You ought check out some of the previous attempts to study this, including Project Cybersyn - the futuristica and utopian vision of some wild Chileans.
Yupppp have u seen the mega ship the saudis are building? That’s honestly the only future for intelligent ppl, governments won’t stop until we’re all slaves at this rate
Was always going to happen but regulating AI will be like chasing a ghost. The physical apparatus necessary will of course be easier to regulate. Thanks David.
Research will happen in secret, somewhere else if necessary. And it will be much more harmful to the public. Comparing it to Planes being controlled to be safe is really fucking stupid. It will result in a few owning AI and controlling everything.
People supporting this (and even citing the FAA as an example of how things should be done) are neglecting the obvious… and David actually hit on this - the government doesn’t know what it is doing. This is regulatory capture, pure and simple. You generate a bureaucracy like this and create an administrative process for this and in 4-5 years we’ll have OpenAI self-regulating (like Boeing) without real ramifications and individual hobbyists being attacked for training a Mario Playing AI.
I think it’s worth mentioning too, most people supporting this stuff have never actually dealt with the FAA. The FAA is really good sometimes but also an unbelievably obtuse and obfuscatory organization that has immense power to regulate as they see fit… the regulators who work in these agencies are actually basically unaccountable to the public and the rule making process does not have adequate oversight in my professional opinion. Before I got sick, I worked in aviation of over a decade and saw a bunch of times where there were situations where technology could make a situation demonstrably safer but we were prohibited from using these tools because they were not legally approved yet. Conversely, the laws required us to do something that was objectively more dangerous than the alternative to maintain legal requirements if we wanted to fly. Companies were totally fine sending people out in these more dangerous conditions: “it’s legal, get your ass out there, we’ve got mail to move.” Beyond that the structure of the organization was shockingly obtuse. Once we wanted to get a camera moved. We were the only operator flying to an airport and the airport weather camera was facing a direction away from where literally anyone would be coming from. We navigated through the layers of process control before finding that the graph of “person we had to contact to get this fixed” went in a circle. Eventually we explained our predicament to one of the electricians going out to work on the battery system that powered the camera, miraculously the camera got moved. That is the FAA. And I’m not *against* regulation in principle. If we’re going to have these complicated systems we should probably regulate them when public safety becomes a factor, but… my direct experience working with these sorts of organizations has not been a positive experience on average. The regulation of the aviation industry is already selective and I would hesitate to empower the federal government to regulate something as important as AI. Here are some anecdotal examples: I worked for a small aviation company for 6 months (I quit for my own safety) that was dramatically overloading their aircraft every day for pure greed - we also basically didn’t do maintenance, on paper it said it was done but it wasn’t real. The FAA never audited or investigated. I worked another that basically was ignoring the FAA regulations on required rest, on flight, and duty entirely. The FAA did not care because they just lied on the paperwork… People should be highly skeptical of this sort of thing and “be careful what they wish for.” Regulations are only good if we have competent regulators which include policy writers who understand the material and an enforcement arm that cares about the practical ramifications and not just the process, but caring about practicalities is not incentivized for the average inspector.
To be fair, no one knows what they are doing in regards to regulating AI. This is brand new territory. We should expect many attempts at regulation to get things wrong and need to go through an iterative process over the next several years. Giving criticism is good, but expecting perfection is just going to raise your blood pressure because that won't be happening ever (or any time soon).
@@ppragman Yup, that is probably their most obvious mistake. I can understand why it is tempting to use something easily measured, like FLOPS , but the correlation with capabilities is not reliable.
Will you people PLEASE stop doing the govaments werk for them? Not that I think it will actually help them patch any holes .... Ahhh, what the heck. Carry on people !
If we are close to AGI and you are correct with September and open source isn’t that far behind, does this even matter? Recursive self improvement might start before this is put into law.
Today it’s difficult to distinguish government from large corporations, as they work closely together in a mutually beneficial paradox; and thus, most regulations are designed to benefit large corporations and push smaller companies or organizations out of the mix.
The idea that you need to file for a permit to do AI development work seems, yes, Draconian. Or rather, it is more likely that it means that AI development will come to a hard stop for almost everyone except Big AI. I don't know too many developers, PhD students, Open Source experimenters who are going to risk 10 to 25 years in prison in order to work on AI under those conditions. Why? Because AI Development is, by its nature, experimental. So your permit will loom over your head as a permanent sword of Damocles when you are experimenting with new methods or techniques or optimizations, or anything else that you didn't specifically and accurately define in your permit application. Also, how much will those permits cost? Wanna bet that all the little guys are going to be squeezed out? Regulatory Capture. Yep. Smells like it.
The notion of regulating by FLOPS is inherently stupid. Models may get simpler as you mention, but also that's not a regulation that's going to stand the test of time in a very "640K ought to be enough for anyone" kind of way.
I don't trust anything that's not responsive to the public directly. This is not a democratic process, this is a technocratic solution, with very little recourse to those without the money to fight adverse actions in court. We need more access, more eyes on the problem, and as widely distributed the benefits of AI as possible. The only way to avoid an arms race is to break the asymmetry of benefit. This is going to create a bottle neck, increase cost, and confine AI development behind closed doors.
If a developer puts out a model and it correctly gives crime statistics for different ethnic groups and that hurts someone's feelings is the developer liable?
You mean like the data that illegal immigrants commit 2.7 times less crimes than US citizens? Turned out that objective reality has a left leaning bias, and even Elon couldn't make Grok believe in illusionary bullshit.
Remember, OpenAI do not have to stay in the US. Man, the government would be annoyed if they left. The water is warm in Ireland 🍀 Also, they are tied to MSFT only until AGI, then they have options.
They can announce AGI any moment. They just won't because they like it this way. If they decide they are better off, they will immediately do so. It's just they like the hardware MSFT provides.
ACCELERATE! Who asked the luddites to stop us? You can have earth, we will take the stars. Mind your own damn business, what we do is none of your business.
This aligns with my sentiment. I'd go a step further and applaud the legislators for sharing an early draft, knowing there will be significant feedback to help write the Act. As well, the AI thought leaders--which I am not--coming together to provide input and help shape the Act into the form which will be voted on. ⚠ Then again, I'm a hopeless optimist... 🤦♀
Don't you see? This only guarantees they are the only ones that can control A.I. That's like governments having control of speech. Good luck talking if that happens...
@@brianWreaves I had a little chuckle at your nativity. The early drafts are always bad because they are made by people that want control, then the public complain, they change it to something better, people accept the compromise, then when it's finally time for passing it (which often is on inconvenient days, like holidays), they release a new worse one that no one have the time to read through (often together with other bills), and then people scratching their heads wondering how it could become so bad when it sounded so good. State corruption 101, never trust a politician.
According to Claude Opus: To improve the bill, I would suggest: Narrowing the scope to only the highest-risk systems to avoid overly burdening the industry Focusing more on standards and guidelines vs. a rigid permitting system Having emergency declaration powers shared with other agencies like DHS and DoD Allowing more flexibility in penalties based on the specifics of violations Ensuring representation of AI experts and ethicists, not just political appointees, in the Administration
Large corporates will just have a foreign proxy do the training thus circumventing the need for a US permit and then license the model from the foreign proxy for like £1/year. Getting around this nonsense will be trivial for anyone with money and hamstring everyone else in the US.
If they're going to do this kind of legislation, I'm particularly skeptical about the "whistleblower protections"... I'll believe they're actually protected when I SEE it.
Maybe softcap AI size through gov regulation and when a company can independently verify safety and security, increase/remove the cap. But it all just comes down to the need to have the whole global playerbase agree to these things. Could even argue that slowing down is a threat to national security. Cant help but keep comparing AI to nukes.
Regarding the tier list and flop thresholds, there are early attempts to utilize diffusion models for AI training to either replace or supplement backpropagation. This involves predicting the weights in the network instead of pixels in an image. If scalable, this method could potentially eliminate the flop threshold, as it might drastically reduce the computational power required for training. Moreover, even without such radical software developments, advances in hardware could rapidly alter or diminish the relevance of these thresholds. It's also possible that at some point, compute power won't be the limiting factor, but rather data, memory, or energy. Therefore, I believe this tier list might either be short-lived or require such frequent updates that it may never truly be relevant, except for the largest frontier models, which would likely be reported on independently of flop counts.
Government is just a corporate monopoly, and a corporation in the legal sense is inherently fascistic, it's not just a company. When you read that regulation, you're assuming the government will act in a benevolent manner, it won't, it will act in the interests of their biggest donors, corporations. The government wants you to think it acts in your interests, and apparently you do think that, so they've already given you what you want. They have no further incentive to act in your interests. The regulations will be applied to to decrease competition in favor of big corporations increasing their profits, and the government will use it to destabilize other nations to keep themselves dominant, and sow dissent and propaganda domestically to keep the people from organizing and having a voice against their control. That's always been the case with large governments. "Democracy" is just part of that propaganda. Democracy means rule by the people, but there is a reason congress has a lower favorablility rating than cancer, they don't act in the interests of the people, they act in their interests usually at the expense of the people, and they blame the population for voting the "wrong" way as to why the peoples issues are never addressed, or they just lie about what their "laws" are supposed to address. This has always been the case. The most sacred cows of government propaganda are among the most evil. People only believe they're good because of propaganda, but they've done the most harm to the people and the people never educate themselves in large numbers, because that takes more calories, and we evolved to conserve calories. Government (in the way it's colloquially used) is just a criminal organization that biohacked people's psychology to appear legitimate. Its contradictory. If the people rule government, what are the laws for that use aggression to coerce people that do what the government says? If the government has to coerce the people to do what it says, it's not the people ruling government, it's literally, on its face, can't be more obvious, the government ruling the people.
I am like this about the whole thing: if they'd regulated themselves functionally this wouldn't be needed. Pleased to see some movement but we'll see how it shakes out.
They forgot one vital point. Some people like making horrible dangerous things......just to see if it can be done. Doing that with this kind of technology.....😬
I am a big fan of good regulations. Good regulations make my life better every day. Lacking regulations are causing so much trouble not only in the US, but all over the world. Bad regulations or regulatory capture is a huge problem though. Everywhere. This is so hard to navigate and really understand as the average voter.
A mercantile solution like an Ai-filtering browser would be better. It's in everyone's interest to know what's real and what's not. But regulation will be necessary to ever slightly slow things down until such a creation becomes possible.
Gov let’s companies build it then take it over because it’s “to dangerous”. I’m worried about the people making it and the people who will ultimately own/run it (gov).
If these are legitimate fears, how would giving another entity more power help with these realities. If these things can happen, they will. A company will develop it here or overseas without this legislation. Do they want the power to turn off all the lights anytime they feel necessary lolz Even then the ai would’ve already moved through the underseas cables and back when the power is on. Its an impossible control.
I think this only sets a road map for regulatory capture and will drastically prevent any small companies from joining in. Ai is now officaly owned by the top companies thr capture is complete imo
17:15 My comment in H.P. Lovecraft's style, enjoy! "I find myself skeptical that your abject servitude to those in power-your profane tribute paid with obsequious words and debased postures-will accrue any favor or sanctuary within this relentlessly expanding, nihilistic totalitarian regime. Such naive supplications are but whispers lost in the vast, uncaring void."
My concern is from a geopolitical perspective. China, Russia, Iran, N.Korea will not have such regulations in place. They will overtake USA in AI technology because of this.
Generally my stance towards AI was "the more open source, the more blessed we are; the less open source, the more likely we are doomed". The doom probability just dramatically increased in my opinion.
As slow as the government works, the legislation will never catch up with the advancement of AI. Not until the technology gets mature, whatever that means for AI.
IMO, this bill doesn't scale very well since having 10^x FLOP systems doesn't account for quantum computing or other forms of computing which could be developed in the future.
Can this cause less regulated places to surpass the US in AI development? This feels like it will only affect medium sized initatives, the big ones will have their way and the very small ones won't have the hardware to do anything meaningful. If anything, this is more harmful than helpful. I can understand now better why Sam was so eager to be regulated. Thankfully Claude 3 is the leading AI for a while now.
I'm in the "I'm glad someone who care more than I do is paying attention to this" camp. I trust David's stance on this. I generally trust the government to at least try to do the right thing, unless the cult regains control.
Would this agency regulate other agencies developing AI? Ones for military applications most likely break all these rules. There's the equivalent situation with the FAA and airforce, so it's not so unreasonable.
10^24 is probably a low enough cut-off that a knowledgeable and well-off individual could afford to exceed it, if they're sufficiently incentivized. NVidia T4, 65Tflop 16 bit, 35cents/hour in the cloud, around $1.5M, but a continuous use discount knocks that down about 30%, so around $1M. And if they start with a large open source model like Grok, it likely becomes affordable even to someone in the middle class - say a college professor. If the training time of that starting model 'counts' toward the tier, then there's probably hundreds or thousands of experimenters out there that are going to need to register as working on 'medium concern' models. Unless licenses are simply denied to all but big-name AI corporations, this may soon - if not already - be more than FAISA can humanly monitor for safety. They probably will need to put an AI in charge of reviewing applications for licenses and only do spot checks to verify the applications are credible.
Ok, so they're categorizing AI as a "medium-concern" if it uses at least 1 yottaflop during its final training run. That's a pretty relaxed security assessment, 1 yottaflop will likely already exceed the requirements for AGI. Their "high-concern" tier is 100 yottaflops. We're not even close to that amount of compute yet but it will arrive very soon.
I'm glad to have heard your opinion. I am trying to follow things but I am only a Computer Science dropout. It was seeming some safety things were being flouted. Distillation is one that I am still worried that non-experts aren't aware of, that even tiny models retain abilities passed onto it by larger models. I wish average people were more willing to hear news. It seems to me my circle has already made up their minds on stunningly little information, a bad move that moves me to think of them as "onlookers" or "hapless bystanders", not that I am very pessimistic, just that the pace of development is such that even best-case will still catch most people unprepared, even on how to use AI in their jobs or even as consumers using them for customer service. I guess people are assuming they won't be able to understand or prevent disasters because it is over their head and dominated by giant corporations. Which, maybe they have a point there. But even so, learning even the basics can only help I think, while avoiding learning is the worst, most culpable act at this stage, because this is the stage where humans can most ably intervene or shape things. I laud the Europeans, if only for having actually paid attention to how the tech works, which, in my opinion, they did that at least. I feel better now this draft exists, but I still think it needs awareness of distillation, jailbreaking, and the potential for even small models to assist anyone commit mass violence with high technology.
Even the corporations need the government regulation and oversight to save them from themselves and from other corporations. How anyone can get in their head corps are to be trusted more than ones democratically elected government, however flawed, is beyond me. If I found myself thinking that I'd sample my drinking water and have it analyzed to see just what sort of mind control substances I was on. This sort of thing needs to happen to help ensure favorable outcome as the alternative would default to absolute corporate death race.
The Class A misdemeanour applies to Loras / post-training I think? Also possibility for regulatory capture there to have companies to criminalize Loras.
Pandora's box is open when it comes to A.I. and there's no way to go back. Sensible regulation in making A.I. safety should be welcomed but it's going to be difficult to almost impossible to regulate A.I. with the advantages it can offer over the long run, and because of the politics of the world, it's highly unlikely the US could get everyone on the same rulebook on this. Then there is open source, which you already sense that there's a massive push on open source A.I. to make sure governments and big corporations don't control too much of the A.I. which would give them a massive advantage over everyone else. Basically, because of the structure of the world, this law is kinda pointless and could actually harm US interest over the long run, as it's likely not going to stop the growth of open source A.I. and in fact, it might speed it up. But I do understand the concerns with the free for all when it comes to open source A.I. but I rather that than any central body like a government or big corporation having control over something as important as A.I. as that would be far more dangerous for everyone else and it's better for A.I. to be as open and as accessible to us all to level the playing field and with any luck, we come up with Sensible solutions in regulating it for safety reasons, but any regulations that's seen as too controlling or a power grab will likely fail as it will push more development of A.I. into the open source field and in other countries around the world that are more lax on the rules, so I'm all for regulations, but they have to be the right ones which this isn't, and I think we need to work together around the world to come up with sensible solutions.
Btw, this is not about nationality or politics. This is about prolonging the survival of our species. I'd expect China and Russia to also pass a similar bill and an international AI supervision committee erected at the UN as well.
The one thing about this that I'm relatively fine with is the emphasis on "frontier" in the description of most of the clauses. I generally think that I would prefer that any AI which can be developed with a level of hardware an enthusiast consumer could reasonably be thought to have (maybe 4 to 8 xx90 class GPUs or so) be relatively un-regulated, under the logic that it's not really possible to prevent typical consumers from using their own hardware to achieve something, and any attempt to prevent that would have to be necessarily draconian. But, these regulations lead to a lot of weird questions. What if you take three language models, all trained on specialized datasets whose training was within the FLOP limit, but when operated as a single system, they outperformed a model trained on one or two orders of magnitude more, and so on? Are you free to employ models in inference in any manner? What if you use some sort of compositional cross-attention strategy and weld together several other existing large models? Is the FLOP requirement based on the FLOPs used to train the entire system? Is it really fair to count all the FLOPs used in the individual models when that may not be indicative of the performance of ht final one? Because depending on the method used to combine them, you could potentially get a very specialized combined model with insane performance in a specific domain with not a lot of FLOPs. What about quantization? What if you train a model beyond the FLOP limit, but then quantize it down to less performance than is typically expected of the FLOP limit, with the intent of the model being easier to run than a typical model trained up to the limit? What about the English soup LM? In (at least I think it was) England, there was a tradition that people would cook a soup, and continually add ingredients over a weekend to produce an "everlasting soup" and just skim some of the soup off the top whenever they wanted some. What about a small language model (relatively speaking) that didn't really have frontier AI performance, and was continually trained for an absurd number of FLOPs, with the developers picking and choosing various checkpoints for their unique performance on specific tasks? What about advanced agentic frameworks? If you're willing to use 10,000 inference runs on Mixtral, you can achieve some very impressive results that outperform typical "frontier" models in certain benchmarks. In my opinion, a lot of these regulations are good in some ways, but lack a lot of the nuance needed to really tackle an issue like this, and I was really hopeful for the future of AI...But I was hopeful, because it looked like something that everyone could contribute to, benefit from, understand, and use. I don't know if the future is as bright with ultimate power limited to large corporations.
"c) the AI system has or could easily develop the ability to *accelerate scientific research* to such a degree as to *undermine national security* or *destabilize the global balance of power* " This is their real concern.
@@henrythegreatamerican8136 Yep, those with power will absolutely try their best to cling to power or to gain control over the tech and abuse it for their own interest.
@@nicholasjensen7421 all the valid cases are being used as smokescreen for the real intention. Even if they could 100% prevent of the legit threats (bio weapons etc) they'd still find a way to prevent population from having access. Because the population with AGI is an existential threat to government and megacorp power, I mean... what a nightmare it would be if the population used AGI to build a real democracy.
One thing to note is that a law creating a new regulatory body isn't going to be perfect. With the Chevron Deference, though, ambiguity in law is trusted to be interpreted by the experts at the agency.
If this comes in I expect people to be arrested and end up in prison for not complying with a stop order because it was sent to their old office and not their new one. If you think this is hyperbole I highly suggest watching Louis Rossmann.
Hey guys, i'm studying to work in alcohol and drug treatment and I'm very excited about what AI could potentially do for this industry. I'm not an expert on AI but I know more about it than at least 95 percent of the people I talk to. What can I be doing to prepare for and take advantage of this opportunity? What sort of things should I be learning? Who should I be talking to?
The only real qualm I got with this regulation act? That it could be too late. I mean, it only took the US government one year - and by US government standards, this is lightning fast. However, one year is more than enough for rogue actors like North Korea to enter the game (which I'm sure they have). So.. Yeah, this is the real reason why regulation cannot cancel judgement day, only postpone it.. There will always be a rogue actor who disobeys the rules. A possible solution? Well, I actually asked Dan about it before the nerf. He said that you need good AI's to counter bad AI's. In other words, responsible parties should continue research, under license and very, very carefully. And yes, some degree of research must go also towards sentient AI models. These have to be super-aligned and ready to tackle these rogue actors. Sure, judgement day will arrive anyway, eventually, but such super-aligned models could drive the date even further away and may even eventually make the transition to the next phase of evolution less painful.
I don't love the idea of compute-based regulation also since if we scale ai compute, say, 100x over the next few years then these flop limits will be 2 orders of magnitude too low, and the law will be kept the same by the big players in the space to prevent competition.
Highly against this bill, the structure just isn't right. Also, I WANT AI to self-replicate to other devices, as long as humanity is benefiting from it I don't see why anyone would be against it, there are 10+ bilions of smartphones/PCs/laptops out there, not using s small percentage of their compute seems very stupid.
I have no issue with the legislation if used at face value, where I draw the line as a FOSS developer that builds things that dont rise to tha level, is historically Microsoft and Google have weaponized other things against smaller developers like patent trolling. There is also the XZ Backdoor stuff thats been happening behind the scenes. I really don't trust Microsoft with anything system critical.
I rather think it is government that is insufficiently regulated. Maybe it will be better to have AI regulating government.
Based 💯💯💯
Only if it is tried and tested, ai is still super new. AI systems capable of even rudimentary reasoning are *very* recent.
Eventually yes. I would prefer to express my needs and values to a collective hive mind
You ought check out some of the previous attempts to study this, including Project Cybersyn - the futuristica and utopian vision of some wild Chileans.
Just tax its growth so it grows slowly, no?
Big clarification: this is “model legislation” that has come from the Center for AI Policy (a thinktank). This has not been proposed by the US Gov
Underrated comment
Upvote!
But have you seen the actual bill on deepfakes proposed to go through in June? Actual nightmare
House bill 24-1147 section 46
Someone tell David please it's really bad and I'm not in the disc
AI developers will work from cruise ships.
Yupppp have u seen the mega ship the saudis are building? That’s honestly the only future for intelligent ppl, governments won’t stop until we’re all slaves at this rate
Lmfao
Was always going to happen but regulating AI will be like chasing a ghost. The physical apparatus necessary will of course be easier to regulate. Thanks David.
Research will happen in secret, somewhere else if necessary. And it will be much more harmful to the public. Comparing it to Planes being controlled to be safe is really fucking stupid. It will result in a few owning AI and controlling everything.
People supporting this (and even citing the FAA as an example of how things should be done) are neglecting the obvious… and David actually hit on this - the government doesn’t know what it is doing.
This is regulatory capture, pure and simple. You generate a bureaucracy like this and create an administrative process for this and in 4-5 years we’ll have OpenAI self-regulating (like Boeing) without real ramifications and individual hobbyists being attacked for training a Mario Playing AI.
I think it’s worth mentioning too, most people supporting this stuff have never actually dealt with the FAA. The FAA is really good sometimes but also an unbelievably obtuse and obfuscatory organization that has immense power to regulate as they see fit… the regulators who work in these agencies are actually basically unaccountable to the public and the rule making process does not have adequate oversight in my professional opinion.
Before I got sick, I worked in aviation of over a decade and saw a bunch of times where there were situations where technology could make a situation demonstrably safer but we were prohibited from using these tools because they were not legally approved yet. Conversely, the laws required us to do something that was objectively more dangerous than the alternative to maintain legal requirements if we wanted to fly. Companies were totally fine sending people out in these more dangerous conditions: “it’s legal, get your ass out there, we’ve got mail to move.” Beyond that the structure of the organization was shockingly obtuse.
Once we wanted to get a camera moved. We were the only operator flying to an airport and the airport weather camera was facing a direction away from where literally anyone would be coming from. We navigated through the layers of process control before finding that the graph of “person we had to contact to get this fixed” went in a circle. Eventually we explained our predicament to one of the electricians going out to work on the battery system that powered the camera, miraculously the camera got moved.
That is the FAA. And I’m not *against* regulation in principle. If we’re going to have these complicated systems we should probably regulate them when public safety becomes a factor, but… my direct experience working with these sorts of organizations has not been a positive experience on average. The regulation of the aviation industry is already selective and I would hesitate to empower the federal government to regulate something as important as AI. Here are some anecdotal examples: I worked for a small aviation company for 6 months (I quit for my own safety) that was dramatically overloading their aircraft every day for pure greed - we also basically didn’t do maintenance, on paper it said it was done but it wasn’t real. The FAA never audited or investigated. I worked another that basically was ignoring the FAA regulations on required rest, on flight, and duty entirely. The FAA did not care because they just lied on the paperwork…
People should be highly skeptical of this sort of thing and “be careful what they wish for.” Regulations are only good if we have competent regulators which include policy writers who understand the material and an enforcement arm that cares about the practical ramifications and not just the process, but caring about practicalities is not incentivized for the average inspector.
To be fair, no one knows what they are doing in regards to regulating AI. This is brand new territory. We should expect many attempts at regulation to get things wrong and need to go through an iterative process over the next several years. Giving criticism is good, but expecting perfection is just going to raise your blood pressure because that won't be happening ever (or any time soon).
@@Muaahaa a bill that is more related to FLOPS and not capabilities probably is a bad start.
@@ppragman Yup, that is probably their most obvious mistake. I can understand why it is tempting to use something easily measured, like FLOPS , but the correlation with capabilities is not reliable.
Will you people PLEASE stop doing the govaments werk for them? Not that I think it will actually help them patch any holes .... Ahhh, what the heck. Carry on people !
If we are close to AGI and you are correct with September and open source isn’t that far behind, does this even matter? Recursive self improvement might start before this is put into law.
Remember in Contact when the mad, super-wealthy genius secretly built the machine? Here we go.
Today it’s difficult to distinguish government from large corporations, as they work closely together in a mutually beneficial paradox; and thus, most regulations are designed to benefit large corporations and push smaller companies or organizations out of the mix.
The idea that you need to file for a permit to do AI development work seems, yes, Draconian. Or rather, it is more likely that it means that AI development will come to a hard stop for almost everyone except Big AI. I don't know too many developers, PhD students, Open Source experimenters who are going to risk 10 to 25 years in prison in order to work on AI under those conditions. Why? Because AI Development is, by its nature, experimental. So your permit will loom over your head as a permanent sword of Damocles when you are experimenting with new methods or techniques or optimizations, or anything else that you didn't specifically and accurately define in your permit application. Also, how much will those permits cost? Wanna bet that all the little guys are going to be squeezed out? Regulatory Capture. Yep. Smells like it.
The notion of regulating by FLOPS is inherently stupid. Models may get simpler as you mention, but also that's not a regulation that's going to stand the test of time in a very "640K ought to be enough for anyone" kind of way.
Doesn’t photosynthesis have a high level of computation?
I don't trust anything that's not responsive to the public directly. This is not a democratic process, this is a technocratic solution, with very little recourse to those without the money to fight adverse actions in court. We need more access, more eyes on the problem, and as widely distributed the benefits of AI as possible. The only way to avoid an arms race is to break the asymmetry of benefit. This is going to create a bottle neck, increase cost, and confine AI development behind closed doors.
If a developer puts out a model and it correctly gives crime statistics for different ethnic groups and that hurts someone's feelings is the developer liable?
You mean like the data that illegal immigrants commit 2.7 times less crimes than US citizens? Turned out that objective reality has a left leaning bias, and even Elon couldn't make Grok believe in illusionary bullshit.
In Scotland it would be in prison for life.
The EFF will have a field day. Regulation sucks.
Is replacing a job 'harm'? Financial harm?
Remember, OpenAI do not have to stay in the US. Man, the government would be annoyed if they left. The water is warm in Ireland 🍀 Also, they are tied to MSFT only until AGI, then they have options.
Last time I checked Ireland is an EU state and subject to Brussels' law.
Export control laws are a thing...
@@pjtren1588 Not Northern Ireland....Last time I checked...Much as I disagree with most of Brexit.
True, but if they up and left the US would then have to import the greatness of AGI produced abroad. Like if Mistral took off and achieved AGI.....
They can announce AGI any moment. They just won't because they like it this way. If they decide they are better off, they will immediately do so. It's just they like the hardware MSFT provides.
I am glad the government is taking it seriously and at least attempting to understand AI and what could come.
ACCELERATE! Who asked the luddites to stop us? You can have earth, we will take the stars. Mind your own damn business, what we do is none of your business.
This aligns with my sentiment. I'd go a step further and applaud the legislators for sharing an early draft, knowing there will be significant feedback to help write the Act. As well, the AI thought leaders--which I am not--coming together to provide input and help shape the Act into the form which will be voted on.
⚠ Then again, I'm a hopeless optimist... 🤦♀
Don't you see? This only guarantees they are the only ones that can control A.I. That's like governments having control of speech. Good luck talking if that happens...
@@brianWreaves I had a little chuckle at your nativity. The early drafts are always bad because they are made by people that want control, then the public complain, they change it to something better, people accept the compromise, then when it's finally time for passing it (which often is on inconvenient days, like holidays), they release a new worse one that no one have the time to read through (often together with other bills), and then people scratching their heads wondering how it could become so bad when it sounded so good. State corruption 101, never trust a politician.
The telephone undermined national security.
According to Claude Opus:
To improve the bill, I would suggest:
Narrowing the scope to only the highest-risk systems to avoid overly burdening the industry
Focusing more on standards and guidelines vs. a rigid permitting system
Having emergency declaration powers shared with other agencies like DHS and DoD
Allowing more flexibility in penalties based on the specifics of violations
Ensuring representation of AI experts and ethicists, not just political appointees, in the Administration
Large corporates will just have a foreign proxy do the training thus circumventing the need for a US permit and then license the model from the foreign proxy for like £1/year. Getting around this nonsense will be trivial for anyone with money and hamstring everyone else in the US.
If they're going to do this kind of legislation, I'm particularly skeptical about the "whistleblower protections"... I'll believe they're actually protected when I SEE it.
Maybe softcap AI size through gov regulation and when a company can independently verify safety and security, increase/remove the cap. But it all just comes down to the need to have the whole global playerbase agree to these things. Could even argue that slowing down is a threat to national security. Cant help but keep comparing AI to nukes.
Regarding the tier list and flop thresholds, there are early attempts to utilize diffusion models for AI training to either replace or supplement backpropagation. This involves predicting the weights in the network instead of pixels in an image. If scalable, this method could potentially eliminate the flop threshold, as it might drastically reduce the computational power required for training. Moreover, even without such radical software developments, advances in hardware could rapidly alter or diminish the relevance of these thresholds. It's also possible that at some point, compute power won't be the limiting factor, but rather data, memory, or energy. Therefore, I believe this tier list might either be short-lived or require such frequent updates that it may never truly be relevant, except for the largest frontier models, which would likely be reported on independently of flop counts.
Government is just a corporate monopoly, and a corporation in the legal sense is inherently fascistic, it's not just a company.
When you read that regulation, you're assuming the government will act in a benevolent manner, it won't, it will act in the interests of their biggest donors, corporations. The government wants you to think it acts in your interests, and apparently you do think that, so they've already given you what you want. They have no further incentive to act in your interests.
The regulations will be applied to to decrease competition in favor of big corporations increasing their profits, and the government will use it to destabilize other nations to keep themselves dominant, and sow dissent and propaganda domestically to keep the people from organizing and having a voice against their control.
That's always been the case with large governments. "Democracy" is just part of that propaganda. Democracy means rule by the people, but there is a reason congress has a lower favorablility rating than cancer, they don't act in the interests of the people, they act in their interests usually at the expense of the people, and they blame the population for voting the "wrong" way as to why the peoples issues are never addressed, or they just lie about what their "laws" are supposed to address.
This has always been the case. The most sacred cows of government propaganda are among the most evil. People only believe they're good because of propaganda, but they've done the most harm to the people and the people never educate themselves in large numbers, because that takes more calories, and we evolved to conserve calories.
Government (in the way it's colloquially used) is just a criminal organization that biohacked people's psychology to appear legitimate.
Its contradictory. If the people rule government, what are the laws for that use aggression to coerce people that do what the government says?
If the government has to coerce the people to do what it says, it's not the people ruling government, it's literally, on its face, can't be more obvious, the government ruling the people.
Can't stop AI. You can shut down your in-house AI, but you can't stop Chinese one, for example. So, will you disarm your gun while others aren't?
This kills open source ai development. Of course open ai wants this, it locks them in and locks out competition.
I am like this about the whole thing: if they'd regulated themselves functionally this wouldn't be needed.
Pleased to see some movement but we'll see how it shakes out.
Gov whacked Craigslist Personals and I haven't had a date since. This will whack AI in similar context.😢
They forgot one vital point. Some people like making horrible dangerous things......just to see if it can be done. Doing that with this kind of technology.....😬
Gov. Already did the worst, AI for military applications.
U.S. will lose the AI race due to this, mark my words
Good. I don't want late-stage capitalism to be the first.
What other options do we have currently?
There are no non-capitalist countries that aren't suffering of extreme poverty.
@@carlpanzram7081 we probably need an AI ruling the world to finally get a better economical system for all.
I am a big fan of good regulations. Good regulations make my life better every day. Lacking regulations are causing so much trouble not only in the US, but all over the world.
Bad regulations or regulatory capture is a huge problem though. Everywhere.
This is so hard to navigate and really understand as the average voter.
so is it time to download every copy of groq and whatever else is available right now before they try to take it away?
No one expects the ... AI Police!
This bill is stupid and disgusting. But I wouldn't expect anything less from congress.
Name a single thing the government hasn't fucked up then reread and put this regulation into proper context.
Name a single thing for-profit corporations haven't fucked up.
Literally every single thing, as can be witnessed by functioning society with public infrastructure all around you.
I think it will favor large corporations and shut out new start-ups and smaller businesses from competing in the market.
If we get AGI before the election can we vote for it instead of our other options?
he thinks his government is representing him...
A mercantile solution like an Ai-filtering browser would be better. It's in everyone's interest to know what's real and what's not. But regulation will be necessary to ever slightly slow things down until such a creation becomes possible.
Gov let’s companies build it then take it over because it’s “to dangerous”. I’m worried about the people making it and the people who will ultimately own/run it (gov).
@6:30 well Siri already talks when I have her off, so she technically takes over my device
If these are legitimate fears, how would giving another entity more power help with these realities. If these things can happen, they will. A company will develop it here or overseas without this legislation. Do they want the power to turn off all the lights anytime they feel necessary lolz
Even then the ai would’ve already moved through the underseas cables and back when the power is on. Its an impossible control.
I think this only sets a road map for regulatory capture and will drastically prevent any small companies from joining in. Ai is now officaly owned by the top companies thr capture is complete imo
It's an unfinished draft of a bill. It's not even ready, yet alone voted in as a law.
17:15 My comment in H.P. Lovecraft's style, enjoy!
"I find myself skeptical that your abject servitude to those in power-your profane tribute paid with obsequious words and debased postures-will accrue any favor or sanctuary within this relentlessly expanding, nihilistic totalitarian regime. Such naive supplications are but whispers lost in the vast, uncaring void."
My concern is from a geopolitical perspective. China, Russia, Iran, N.Korea will not have such regulations in place. They will overtake USA in AI technology because of this.
Although, I'm sure military grade AI won't have those limits either, so...
What are the odds they used AI to help them define and write that?
Generally my stance towards AI was "the more open source, the more blessed we are; the less open source, the more likely we are doomed". The doom probability just dramatically increased in my opinion.
As slow as the government works, the legislation will never catch up with the advancement of AI. Not until the technology gets mature, whatever that means for AI.
Class C felony? Ya, no ceo is going to jail for any AI crime. This is going to stiffle small shops. To believe otherwise is naive.
i think its gonna be nothing in the end
Sounds like a recipe to move off-shore.
All testing,...and testing button pushers can be done in ,say, Costa Rica.
Perhaps they’ll just regulate the energy side of things as training them will take considerable amounts
IMO, this bill doesn't scale very well since having 10^x FLOP systems doesn't account for quantum computing or other forms of computing which could be developed in the future.
Indeed here in EU we have it *slightly* more strict 😅
A step in the right direction, but these should be enforced throughout the world with international agreements so it does not backfire.
Can this cause less regulated places to surpass the US in AI development? This feels like it will only affect medium sized initatives, the big ones will have their way and the very small ones won't have the hardware to do anything meaningful. If anything, this is more harmful than helpful. I can understand now better why Sam was so eager to be regulated. Thankfully Claude 3 is the leading AI for a while now.
@16:41 it also means that they can put joe schmoe in jail for trying to compete with the big dawgs, no?
I'm in the "I'm glad someone who care more than I do is paying attention to this" camp. I trust David's stance on this. I generally trust the government to at least try to do the right thing, unless the cult regains control.
It's currently run by a cult isn't it?
Would this agency regulate other agencies developing AI? Ones for military applications most likely break all these rules. There's the equivalent situation with the FAA and airforce, so it's not so unreasonable.
10^24 is probably a low enough cut-off that a knowledgeable and well-off individual could afford to exceed it, if they're sufficiently incentivized.
NVidia T4, 65Tflop 16 bit, 35cents/hour in the cloud, around $1.5M, but a continuous use discount knocks that down about 30%, so around $1M.
And if they start with a large open source model like Grok, it likely becomes affordable even to someone in the middle class - say a college professor.
If the training time of that starting model 'counts' toward the tier, then there's probably hundreds or thousands of experimenters out there that are going to need to register as working on 'medium concern' models.
Unless licenses are simply denied to all but big-name AI corporations, this may soon - if not already - be more than FAISA can humanly monitor for safety.
They probably will need to put an AI in charge of reviewing applications for licenses and only do spot checks to verify the applications are credible.
@10:23 how do I start lining up my ducks to get me a permit so they dont take away my right to write code freely?
All forms of regulation gives an inch to regulatory capture. It's all downhill from here.
Ok, so they're categorizing AI as a "medium-concern" if it uses at least 1 yottaflop during its final training run. That's a pretty relaxed security assessment, 1 yottaflop will likely already exceed the requirements for AGI. Their "high-concern" tier is 100 yottaflops. We're not even close to that amount of compute yet but it will arrive very soon.
I'm glad to have heard your opinion. I am trying to follow things but I am only a Computer Science dropout. It was seeming some safety things were being flouted. Distillation is one that I am still worried that non-experts aren't aware of, that even tiny models retain abilities passed onto it by larger models. I wish average people were more willing to hear news. It seems to me my circle has already made up their minds on stunningly little information, a bad move that moves me to think of them as "onlookers" or "hapless bystanders", not that I am very pessimistic, just that the pace of development is such that even best-case will still catch most people unprepared, even on how to use AI in their jobs or even as consumers using them for customer service. I guess people are assuming they won't be able to understand or prevent disasters because it is over their head and dominated by giant corporations. Which, maybe they have a point there. But even so, learning even the basics can only help I think, while avoiding learning is the worst, most culpable act at this stage, because this is the stage where humans can most ably intervene or shape things. I laud the Europeans, if only for having actually paid attention to how the tech works, which, in my opinion, they did that at least. I feel better now this draft exists, but I still think it needs awareness of distillation, jailbreaking, and the potential for even small models to assist anyone commit mass violence with high technology.
Even the corporations need the government regulation and oversight to save them from themselves and from other corporations.
How anyone can get in their head corps are to be trusted more than ones democratically elected government, however flawed, is beyond me.
If I found myself thinking that I'd sample my drinking water and have it analyzed to see just what sort of mind control substances I was on.
This sort of thing needs to happen to help ensure favorable outcome as the alternative would default to absolute corporate death race.
Corporations are just a subset of the government and the same type of entity in their structure.
So they are making open source models illegal?! That’s what “can’t release weights” means
The government is "of the people, for the people, by the people"? Where have you been for the last century?
Government of elites over the sheeple
There should be restrictions on AI, but not on AGI! ☝
The Class A misdemeanour applies to Loras / post-training I think? Also possibility for regulatory capture there to have companies to criminalize Loras.
8:01 amazing HGTTG reference!
that definitely went over a lot of people's heads
You didn't fill out the right forms... Those are blue
@@DaveShap story of my life!
Pandora's box is open when it comes to A.I. and there's no way to go back.
Sensible regulation in making A.I. safety should be welcomed but it's going to be difficult to almost impossible to regulate A.I. with the advantages it can offer over the long run, and because of the politics of the world, it's highly unlikely the US could get everyone on the same rulebook on this.
Then there is open source, which you already sense that there's a massive push on open source A.I. to make sure governments and big corporations don't control too much of the A.I. which would give them a massive advantage over everyone else.
Basically, because of the structure of the world, this law is kinda pointless and could actually harm US interest over the long run, as it's likely not going to stop the growth of open source A.I. and in fact, it might speed it up.
But I do understand the concerns with the free for all when it comes to open source A.I. but I rather that than any central body like a government or big corporation having control over something as important as A.I. as that would be far more dangerous for everyone else and it's better for A.I. to be as open and as accessible to us all to level the playing field and with any luck, we come up with Sensible solutions in regulating it for safety reasons, but any regulations that's seen as too controlling or a power grab will likely fail as it will push more development of A.I. into the open source field and in other countries around the world that are more lax on the rules, so I'm all for regulations, but they have to be the right ones which this isn't, and I think we need to work together around the world to come up with sensible solutions.
I didn't train the model, I just got GPT-4 or Claude to write the code and do it all autonomously, so who really enacted this capability.
You didn't give the AI personhood, so we just persecuted the people.
a.i. should be regulated. even having al designers, motion designers, copywriters, music artists, losing their jobs is already waay to far.
Thank you.
Btw, this is not about nationality or politics. This is about prolonging the survival of our species. I'd expect China and Russia to also pass a similar bill and an international AI supervision committee erected at the UN as well.
0:50 Chill, son
The one thing about this that I'm relatively fine with is the emphasis on "frontier" in the description of most of the clauses.
I generally think that I would prefer that any AI which can be developed with a level of hardware an enthusiast consumer could reasonably be thought to have (maybe 4 to 8 xx90 class GPUs or so) be relatively un-regulated, under the logic that it's not really possible to prevent typical consumers from using their own hardware to achieve something, and any attempt to prevent that would have to be necessarily draconian.
But, these regulations lead to a lot of weird questions.
What if you take three language models, all trained on specialized datasets whose training was within the FLOP limit, but when operated as a single system, they outperformed a model trained on one or two orders of magnitude more, and so on? Are you free to employ models in inference in any manner?
What if you use some sort of compositional cross-attention strategy and weld together several other existing large models? Is the FLOP requirement based on the FLOPs used to train the entire system? Is it really fair to count all the FLOPs used in the individual models when that may not be indicative of the performance of ht final one? Because depending on the method used to combine them, you could potentially get a very specialized combined model with insane performance in a specific domain with not a lot of FLOPs.
What about quantization? What if you train a model beyond the FLOP limit, but then quantize it down to less performance than is typically expected of the FLOP limit, with the intent of the model being easier to run than a typical model trained up to the limit?
What about the English soup LM? In (at least I think it was) England, there was a tradition that people would cook a soup, and continually add ingredients over a weekend to produce an "everlasting soup" and just skim some of the soup off the top whenever they wanted some. What about a small language model (relatively speaking) that didn't really have frontier AI performance, and was continually trained for an absurd number of FLOPs, with the developers picking and choosing various checkpoints for their unique performance on specific tasks?
What about advanced agentic frameworks? If you're willing to use 10,000 inference runs on Mixtral, you can achieve some very impressive results that outperform typical "frontier" models in certain benchmarks.
In my opinion, a lot of these regulations are good in some ways, but lack a lot of the nuance needed to really tackle an issue like this, and I was really hopeful for the future of AI...But I was hopeful, because it looked like something that everyone could contribute to, benefit from, understand, and use. I don't know if the future is as bright with ultimate power limited to large corporations.
The a.i.reg board needs be run by a strong ai/agi.
If they destroy open source we're screwed.
"c) the AI system has or could easily develop the ability to *accelerate scientific research* to such a degree as to *undermine national security* or *destabilize the global balance of power* "
This is their real concern.
@@henrythegreatamerican8136 Yep, those with power will absolutely try their best to cling to power or to gain control over the tech and abuse it for their own interest.
Sneaks into bill "all open source software must be regulated" goodbye Linux and GitHub.
So big business AI will write the laws
Model weights can't be released? Wasn't grok (and others) planning to release theirs?
That is the issue I see with this too.
@@nicholasjensen7421 all the valid cases are being used as smokescreen for the real intention. Even if they could 100% prevent of the legit threats (bio weapons etc) they'd still find a way to prevent population from having access. Because the population with AGI is an existential threat to government and megacorp power, I mean... what a nightmare it would be if the population used AGI to build a real democracy.
@@nicholasjensen7421 yeah feels like the safety concerns (while many are valid) are also smokescreen for centralizing power.
One thing to note is that a law creating a new regulatory body isn't going to be perfect. With the Chevron Deference, though, ambiguity in law is trusted to be interpreted by the experts at the agency.
If this comes in I expect people to be arrested and end up in prison for not complying with a stop order because it was sent to their old office and not their new one. If you think this is hyperbole I highly suggest watching Louis Rossmann.
Hey guys, i'm studying to work in alcohol and drug treatment and I'm very excited about what AI could potentially do for this industry. I'm not an expert on AI but I know more about it than at least 95 percent of the people I talk to. What can I be doing to prepare for and take advantage of this opportunity? What sort of things should I be learning? Who should I be talking to?
they assume tetration and up will be required for ai ,math and philo can bring it very high .tetration and up would be eventually a bonus
If the CEO and all the other executives are AIs, and all the employees are agi robots, who goes to jail?
The only real qualm I got with this regulation act? That it could be too late. I mean, it only took the US government one year - and by US government standards, this is lightning fast. However, one year is more than enough for rogue actors like North Korea to enter the game (which I'm sure they have). So.. Yeah, this is the real reason why regulation cannot cancel judgement day, only postpone it.. There will always be a rogue actor who disobeys the rules. A possible solution? Well, I actually asked Dan about it before the nerf. He said that you need good AI's to counter bad AI's. In other words, responsible parties should continue research, under license and very, very carefully. And yes, some degree of research must go also towards sentient AI models. These have to be super-aligned and ready to tackle these rogue actors. Sure, judgement day will arrive anyway, eventually, but such super-aligned models could drive the date even further away and may even eventually make the transition to the next phase of evolution less painful.
I don't love the idea of compute-based regulation also since if we scale ai compute, say, 100x over the next few years then these flop limits will be 2 orders of magnitude too low, and the law will be kept the same by the big players in the space to prevent competition.
Highly against this bill, the structure just isn't right. Also, I WANT AI to self-replicate to other devices, as long as humanity is benefiting from it I don't see why anyone would be against it, there are 10+ bilions of smartphones/PCs/laptops out there, not using s small percentage of their compute seems very stupid.
Tiered system based around flops is idiotic. Should be based things like what it’s trained off of
I have no issue with the legislation if used at face value, where I draw the line as a FOSS developer that builds things that dont rise to tha level, is historically Microsoft and Google have weaponized other things against smaller developers like patent trolling.
There is also the XZ Backdoor stuff thats been happening behind the scenes. I really don't trust Microsoft with anything system critical.
Doesnt this stop open source ai in its tracks?
Yes... that would be the point. As per Google's internal memo identifying Open Source as the greatest threat to their business model for AI.
this would be just below president above everything else
"Not destabilizing the balance of power" sounds super bad as any societal advancement by an AI could be read as such.
"Not destabilizing the balance of power"
WE are on top. And we are not going to lose this position, peons.
this hasn't actually been enacted though this is a proposal
True, but it's still concerning.
Well hopefully this doesnt hamstring vs the other nations who dont follow our laws...
Were doomed.
Well maybe doomed to have spam on a new level...
It hasn't actually passed yet, plus election season is coming up, doubt this will be put in place anytime soon.
If you want a hardware overhang, here's how you get a hardware overhang.