Talk & transcript
AI and the Laws
Assemblymember Alex Bores
Transcript
Download textAutomatically generated transcript. May contain errors. Select a timestamp to watch that moment on YouTube.
[Music] International.
[Music]
That's right. [Music] [Music] What? What? Where's that? [Music] [Music] Oh, do you want to take pictures just with your phone? Oh, just your phone.
[Music] Um, a former coworker Ready? What? We're running out of time. Well, let's start. All right. [Music] Okay, everybody, we're going to start. Are you guys Can you hear me? If you want to see if you guys um [Music]
[Music] Okay. Quiet. Quiet. Quiet. Quiet. Quiet. Turn to your neighbor and say, "Neighbor, could you be quiet?" Or could you um There's actually like room outside if you guys want to uh Everyone turn to their neighbor and say, "Neighbor, can you be quiet? It'll work in a distributed fashion. We're getting a tad of feedback, too. Um, and can someone figure out how to turn off the music?
All right, we'll figure that out. Um, just bar. Um, thank you guys all for coming out to DCYC. Um, we're so honored to have you all here. Um, it also there's room outside. It's not as fun, but you can definitely hear um things going on out there. Um I'll just quickly go through these slides for everyone. Um so this is DIN NYC. We host these monthly meetups um about once a month. Sometimes we do them twice a month. Um and we have lots of people coming to them and it's about bringing people in New York together over science
instead of just like alcohol or run clubs or something like that. you can have like an intellectual discussion here. Um, here's some of our upcoming events. Oh, here's some of our previous speakers. Um, oh, I'll tell you that a second. So, here's our coming upcoming events. We have one on nutrition, transport, and cell metabolism, which will be really cool. Um, that is on August 12th, which is coming up. And then we have another one. Does anyone get that? Who wants it?
Okay, we have another one on embryo editing which is going to be very interesting. Um, thank you to Jana for organizing that one. And thank you to Conrad for also helping check people in. Um, this was very helpful. We wouldn't have be able to do this without you guys. Uh, so that's very much appreciated. Um, does everyone have that? Couple more people. Cool. Cool. Okay. Um, who here knows what decentralized science is? It's like using crypto and the blockchain to improve science. So, academics today have just so much going on. It looks like they're getting defunded at every turn. There's a lot of reproducibility
issues. What if you could use like crypto incentives to change how science is done? So, every time someone submits data to a data set, they earn crypto. or anytime someone does some analysis that's really interesting, they earn crypto. And if we the idea is if we change the incentives around how science is done, then you can actually change how science is done. Um, we can talk more about it. I'll run quickly through this. Um, we're we're conducting an experiment right now around sleep. So, in New York City, sleep is so important.
There's so much noise pollution. We're running the first ever distributed sleep study where people are earning crypto for their sleep. and we hope to create a like a a sleep science platform. So if you're interested um you can download the app uh here and you can uh take part in the world's first distributed sleep study. I think this is very cool and um it's like it's like imagine you're part of the first university in the world.
This is kind of like what that is like. How many people here are in the room already on this platform? Oh, okay. a couple a handful. Very cool. Thank you guys for signing up. Okay. Um Oh, and then one other plug just because um Oh, did Sorry, I didn't mean to. I'll go on the website. What? No. What? You have to get in here now. Yeah. Oh, and the other thing that I want to mention is um is we're hiring. So if anyone is interested in um is looking for a job as a programmer, as a designer, as a developer or as a business person, um scan that QR code
over there and it would be lovely to um stay in touch. Okay. Um and then one other slight plug. If you're interested in AI and regulation, um a couple years ago I wrote this book on AI and law. Um, I didn't have any copies with me, so Oh, we have one right here. Thank you, Anthony. It's signed. Oh, it's signed. Um, you can get this on Amazon if you want to learn more after the event. It's very cheap. It's like $10 or something. Um, uh, here we go.
Thank you. Um, so you can get the QR code there. Okay. Um, without further ado, um, co-host of this tonight is, um, David, who's with Freedom. So, I'll let David have a couple seconds. Uh, yeah. So thanks to Mikey and Dside for helping put this together and host the event. So as Mikey mentioned, I am David Gobod, the founder of a company called Freedom. And what we do is we create open-source local first and encrypted technology that lets you get your data out of big tech cloud and own it so that it's truly private and gives you truly private AI. And as I was working on this, uh, I discovered Alex and his
bills actually and wrote a little, uh, blog article or Twitter, um, and then emailed him because I wanted him to talk more, um, about what, you know, what he sees as, uh, the future of AI and how it should be regulated because I think it it's an extremely important question. And what we say at freedom is uh the only way to be free in the AI age is open source and private AI because otherwise we all basically will end up plugged into some closed source cloud AI. It'll be like being plugged into the matrix and you just sit on the wall
conveyor belt. Uh so I but it's very important how the regulation is done so that we also don't actually lead the AI age in the US. So I think it's the most important question really uh facing us today to get this right and I'm super excited to introduce assembly member Alex Fors and I want to thank you for doing this. He's the assembly member for the uh 73rd district which I think includes Murray Hill. is that that's sort of that area over there. Uh just a little bio before that he worked at Palunteer and I believe he also did I forgot the number some COVID fundraiser effort on you can say more
about that but thank you for uh being here and very excited to learn about your bills and your thoughts on the future of AI. Thank you. Thank you. How's it going? I think we do better than that. How's everyone doing? Cool. This is a packed room. This is a really cool event to bring people together to have conversations about science, about emerging issues. I mean, the range of topics that you've covered in past events has been quite broad. Uh but also each of the talks have been quite deep. So that is a rare combination. And to see this many people interested in having these sorts of conversations is a really exciting
thing. Um, just to get a sense of the room, how many people work in AI as you define it, broadly defined? Okay, cool. How many people work in politics? Couple. Half of which are my staff, but okay. Um, uh, how many people is this their first ESI event? Awesome. Really cool. That's really exciting to see. It's my first two, so we'll figure this out together. Um, I want to make sure there's a lot of time for Q&A. Um, I I believe, correct me if uh I'm wrong where Mike could go. Um, right in front of me. 20 minutes. No, no, no. I was going to say, am I the
first elected official that you've had here? Yes. Okay. So, I want to make sure there's a lot of time for Q&A on any sort of topic, AI, regulation, politics, government, whatever. Um, but I want to just tell you a little bit about Do you have his slides? Yeah. Great. Uh I want to tell you a little bit about who I am and and how I came to this. How I think about AI regulation broadly, how I think about the role of states in doing that and then focus a bit on frontier models in particular because I know that was a lot of what was put out here and probably what a lot of people want to hear about. So I want to make sure we have a lot of time for that. Sound like a good plan? Yes. Cool. Excellent. Let
me just check my time here. Great. So um my name is Alex Boris. I'm the state assembly member for the 73rd district which is much of the upper east side midtown east all the way down to Sutton Place. So 32nd Street to 93rd, second or third avenue to 5th Avenue uh plus Sutton Place. So in the 50s all the way to the river. How many of you are or might be a constituent?
Okay. So I got to be careful. I have voters in so that's good. Um but uh I worked in tech my whole career before I ran for office which is pretty unique. Um I was at uh yeah cool that comes up. Um I was at Palunteer for 5 years or so and then a couple of startups after that. Um the last of which was helping to sign up people for COVID relief which was being referenced before helping them keep their water on, their electricity on make sure that funds got to the people that it actually depended upon.
But I was always working in and around government technology. And so when this seat opened up, I had a lot of hard conversations and fun conversations with friends about do I actually want to run for office? Is this the sort of thing I want to do? I saw it as a chance to uh instead of being downstream of policy and trying to fix bad policy through tech, actually go upstream and design policy right the first time. And um you know, a friend of mine finally was like, "Listen, run for the seat. you don't know if you're going to win. And if you win, if you hate it, if you can't be effective in two years, in four years, you quit. But in two or four years, you can't say, "Now I'm going to run for the open seat." Like these opportunities happen when they happen. And so I
decided to run. I would do a side note that if any of you are interested in running, like please come talk to me. And also like I encourage you to again, what's the worst that can happen? You get there and then you get to have a fun time. But I decided to run. And during the campaign, I would talk about, hey, if I'm elected, I will be, and this is true, I became the first Democrat elected in New York State at any level with a degree in computer science, which is it's a fact that I find 10% cool and 90% horrifying. Um, but but it is a true fact. And also when I'd say that to some people, you might
be surprised you that uh the reception wasn't always what the reception was in this room. Sometimes when I'd mention that to voters, they would say, "What what is computer science have to do? What does AI have to do with government?" Um and then of course Chad GPT comes out right around election day, right around November 2022. And I haven't gotten that question since. uh I work on a lot of things as a legislator but given my background and given the speed with which this is moving tech has become a large portion of that portfolio. So uh I want to talk about um uh AI broadly um as I said and then the states and then the frontier models. But when I think about AI and for those of you that work in it, apologies at how
basic this will be and for others I hope this is interesting. But I think the thing that most people miss about it is that this is going to move faster than any technology that we have ever developed as humans. And the bounds of what is possible are as wide as any technology that we've ever developed as humans. And the only thing that comes close and I think it's useful as that analogy is nuclear fishision. Right? If you place yourself in the 1930s and there's this new technology coming and on the one hand people will tell you this will lead to you know utopia because we will have clean unlimited nearly free energy for everyone through
fision or or fusion was promised at the time uh which we're still waiting on right that's the upside and the downside is you know Armageddon and nuclear apocalypse and bombs raining down but both of those in the 30s and the 40s were reasonable thoughts as to what the future would happen on any time frame and obviously we've charted some path that is neither of those things. AI is itself that broad and so I don't like thinking of individual legislators or individual people or individual you know uh views as oh you have to be pro AAI or antiI. When people ask me should I be excited by AI or terrified of AI I say
yes. uh and that actually I think leads to a useful breakdown in the kinds of policies that all of us are trying to pursue. There's basically three questions that I ask when I'm looking at an AI policy. I would encourage you to do the same. So the first question is is this policy optimistic or pessimistic on AI? Right? Is it thinking, hey, this is a useful building block and maybe we just need to clean up the edges or maybe we need to encourage adoption. That'd be optimistic. Are you mostly worried about protecting the downside risk? Is it short-term or long-term? Is it dealing with things that are already here right now? Uh it's dealing with bias. Is it dealing with discrimination? Is it dealing with data transparency? Or is it
dealing with long-term? I don't necessarily mean things that are, you know, way out in the future, but perhaps where it's going and trying to skate to where the puck is headed. And that's what I think the frontier model bill is. And then whether it's use case specific or whether it's general. Is this something that is just applying to a particular industry or is it applying to AI models themselves? And I think we have a lot of philosophical debates where people stake out one of those eight potential places and say that is where all regulation should be. Um and I think that's a pretty naive way to look at the world because in general we should be focused on bills in each of those places. So I have bills that are working on um giving uh encouraging the
adoption of an industry standard like C2PA that is a free open standard that industry has developed to establish content provenence which is an optimistic general short-term bill right I have bills that are focused on training data transparency to know what's going into an algorithm which is a short-term general probably pessimistic bill um and then I have things like on frontier AI models where I'm really focused on what is the risk that is coming from some of this AI development and how can we ensure that we keep us all safe while we get the many many many benefits that we all want from AI right so that's where I'm going to focus today um and I'm going to use a little bit of my slides here on this
thank you okay we'll scroll a little the punchlines will be killed as I scroll but as long as you just laugh just as hard as if it had popped up in a transition but I can scroll oh thanks Thanks S. Um so why why am I focused on frontier model policy now? Well the slides 7 months ago anthropic said government urgently needs to take action on regulation in the next 18 months. Right? So that gives us 11 months left. That's April of 2026. We prefer it at the federal level. The federal level isn't moving fast enough.
So we'll take it at the states. Uh scroll down to the next one. Uh last year California worked on a bill SB1047 that passed legislature overwhelmingly but then got vetoed by the governor. He impanled a commission to study the question of AI and say what should be done at the state level. And two months ago that group came out and said no no no go ahead go up. Uh that group said you know policy windows do not remain open indefinitely. That with AI moving as quickly as it is if you set standards right when it's starting that everyone can use well then the cost of adoption is pretty cheap. But if you wait too long and people get uh on specific paths and you have path dependencies built in, maybe you can't actually do regulation
that would have made sense and they encourage action to happen right now. And then of course uh this says nine days ago, but it's 8 days ago. I'm giving this talk tomorrow, but if you scroll down, uh 8 days ago, chat GPT when told, hey, allow yourself to be shut down, said intercepted and didn't allow itself to be shut down. Um uh so we need action now. Um so let's talk about why the state level is important and then we'll talk about what we're trying to do about it. Um yep. Go to the next slide here.
Oh, okay. Sorry. No, thank you. Um so when people think about policy, most people think about the federal government. Obviously, um, most people don't think about state level regulation or Well, how here's a fun test. And remember, the camera's on me, not on you, so you can be honest on this. Raise your hands if you know who your state assembly member and state senator are. Okay, we got a few about the same numbers of people that work in politics.
So, uh, so but and that's fine. That's normal. That that's the way most people think about these things. But actually most of the laws that you live under, most of the laws that are passed every year, most of the things that directly impact your life happen at the state level. Um New York uh last year I think passed some 600 laws just last year. Uh Congress as a whole in 2023 passed 27.
Uh so most of what's going to impact how we work actually happens at the state level, which is a good reason to be focusing on it. If you are someone who wants reasonable regulation, whether it's encouraging AI adoption like my C2PA bill, or it's worrying about the risks like my frontier model bill, I would encourage you to focus on the state level. You get 50 bytes at the apple instead of just one. You can actually talk to your legislator a lot more easily than you can to your federal one. Uh there are different coalitions that are possible at the state level that aren't as entrenched at perhaps the federal level where you have two parties that will never talk to each other and always operate that At the state level, we tend to be much more cooperative and actually can get interesting things through. And frankly, we move faster.
Getting a bill passed at the state level is generally easier, which is why they're referred to states as the laboratories of democracy. We can try these things out. I don't think any of my bills are necessarily the right one forever for every state. But I think they're steps in the right direction, and if they're wrong, we can adjust them. And if they're right, I would love nothing more than the federal government to take them on and to pass them. Right?
that is the laboratories of democracy functioning. Um so let's talk a little bit about frontier uh models and the focus here. So these are current proposals. I know you can't read this in the back. I can't read it in the front but these are current uh proposals at the state level to regulate frontier models. I think there's eight that were there at the start of session. Um open to feedback if people know of other ones. Um but we'll go through very quickly. um they're listed in order of when their session ends. So these first four basically session's already ended. These bills are not passing. Colorado on May 7th, Vermont on May 30th, Illinois May 31st,
Texas on June 2nd. Um different approaches on how to regulate AI. um with uh Illinois really focused on transparency around the safety your safety and security plan uh I think was the had the most chance of passage this year. Colorado is really just focused on whistleblower protections. Um Texas started off as a really broad frontier bill but was edited down to not deal with frontier models. So listed but not there. Uh and if you scroll to the next slide, these ones are all in some ways still live. I'll work backwards.
Massachusetts is the last November 19th. It is basically a copy of SB1047 um for those that follow that debate last year, but we'll see how that evolves in Massachusetts, California, where SP 1047 happened. Uh they the same senator Scott Weiner has a new bill that is basically exclusively focused on the whistleblower protection piece. Making sure that if you work at one of these labs or you're a contractor for one of these labs and you see a truly catastrophic risk that you can raise that up without fear of losing your job even if what you're raising up is not behavior that is illegal but is that risky for everyone. Uh Rhode Island has a very different approach but one that's interesting which is strict liability
for AI models. So it basically says government isn't going to tell you how to do anything, but everything that goes wrong is your fault. And so you just internalize all of the risk. Uh we trust you to make the right determination on what it is because you're the experts, but we're just going to be clear that you are responsible. So do with that what you will. Um and then you have my bill in New York, the Raise Act, which is where I'm going to spend a little bit of time here. Great. So the raise act is focused on frontier model developers. So you need to have spent $und00 million in training frontier models. Frontier models are further defined as either $100 million on one model and 10 to the
26 flops uh mathematical calculations in order to generate the model. That's what I mean by flops. um or you were created via the process of knowledge distillation from another frontier model. And knowledge distillation again for those not is a very specific process where you were taking the final nodes of a neural net and using that as output to train a smaller neural net based on a larger one. Um if it's the process by which deepseek was created. uh if you were doing that specific process to make a nearly as capable LLM from another frontier model and spending 5 million to do that that would also apply but largely regardless of which method you
use it only applies to companies not academic research that are have spent $und00 million in training costs for their models. uh and it asks you to do four things which are largely in line with commitments these companies have already made to the Biden voluntary commitments and to uh the sole commitments commitments many ones that they've already made publicly which is they have to have a safety plan they have to have that safety plan audited by someone who isn't them doesn't it's not government I'm not insisting that but someone who doesn't have this sort of direct financial interest they need to disclose critical safety incidents so things that are truly raising the risk of something happening, something bad happening. We define that very strictly
in the bill. Uh and then they have to not fire employees or contractors that raise up catastrophic risk that whistleblower protection. Largely what this is aimed at is saying the cigarette companies, the tobacco companies that knew their own testing showed that cigarettes cause cancer and then publicly they went out and said it doesn't cause cancer and they kept selling it. or the oil companies that their own research told them, hey, this is going to cause climate change and then they went and testified publicly and said doesn't cause climate change. Let's keep selling it. This is saying you write up your own safety plan ahead of time when you don't have economic pressures. You state what would be a reasonable and unreasonable risk and if your own testing comes back by the standard that you've set up and
says it's unreasonable, you don't get to change that after the fact, right? That you're actually held to that standard that you designed in the first place. Um, yeah, I think let's go on to where it is. So, uh, this bill is something we've been working on for almost a year at this point. Uh, I generated the first version back in August based on a lot of the feedback to 1047. I actually accepted about 90 95% of the feedback that was out there. I then sent a version of it to six or seven major labs, got feedback, got red lines on it, made a new version in December, sent it to them again, got more red lines on it, introduced this bill in March. We've
been working to pass in this in the legislature, which again, our session ends June 12th for the Senate, June 17th for the Assembly. Don't ask or do in Q&A. Um, but uh today in Albany, I was there. We just had the debate. I just got off the train. Uh it passed the science and tech committee in the assembly uh so we're on our first step there. It'll go to the codes committee hopefully next week uh and then eventually be scheduled for debate on June before June 17th. The Senate, it is still in the Internet and Tech committee uh but we are similarly hopeful that it'll move there and the Senate tends to move bills more quickly. So before I open it up to questions, I'm coming right up on 20 minutes here. Uh I want
to do a quick pitch. You have four minutes. Sorry. You have four minutes. Thank you. Uh I want to do a quick pitch on how anyone who's interested can help. Um which is legislators can seem really far away and really disconnected. Um I won't speak about the federal one, but at the state level, we really do respond to input and to feedback. Uh, I can give you countless examples of bills I've changed of mine because of feedback, conversations I've had with colleagues that have le led to bills being changed.
Um, bills I even even decided not to do after some feedback. Uh, and votes that have been affected based on hearing from constituents. And so, if this is a topic that interests you, and since you're here, I assume it's a topic that interests you, uh, a you can always reach out to me if you have a specific thing on my bill. My email, it's not on there, but we have cards and it's my last name and then my first initial. So, Boris ANYsembly.gov. We go through everything.
Send any feedback. Um, if this is a concept that you find interesting, and I'll I'll say again, whether or not like if you're not 100% in agreement with the bill, but you're 80 90% in agreement with it, send me an email with all the things you'd like changed. We'll look through it. But also, we're in a really critical point where we might get this or nothing. And so your input, especially if it's to someone other than me, uh, because I'm in support of the bill, but if it's to any other legislator, can really make a big difference. And I want to in I want to emphasize to you how much you have power in this circumstance to really change where regulation of tech is going. And so, uh, you can, the first QR code will
take you to a page, uh, sorry, the second QRQ code will take you to a page where you can email your legislators directly. Uh, it'll you put in your address. It will tell you who your assembly member and senator is and you can send them a message to support this bill. The first one, if you are part of a nonprofit or an academic institution or a company, you're some part of some organization and can speak for that organization as to why you would like a bill like this. That will help generate what's called a memo of support which goes to our internal staff experts to say, "Oh, hey, people in industry are caring about this and looking at this."
So, do both. Do either. Uh, also always email me. uh you have I I think to go back to the theme of decentralization of science and how individual people are going to have a lot more power going forward. You as individuals have a lot of power in what happens in the next few weeks in terms of regula regulation of tech, regulation of AI, are we building a system that protects us from the worst while enabling and empowering the best?
Um and I welcome you all to that fight. So thank you for coming to this. Thank you for being involved. Thank you for reaching out and I'm really excited to answer these questions. Uh so thank you. So yeah, I just I'll start with just two questions. One is there is the uh federal proposal to enact a 10-year uh ban basically on states uh regulating AI because of the concern of if you end up with 50 different state laws, you know, how can you make an AI app or foundation model and the concern is it'll really hold the US back like you already look at Europe and Apple and companies are not
releasing features over there because they don't want to deal with the locks. And then uh I think the second really important question is on export controls, right? So you you had the what was known as the Biden diffusion rule. It sort of you know limited the export of call it the H100s or the leading Nvidia chips to China. And you know there the argument is yeah that's the right thing to do. You don't want to give them like nuclear bomb tech. On the other hand, you know, it's unclear how behind China is. They're they have now, I think, the leading open source models.
They've got five of them. What's super interesting is that China has adopted the open- source approach, which is what the US really did sort of with the Linux and during the internet age, which let us win. The concern is here, you don't give them the Nvidia chips, they will develop their own and then the rest of the world that's cut off will then use the China chips. It incentivizes them to spend tons of money and resources developing their own chips that will become better and like Huawei become run the rest of the world. And I think even with the open source models today, it's a great risk if you know
China trained models end up running the majority of the world because I think that's what will happen. Um so I'll do them in reverse order. Uh I agree. I think it'd be bad if China made ships run the world. Um, and I don't know that there's as direct a link perhaps between the export controls and that. Uh, if the export controls were only the US and no one else can have them, you start to have more of that risk. If you think about how you engage allies with it and give access, maybe you can build a system that's uh that's broader. You can think about what we are able to do with the dollar in the financial system where you know if you play by the rules and you
there's certain privileges that come with that. uh that the current administration seems uh aiming to take down. But um so generally I think export controls on chips to China are something I support, but they can't be your entire strategy, right? You also need to follow that up with deep investment in AI and AI research here in the US and making sure that we are finding ways to get more renewable energy and to get transmission lines so that we can build data centers so that we can actually do some of this research. uh in New York we uh did something called Empire AI which is we put a big investment in buying a number of the chips and making them accessible
at cost of running them not a cost of buying them to universities and so we've been able to accelerate the research of universities in New York and over time we're expanding out of who has access to that so you need both sides of it you can't just do the restrictions the export side on the federal piece um I agree this should be done at the federal level. Full stop. Come to me when it is.
Um, and I think what doesn't work is saying we're not going to do it at the federal level. And also, none of the states can do it, aside from it being blatantly unconstitutional and violating the rules of the Senate in terms of having a fiscal impact and many other things that may or may not matter in the current world. Um, you can think about the debate over say Obamacare where forever one party was able to run on we're going to replace it with something better. And everyone loved that. And it's easy to say, "Hey, it shouldn't be 50 states. It should be one federal standard." Great.
What's the federal standard? I think you saw the debate with Obamacare when it came to wait, what's the replacement? They were like, "Oh, wait. We don't have one. Okay, we're going to keep this." Uh, so if any Congress member or senator who's watching this wants to take the raise act and pass it, I would love that. Please do it. Uh, if you want to have a conversation about changes, great. Let's do that. But to say you can't do any, we're not going to do anything. And also, you have to bury your head in the sand is like a real way to ensure that we are not taking the lead in AI.
Should I call on is that? Okay, go ahead. So let's say AI goes sensit right. Uh who takes liability there? What's what are your thoughts on that? Uh so I I think that's interesting with the Rhode Island proposal which is like strict liability for the developers, right? My proposal doesn't change liability after the fact. If if you develop a thing and it goes wild and it what I'm focusing on is killing a hundred people or a billion dollars in damage, like courts already know how to deal with that liability. uh that is that is there. Now, if an AI model is sentient, I don't think you're going to find a court that says the developer isn't responsible for that because you're telling me it's sensient. Um I
think there's still some implied responsibility there. But it is an important question of when you release a model out into the world and someone modifies it, how much is the original deployer? How much how much is the original developer? How much is the person who deploys it? I think those are tough questions that that Rhode Island bill is trying to resolve and I have a separate one thinking about that but not one that that I'm emphasizing right now.
Yeah. Um, cool. I'll jump over here. Okay. I think it's great that we have you. So, I'm a data scientist and founder in medicine. I will say I think this is more regulatory capture sort of legislation. And I'd say that because you're focusing on model development and not the application layers. Model development. This is exactly what the anthropics, open AIs and Microsofts want. This is an anti-D really it's anti- decentralization. You want to talk about a practical application uh level legislation force these companies to align out real images of people out of the generators right we had company we had open AI sort of self-regulating this early image generation you could not get
a real person out that was smart of them then Elon Musk comes around with grock blows that open now they react and now you can get real politicians and real people out of chat GPT that is a very practical application layer Stephanie could take why are we focusing on on flops and prints. So sorry the the second part is government should require that the frontier models make real images of real people. That's application layer, right? Why focus on model training when you should be regulating application layer? Oh, because you should do both is the short answer. But but it would come to great news to Meta and Openthropic that they want this and love it because they've been very public that they don't want it and don't love it. So, um, uh, they
would be interested in that take. Uh, they don't want them to have to put out a safety plan and be really held to it has been the experience and what they're saying publicly. Um, there's always a threat of regulatory capture, right? And and that's why it's important to have real conversations that are based on the details of the bill and the nuances of this bill. Um, I I I I have heard the argument that they love it and they make the argument that they don't.
back over here. Food for thought. Um, what what do you think the impacts would be if we held investors liable as well as developers? Just out of curiosity. I don't know. That's a that's a really good question. I I think I would dive in again to liable for what and when, right? And and I think that you don't want to create a situation that dissuades people from investing in new risky ideas. I think that's a very important thing to develop. Um there's certainly some standard by which they're not doing any due diligence and putting
money into a thing that maybe it crosses over. But but that is a thing you want to be careful of um because you want to encourage a lot of competition and a lot of robust aspects. And that's why this bill only kicks in when you have trained a frontier model of a hund00 million. Right. This is going to apply to 8 to 12 companies right now. Um I want to encourage a lot of innovation at a smaller level.
Maybe we'll yeah go to the back here. I don't I don't know if the mic can go. Go ahead. Hi. Um I just wanted to ask about the enforcement mechanisms in the bill. I also wanted to ask um if any of the folks that would any of the companies that are going to be under the jurisdiction of this bill donated to your campaign and I'd also like to know if um or just what your game plan is for you know the anthropic CEO just said that white collar entry- level jobs are going away. Yeah. What is your plan for the constituents in your district? What jobs you going to create for those white collar workers that are being displaced?
Yeah. Cool. Um uh I have gotten contributions from companies that would uh be regulated by this and by trade associations all of whom really oppose this bill. So there's that. Um but um uh but yes, they've been a very very small part of my fundraising and because I I talk to everyone and I prefer from people. But um but yeah, no, that's that's a great question and it's public record. uh the enforcement mechanisms are through the attorney general. Um so the attorney general can bring a civil suit if if aspects are uh brought here. It's not opening up a
private right of action or allowing for sort of that harassment. Um and I'm sorry the last question was uh in your district white collar. Oh yes. No, I I think I'm incredibly worried about what this will do to jobs in the long term because throughout history as you have created new technology, it has destroyed some jobs and created new ones. But the rate of change of a person's job was usually once a lifetime, right? You would for most of human history, you were a farmer. But but but even you would do one job your whole life and if your job got disrupted, okay, you could retrain to a new job, you'd be good in that job the rest of your life. Uh and the next
change would come afterwards. Obviously, none of us are in one job for our lives anymore. And even so, we're entering a point with AI, which is a step beyond that, which is you get replaced in a job by AI, you go try to retrain for a new job. AI can train faster than you. So by the time you're ready for that new job, AI is also ready for that new job. So what what no what does that mean for a broad workforce? I I don't have the answer to that. I think it's some bit of ensuring that more people own capital in what they are building whether that's through equity that workers get whether that's
equity in current licenses and how we transition that. I think that there are some really unique interesting ideas there. people that have gotten licenses in trades and then that trade is no longer being used. Can you convert that to equity in some in some way? I think UBI is going to be part of the conversation. Uh it is it is a an unsolved problem. I would love I would love deep thought on Yeah. Thanks. Cool.
Hi, thank you so much for taking the time. Credible speech. Um I wanted to ask um so I'm writing my I wrote my thesis. I just graduated. Congratulations. Thank you very much. um on regulatory capture and regulation in the EU and the US and I looked at how the U AI act u actually wholly um implements national security and defense applications and of course the US doesn't even come close. So I know of course at the state level, city level we're not looking at defense um whatsoever we can't. Um but I guess I wanted to think about I have two questions. One is what's the ideal what do you I mean large question but what's the ideal defense um regulation state um beyond what we have with
voluntary and um two is a lot of this is due to lobbying right I mean mine was just tracking the sheer money and how do you think of like improving US lobbying reform? Um, sorry. Really big questions. No, no, no. Thank you. Questions after my own heart. Um, yeah. Uh, [Music] uh, what is the ideal regulatory structure when it comes to national security is a really big question and I think it probably involves some conversations in classified rooms. Um, uh, so I I'm going to I'm going to skip that one if that's okay. It's just not where I'm spending time. Um, but on on
regulatory on on lobbying reform and all of that, like, yeah, we should have really strong campaign finance reform. We should have uh strong incentives to take small dollar donations. We should have matching funds for contributions that are coming from real people. Um, we, you know, I don't agree with Citizens United. I don't get a vote on the Supreme Court, but I think it's sort of crazy that we have this fiction of, oh, a wholly separated thing is spending money and definitely no elected official has any knowledge of that or it's influencing their behavior in any way because it's an independent spending.
Like that's a that's a fiction. And um so there's a lot of changes, some of which are easier to make at the state level because the federal constitution empowers states to regulate their own elections. So, uh, we have done some of that at the state level. I think we should do a lot more. Um, I think it's a a much harder push at the federal level. Um, but but but I will say one thing, money in politics is bad. We should have less of it. Lobbying is uh sometimes bad. Sometimes it's good. Like advocacy groups that come up, right? That's lobbying, right? And and sometimes they bring really interesting perspectives.
But on the whole, yes. Uh what I don't want people to walk away from is a belief that there's like capture of everything that is passed. Um spend a little time in Albany, spend a little time in DC, less so in DC. Spend a little time in Albany. Uh you'll see bills passed every day that the big lobbyists hate. Um ultimately we have to win our elections by going to the people. And that's not to say that people are always following what they they in their districts want, but I think in people's head there is more prevalence of the worst examples. But it's because that's what's covered in the news. Like the whole constituency wanted clean water and the politician
voted for there to be clean water. Doesn't make the 6 p.m. news. But the whole and that's what happens nine times out of 10. But the whole community wanted clean water and the politician voted to pollute it with PAS. That makes the 6 p.m. news as it should. But that doesn't mean that's what happens at that same prevalence in in legislator. Let's do short questions and short answers just so we can get Yeah, short answers is a really good tip. That's uh sorry about that. Yeah.
So uh in New York in your district, we have a lot of IP quality like actors publishing, you know, that kind of stuff. So are any of your bills basically addressing uh AI? If I if I Google Harry Potter with Daniel Radcliffe, right? And my daughter's an actor and starting out in New York. Uh so basically like what what can you do in terms of really protecting IP? Do you believe that LM generated content has any IP protection at all? That what generated content? So, if I use an LLM to generate content, do you think there's actual IP, intellectual property there, or is it not? He got it. He got it.
Um, that last part is a challenging legal question, and I'm not a lawyer, so I'll pass on that one. On the um what we can do to protect it, my bills aren't directly addressing it, but that's because many of my colleagues are. And so, we've been able to pass some things recently and more that we need to do. So, one of the bills that passed last year was ensuring that in the state of New York, if you are a performer, you cannot sign away the rights, your digital rights to yourself unless that contract has been reviewed by your collective bargaining unit, by union, or by your lawyer. So, they can't force any like starting actor like, "Oh, by the way, we own your image forever." Um, including video games. That's a debate right now. I think video games are
included, but I will double check that. I I I will um So, yeah, that's one of the things we've done. Sorry. They're on strike because of that right now. There you go. There you go. You're going to have to repeat what people show. Uh she said they're on strike because of that. And and and it's a follow on from the strike that the director's guild and the writer guild and and the Screen Actors Guild had last summer to get those protections in their union contracts. Um but even people who are not union should have that protection too, which is why that bill's survey. Yes, absolutely.
They are on strike right now for that. Let's let's try Can I include someone like shouting from the back? We see if that works. Question, guys. Yeah. Go. Nuclear fision. You made the analogy earlier. Yeah. How do we avoid those mistakes? For example, right now we haven't built one in 50 years and China has three times our power output. How do we not that up with the uh we don't have Three Mile Island. Uh the question was, hey, nuclear power, we've messed it up. We haven't built one in 50 years. How do we make sure we do that with AI? we don't have Three-Mile Island, which I'm being like a little flippant, but but but what I mean is what I mean is a lot of that what I would describe as overregulation came in
response to a disaster. And so I think one of the things Okay, it wasn't a big deal and something's always going to go wrong, right? Like No, no, no, no, no, no. Sorry. I I just to be clear, I'm saying there were political actions taken because of it, right? And when you have a major disaster, the natural reaction of a government body is to clamp down in ways that I think maybe shifted risk. So I actually think making sure that nothing horrific happens with AI is the most pro- innovation stance because the alternative on the other side is one that really cramps down and harms a lot of that innovation. Yeah.
Question over here. Yeah. I just want to clarify something before I ask my question. So you had a safety plan. Yeah. And then the safety plan reviewed by third party actors. Yeah. The safety plan has to be developed by the company that they're regulating as the frontier model. Um, and then do they get to select a third party actor who reviews it? Yes. Isn't that a bit of a conflict of interest?
Yes. Why is this in the bill with this? Sure. Uh, isn't it a conflict of interest that the third party actor that they get to like third? Yes. And that and that they make decisions themselves. Yes. 100%. Um and that is why I find it so crazy that people are saying, "Oh, this is stringent. We could never comply with it." Um most developers have been very responsible. I expect most of them to be very responsible when the right guard rails are set up. If we pass this bill, if some, you know, uh, set of unscrupulous auditors pop up or we see some large developers really taking advantage of it, say legislature, we can
take other action in the future. But I I I there's a risk of overprescribing that in a way that actually makes it impossible to deal with. Um, and I wanted to make sure I didn't go on that side. But yes, I am worried about that. I I my ideal one would probably be stronger than it is. True. There's only like 10 questions, so let's answer them quickly. There's a lot. Do you go ahead? Yeah. Why Why didn't you um like what would the provision be to make it a bit stronger?
Uh more set standards on what the audit needs to be or some other body that validates them or some set qualities to be an auditor. We didn't really put any of that. Yeah. Um, do you do you want to just call on like two or three people and then I'll do both of them in a row. You can go all three of you and then you'll answer.
Okay. So, if you can shout your questions and I'll try to repeat it and I'll do two or three in a row. So, so shorter questions will be really helpful for my memory. Regarding Halunteer, a company that's come under scrutiny for the export of technology potentially utilizing AI to make decisions that lead to the deaths of people overseas. Have you considered legislation to limit the harm to the or unusable use of unreasonable uses of that technology?
Okay. Thank you. Yeah, go. I'll answer. Yeah, go ahead. Do you bring the AI revolution in New York? You go to ASAP. It looks like the future in Europe. We're pulling startups left and right. We need to do something to enable startups. I'm a startup founder myself. I'm a constituent. Awesome. Really hard to build stuff here. Yeah. Cool. Yeah. Go ahead. Uh how do you see the breakdown of support versus opposition to your your bill particularly like with the California bill that had some sort of strange bed pillows that were there?
So curious what your your mapping of that is. Cool. Okay. I'll try to take them in order. Um uh I don't have any direct legislation on impacts of Palunteer um or things like it. I left Palunteer part and this is part of the public record partially in in in I was concerned with the direction it was going and this was back in 2019. Um sadly there's federal preeemption for a lot of it. So anything that's worked directly with federal government we can't really touch at the state level but I'm open to ideas about how to keep all of us safe. Um how do we bring the AI revolution to New York? I don't know that SF is a panacea or utopia. I would start there. But but but I take but but
we should set our standards higher than SF. So I I take the point well. Um I worked in two startups before I ran for office. Uh my wife's a YC founder. Uh and so I she's way cooler than me. Um and so yes, I think startups are super important. Small businesses are most of the jobs and most of the job growth in New York. I want that technology coming here. I want that. I want people working here. There's a lot of different things we have to do. Part of it are specific investments we can do in AI. I talked about making it easier to build data centers to get transmission to get you hooked into fiber. Part of it is investments like Cornell Tech and Empire AI where we try to create these centers
that can bring people together that you can have those conversations that then spur the next idea from it and continue to put more support into it. One of the things that New York is doing in its budget, and I believe it's the first in any other state is empowering the government to actually make venture capital investments in AI startups. So, we're like directly putting funding back in to encourage growth. Um, but also startups, AI startups like any other business, there's a lot of regulatory red tape and there's pieces that we could work on to make it easier to found a business to keep that open uh that I think legislature really needs to vote.
So, it's quick answer there and sorry, I forgot the third question. and help me. Uh, how what's your mapping? Oh, strange bed fellows in 1047. Yes. Um, similarly strange here. Um, I have the large companies that it would actually apply to telling me this would end the world. Um, uh, which is funny because that's what the bill is meant to avoid. Um, I have small companies telling me that it's regulatory capture and it's just benefiting the large companies. And again, I tell them to talk to the large companies. But we actually have a number of memos of support from New York startups. Like people that actually read the bill and look at it have found that like oh hey this actually makes a lot of the technology we depend on safer. So we
want that. We have support from venture capitalists. I was on Bradley Tus podcast which is coming out in a few days who strongly support as well. So most of the people that have like looked at it that represents fallins actually support it. And so that's maybe been an interesting change from 1047. Um and more and more people will pop up. So, yes, it'll be it'll be odd. Yeah. Okay.
I don't know if Oh, yeah. Go ahead. You and you. Yeah. Go ahead. Sorry. So, speaking of Palunteer, I I just have a political question. When you ran for office, uh did your opponent use the fact you used to work there against you? Um it's like, you know, big deal connection. Uh and if not, do you feel like that was a missed opportunity? Okay, I'll do that one quickly. Uh, they didn't really use it, but that's because I talked to I brought it up to the press the week I launched and I was like, I'm an open book on all this. Let's talk about it. I gave an interview and then people read that and they're like, "Okay, so I don't think it was a missed opportunity. I think it was done well."
All right. Go ahead, Anthony. Hey. Um, so decade of cyber security experience with the federal government. So, awesome. Yeah. So, super interested. But, uh, my first thing is that, uh, thoughts on privacy because it's a hot take. take but you know citizens of this country have had their all their data used for anything right basically we're the ones who are losing out on that so what if anything this do has those assets are even easier to mine and use things for your applications in every single way how can you protect the citizens in a way that they've never been represented I'll do this one quick question I um uh the New York Privacy Act when it was first written however many years ago I think there were two states three states that had a privacy
bill in place now There's We're in the back half. It's over 25. Um that bill is long overdue. We just passed it this week through the Assembly Committee. I don't I think it'll go to the floor. We'll see. I don't think it's going to go in the Senate. I hope I'm wrong, but it is a bill that I am advocating to pass as long. I got it. So, yeah. Cool.
What was the third? Yeah. Go ahead. Raise your hand if you have a question. Yeah. Okay. Uh, so first, how is there a way to prevent the robot apocalypse in a way? Like I'm I'm just I'm just I'm just saying it's like is it possible to have a universal like part of an off switch that would come with these large AI uh models because like we're already seeing that those efforts to shut it down are actually being defeated. And that's the most alarming thing I've seen to date.
Like all the regulation aside, like that's the one thing that I think needs to be put in place in some form. Uh so the question was how do we stop the robot apocalypse? Um and but in particular like having a shutdown switch or something like that and this bill doesn't have a shutdown switch and 1047 did and and partially I think a lot of opposition focused on that. Um I what this bill does is say companies can design their own plan. Some will say yes, we're going to keep a shutdown switch on everything. Some will say no.
And I think over time as there are auditors, as there's others that look at this, we'll get to some consensus on what that should be and what the risk should be. Um I I think the best step that we can take to reduce the risk of something really bad happening is help me pass this bill. Um and that's why I went through, hey, there were eight at the start and now there's four. And honestly, I think this is the one that has the best shot. So, uh, it's probably not 100% the bill you would write because it's not the 100% the bill that anyone here would write because we're all different people, but it is the bill that I think captures the most of this risk and has a real chance of passing. So, um, I don't want to say
scan the QR code, but please uh, get involved because you can really make a difference. Yeah. Final question. Two, you and then you go ahead. All right, Conrad. Um, can you go into a little bit more detail about how you came up with five million or 100 million or like numbers and why you chose to use a monetary number as opposed to an alternative because there is no alternative. Uh, which again somewhat flippantly. So the question was how did I come up with the thresholds?
Um the main one for so the the way we got the knowledge distinction 5 million was deepseek right we didn't want to do a bill that would apply to a lot of the innovators in the US and then when their models is ripped off uh not apply to what's ripping it off so deepse the first version wouldn't qualify because it was trained on 01 which wasn't a frontier model but they spent over 5 million in doing it did it knowledge distillation that's how we came up with that threshold the 10 to the 26 flops and 100 million is based on binding executive order partially and the and 1047 and a number of other places. Um, I've had a lot of conversations with industry as I said at the start. Uh, and they said, "Hey, it's not the perfect definition." I said, "I know. Give me
the perfect definition. Is there a universal capability threshold? I'll plug it into the bill right now. Um, should that threshold be a higher number than it currently is? Let me know when you agree on and I'll put it in right now." And I've made that offer in August and in December and in February and now here we are in May into June. I stand by it. If industry wants to come together on one threshold for what's a frontier model, like let's talk. This is the best that's been published. So, that's what I'm using for now.
Do you hear? Oh, yeah. Go ahead. Last two. Curious to hear your thoughts. Um, or if there's been any discourse at a policy level on the environmental impact of AI, just things like, you know, you reply, "Thank you to chat to BTA uses the same natural resources as like five evaporated lakes." Yeah. Uh, it's a great question. Uh my colleague, Senator Kristen Gonzalez, has been looking at this and I think has introduced a bill on reporting on the environmental impacts of building data centers and of using it in in these ways. Um so it is it is an active part of the conversation. Um I don't know if I explicitly said this at the start, but I don't believe the raise act or even my package of AI bills, I think I have nine
right now, is covering everything that needs to be covered in AI. That'd be like saying I have nine bills on the economy and it covers everything in the economy, right? There's a lot to do and as a legislature we work together as a team on all of those. So that is part of the conversation for sure. Amazing. Well, thank you. Thank you after. Okay. So um thank you all and thank you to uh the saint for uh hosting us here today and I encourage you to uh get a drink, get some food, uh talk to Alex, uh get interested in DI and uh we'll see you next time. Thank you guys.
[Applause]
Yeah, that's right.
[Music] Thank you. I hope you enjoy.
[Music]
Oh, cool. Yeah. [Music] Yeah, probably. [Music] [Music] One [Music] [Music] [Music] prediction is on the way out. [Music] [Music] [Applause] [Music] Of course. [Music] Okay. Oh yeah. [Music]
I'm doing [Music]
Thanks for coming. [Music] [Music] She [Music] said [Music] that for you [Music] I know, but I still feel that everyone is cool. [Music] Hello. I know. [Music]
It's like hard to say. So I like [Music]
[Music] coming to the Are you up to anything after? No. I'm getting go get done. I need a little time here. Yeah. Yeah.