattering-class clickbait, teeming with muddled analogies. Take, for instance, the pervasive comparison likening AI to nuclear weapons — a trope that continues to sweep through media outlets and congressional chambers alike.
While AI and nuclear weapons are both capable of ushering in consequential change, they remain fundamentally distinct. Nuclear weapons are a specific class of technology developed for destruction on a massive scale, and — despite some ill-fated and short-lived Cold War attempts to use nuclear weapons for peaceful construction — they have no utility other than causing (or threatening to cause) destruction. Moreover, any potential use of nuclear weapons lies entirely in the hands of nation-states. In contrast, AI covers a vast field ranging from social media algorithms to national security to advanced medical diagnostics. It can be employed by both governments and private citizens with relative ease.
"""
Let's stop contributing to this "calorie-free media panic" with such specious analogies.
I think the true fear is that in an AI age, humans are not "useful" and the market and economy will look very different. With AI growing our food, clothing us, building us houses, and entertaining us, humans don't really have anything to do all day.
"Hackers aren't a problem because we have cybersecurity engineers". And yet somehow entire enterprises and governments are occasionally taken down.
What prevents issues in redteam/blueteam is having teams invested in the survivability of the people their organization is working for. That breaks down a bit when all it takes is one biomedical researcher whose wife just left him to have an AI help him craft a society ending infectious agent. Force multipliers are somewhat tempered when put in the hands of governments but not so much with individuals. Which is why people in some countries are allowed to have firearms and in some countries are not, but in no countries are individuals allowed to legally possess or manufacture WMDs. Because if everyone can have equal and easy access to WMDs, advanced civilization ends.
What on earth could this possibly mean in practice? Two elephants fighting is not at all good for the mice below.
Do we have reason to believe that giving the ingredients of AGI out to the general public accelerates safety research faster than capabilities research?
It’s no different than inviting an advanced alien species to visit. Will it go well? Sure hope so, because if they don’t want it to go well it won’t be our planet any more
People who are saying AI is dangerous are saying people are dangerous when empowered with AI, and that's why only the right people should have access to it (who presumably are the ones lobbying for legislation right now).
What does that buy us? An extra decade?
I don't know where this leaves us. If you're in the MIRI camp believing AI has to lead to runaway intelligence explosion to unfathomable godlike abilities, I don't see a lot of hope. If you believe that is inevitable, then as far as I'm concerned, it's truly inevitable. First, because I think formally provable alignment of an arbitrary software system with "human values," however nebulously you might define that, is fundamentally impossible, but even if it were possible, it's also fundamentally impossible to guarantee in perpetuity that all implementations of a system will forever adhere to your formal proof methods. For 50 years, we haven't even been able to get developers to consistently use strnlen. As far as I can tell, if sufficiently advanced AI can take over its light cone and extinguish all value from the universe, or whatever they're up to now on the worry scale, then it will do so.
I mean, once the discussion goes THIS far off the rails of reality, where do we go from here?
If we are talking about AI stoking human fears and weaknesses to make them do awful things, then ok I can see that and am afraid we have been there for some time with our algorithms and AI journalism.
at best, maybe it adds a new level of sophistication to phishing attacks. That's all i can think of. Terminators walking the streets murdering grandma? I just don't see it.
what I think is most likely is a handful of companies trying to sell enterprise on ML which has been going on since forever. YouTubers making even funnier "Presidents discuss anime" vids and 4chan doing what 4chan does but faster.
It doesn't take a whole lot of imagination to come up with scenarios like these.
More banally, state actors can already use open source models to efficiently create misinformation. It took what, 60,000 votes to swing the US election in 2016? Imagine what astroturfing can be done with 100x the labor thanks to LLMs.
[1] dx.doi.org/10.1038/s42256-022-00465-9
So you're saying that:
1. the religious nut would not find the same information on Google or in books
2. if someone is motivated enough to commit such an act, the ease of use of AI vs. web search would make a difference
Has anyone checked how many biology students can prepare dangerous substances with just what they learned in school?
Have we removed the sites disseminating dangerous information off the internet first? What is to stop someone from training a model on such data anytime they want?
this is a conspiracy theory that gained popularity around the election but in the first impeachment hearings, those making allegations regarding foreign interference failed to produce any evidence whatsoever of a real targeted campaign in collusion with the Russian government.
It’s true we are fucked if bioweapons become easy to make, but that is not a question of ”AI”.
> Essentially you are advocating against information being more efficiently available.
Yes. Some kinds of information should be kept obscure, even if it is theoretically possible for an intelligent individual with access to the world's scientific literature to rediscover them. The really obvious case for this is in regards to the proliferation of WMDs.
For nuclear weapons information is not the barrier to manufacture: we can regulate and track uranium, and enrichment is thought to require industrial scale processes. But the precursors for biological weapons are unregulated and widely available, so we need to gatekeep the relevant skills and knowledge.
I'm sure you will agree with me that if access information on how to make a WMD becomes even a few order of magnitudes as accessible as information on how to steal a Kia or how to steal a catalytic converter, then we will have lost.
My argument is that a truly intelligent AI without safeguards or ethics would make bioweapons accessible to the public, and we would be fucked.
Though with a tiny bit of Googling you’ll be able to find several Turing Award winners who are saying exactly what I’m saying. In public. Loudly.
However, the possibility of foreign election interference, real or imagined, is not a valid reason to hold back progress on AI.
With superintelligent AI you likely have to have every developed use and every end user get it right, air tight, every time, forever.
Can you imagine how much harder it would be to protect against hackers if the only cybersecurity engineers were employed by the government.
Don't let your limited imagination constrain you're ability to live in fear of what could be. Is that what you mean? So it's no longer sufficient to live in fear of everything, now you need to live in fear even when you can't think of anything to be afraid of. No thanks.
2. Accessibility of information makes a huge difference. Prior to 2020 people rarely stole Kias or catalytic converters. When knowledge of how to do this (and for catalytic converters, knowledge of their resale value) became available (i.e. trending on Tiktok), then thefts became frequent. The only barrier which disappeared from 2019 to 2021 was that the information became very easily accessible.
Your last two questions are not counterarguments, since AIs are already outperforming the median biology student, and obviously removing sites from the internet is not feasible. Easier to stop foundation model development than to censor the internet.
> What is to stop someone from training a model on such data anytime they want?
Present proposals are to limit GPU access and compute for training runs. Data centers are kind of like nuclear enrichment facilities in that they are hard to hide, require large numbers of dual-use components that are possible to regulate (centrifuges vs. GPUs), and they have large power requirements which make them show up on aerial imaging.
If that happened open efforts could marshal tens or hundreds of thousands of GPUs.
Right now the barrier is that training requires too much synchronization bandwidth between compute nodes, but I’m not aware of any hard mathematical reason there couldn’t be an algorithm that does not have to sync so much. Even if it were less efficient this could be overcome by the sheer number of nodes you could marshal.
In my mind the existential risks make regulation of large training runs worth it. Should distributed training runs become an issue we can figure out a way to inspect them, too.
To respond to the specific htpothetical, if that scenario happens it will presumably be by either a botnet, by a large group of wealthy hobbyists, or by a corporation or a nation state intent on circumventing the pause. Botnets have been dismantled before, and large groups of wealthy hoobyists tend to interested in self preservation (at least more so than individuals). Corporate and state actors defecting on international treaties can be penalized via standard mechanisms.
Here’s a legit question: you say pause. Pause until what? What is the go condition? You can never prove an unbounded negative like “AI will never ever become dangerous” so I would think there is no go condition anyone could agree on.
… which means people eventually just ignore the pause when they get tired of it and the hysteria dies out. Why bother then?