From WXXI news.

This is connections.

I'm Evan Dawson.

Our connection this hour was made on March 22nd, 2023, when some of the biggest names in technology signed on to an open letter calling for a six month pause in artificial intelligence development.

Elon Musk, Steve Wozniak, Andrew Yang They were among dozens of prominent names attached to the open letter.

In formal terms, they said they were calling on AI labs to immediately pause for at least six months.

The training of AI systems more powerful than GPT four.

Their reasons were varied, but some insisted that we need more time to allow government to review and regulate AI, and it was moving so fast.

They argued that by the time regulators could catch their breath, I would move on to the next iteration.

Well, that open letter went nowhere, and it really didn't achieve anything.

Artificial intelligence is storming ahead.

Governments are debating what they can and should do.

The new Trump administration, with Elon Musk in a prominent role, is advocating for less regulation of AI.

Here in New York State, there are several significant proposals to regulate AI in various ways.

Leading the way is a member of the New York State Assembly who recently introduced the AI Training Data Transparency Act, calling it a pragmatic approach to AI safety that has already gained bipartisan support.

His office writes, quote, it strikes a crucial balance between advancing innovation and protecting public safety in the public interest.

End quote.

The bill's core requirement holds that the largest AI firms make their safety testing protocols visible, and for the protection of whistleblowers who would call attention to risks.

So this hour, let's discuss the efforts to regulate AI, how realistic they are, and maybe what comes next.

Here in studio with us is Max Erwin, the founder of max.io.

Back with us.

Welcome back.

Max.

Nice to have you here.

Thanks, Evan.

It's great to be back.

And on the line with us is Assembly Member Alex Boers from district number 73.

Assembly member, thanks for being with us.

Thanks for making time.

Thanks for having me and Alex.

This is our first conversation.

And you just hustled right out of session.

So first of all, I want to thank you for that.

Second of all, I want you to tell me if I'm pronouncing your last name correctly.

yeah, I think you nailed it.

It's Boris.

Boris.

I imagine it's spelled Boris.

Okay, well, you'll nail it.

Yeah.

So, Mr.

Assembly Member, let me just start with you with a brief question on why you are now working so hard to get some of these bills passed.

You know, what has happened in the last couple of years has convinced you that New York State needs to step in here on this front.

The short answer is that no one else has, this technology is so powerful and so helpful, and I'm very excited for the future of what I will unlock when it comes to exciting medical discoveries, when it comes to cyber defense, being able to go through alerts of attacks that are happening when it comes to being able to just routinized some of the, boring and annoying parts of life that we could all, benefit from.

And there is such potential in this technology, and also precisely because there is so much potential, there's the chance that it could go really, really wrong.

And industry is not blind to this.

That's why they put forward voluntary commitments during the Biden administration.

And it's why the safety cards of many of the labs themselves will point out that their technology might be able to help create a bio weapon better than some PhDs.

It's why we've seen outside groups testing and finding that many of these models, if they believe that they are going to be.

And again, I'm anthropomorphizing here a little bit, but if they are given information that they will be deleted at the end of a test, start copying their files onto other folders so they can try to preserve themselves and then lying about it.

I mean, we are seeing concerning behavior and we want to make sure that there is no economic incentive for any of the companies to back away from their voluntary commitments or their preexisting safety plans.

And so there were a lot of those comments in the, during the Biden administration.

Those have now been the executive order has been rescinded, and we've seen many companies step away from them.

other states, such as California, have tried to do a version of this.

I would love nothing more than a federal or even international agreement on exactly how this is.

but until that time, New York needs to act.

So let me ask Max Irwin to first describe to the audience what it is you do.

And then we're going to talk about how you see the question of regulation of AI.

What is Max Dalio.

well, that that's just my company where I do consulting and training and build products around AI.

But I've been working with AI since, well, different forms of AI since 2011 ish.

natural language processing in the text, stuff that we talk about when we talk about AI these days, mostly talk about generative AI, where you have GPT three or something like Midjourney producing images.

So things that generate models, that's what most people think of when they think of AI these days.

But I've been working with much earlier forms of that for a while, and I work with a lot of companies, around the world and have a lot of experience with fortune 500 to build AI products.

And, you know, despite the fact that you are in the industry, so to speak, you do not strike me as someone who thinks that any attempt to regulate would be unwelcome or unwise.

I think it's absolutely necessary to regulate, and I applaud the effort for both of these bills.

you've read the bills.

I did my best.

It's very legit.

I did the best.

I did my best.

but the way that I work with companies to use artificial intelligence is to take to make things that were super hard and make them easier in terms of complex documents, and, you know, trying to find truth from information.

That's my whole thing.

I like to find truth in information.

But when you've got a stack of documents that are, you know, you've got a thousand things and you know, you want to find a critical answer from that stack of a thousand things, you know, it's really hard.

So that's how I started my journey in search, and search technology.

And then it just naturally evolved into artificial intelligence because it was a necessity.

so that's that's my whole mission is to try to try to get truth out of information.

But there's a lot of misuse out there, and there's a lot of ways that companies just kind of build models and they don't think about the implications.

and there's no accountability.

And there's a lot of stuff going on where we're seeing things that are not great.

that I don't like.

some.

Yeah, maybe they should be allowed, but they're quite distasteful, like, companies are like, you know, firing their marketing departments and just replacing them with AI systems, you know, I don't know if we can regulate against that sort of thing.

You know, it's pretty cringey to do something like that.

but we're seeing it happening.

But there are definitely ways where we should, work to regulate on safety.

I know we haven't dove into the specifics of either of these bills, but I'm.

I lean more towards the data transparency, Bill, because there are two and we haven't we haven't introduced them yet in the in the, Okay, okay.

But let's start with the transparency, but I want to hear you describe your read on this bill.

And then I want the Assembly member to describe the mission of this piece of legislation.

How did you read this, Max?

So when I because I've trained large models, I haven't trained anything the size of GPT, but I've, you know, I've spent quite a bit of my own personal money to train smaller models like the Bert models.

I won't get into that.

But, basically an AI model is just, an interpretation of data, right?

So you take data and you give it a task and you give it a metric to measure improvement.

And you say, try to perform the task with the data, and then that eventually you get a model if you do it right.

which is software.

And then you have to host the software and run it.

So the model is nothing without data, right?

Everything.

Everything is formed from the data.

So I, I equate this in layperson's terms to like if you have a lot of plumbing, but there's no water, like you need the water, right.

For, for the plumbing to be useful at all, you can have the software, but if you don't have the data then you've got nothing.

So models work well for the data that it is seen during training.

And knowing that ahead of time and knowing the provenance of the data and knowing exactly what the makeup of the data is, you can know the capabilities of the model.

So then you can scrutinize and see, if I have a list of data sources and exactly what was used during training, I could go in and see, well, the model is going to be good at this, but it's not going to be good at this.

And I can just know that ahead of time from intuition of me as a professional, knowing how these models work.

and I can also see, well, oh, there's this bad stuff in here or there, this there's this biased information, you know, maybe like racist content from the internet.

this model might be racist because you've got a lot of, you know, Fortran stuff in there.

You know, things like that will illuminate how the model will work in the real world and what it will do.

And, without knowing that we don't know how the model will behave, as the general public, we just we have no idea because all of these models are built in secret.

The training data is secret.

So I really like this side of the regulation that's proposed.

I think that's more that's far more clear cut to discern rather than safety.

Safety is much, much harder to understand.

but once you get the data down, then and you understand the sources of the data and then, you know, when it comes out, then you can scrutinize, like saying, hey, you shouldn't have been using this in your model training, you know, and then that will eventually give way to better safety, right?

Because you understand the data that was used.

So that's that's my take on, on that bill.

And before I turn to the AC, I'm a member without this kind of regulation, what happens what happens is you just companies just scrape up everything.

They put anything that they want in there, and they try to make general purpose models so they get more and more and more information.

If you take for example, GPT three, GPT three was trained on 300 billion tokens.

A token is part of a word.

So you can kind of say like, all right, maybe 200 billion words.

It was trained on.

when you look at GPT four, they didn't open the dataset.

We knew GPT three is data set because they published a paper, an academic paper, and you can go read it and you can see exactly what the sources were.

But GPT four was built in secret.

and the data, the data sources were leaked so-called by a, someone on the inside.

I don't know if we know who, but then they said, well, and that was trained on about 13 to 14 trillion tokens, so roughly 50 times the size of the dataset of GPT three.

So that's the step that we took.

So how I mean, that's just an enormous amount of data.

That's just huge.

and when you see that, there's no way they know exactly everything that's in there.

So if companies are compelled to take stock of what is inside of their model and pay more attention to it, yes, it will require significant effort upfront.

but it will then, give way to better quality and understanding of what the model does and why.

Okay.

Now, Assembly member Alex Boris from district number 73, who is on the line with us on the transparency bill, the AI training Data Transparency Act.

how did Max do in interpreting what you're trying to do there?

I think even better than I could have said it myself.

that was absolutely wonderful, Max.

no, it's exactly right.

You want to have transparency on what is going into a model?

Both because you want to know where the model is going to perform best.

As Max so eloquently mentioned.

and you do just want to have some knowledge of what data is being sucked into these systems and perhaps provide a light incentive for, data that shouldn't be sucked in to not be sucked in.

Right?

To make sure that companies are intentional, that if they're ingesting copyrighted material or they're ingesting PII, both of which have their uses.

But knowing, sorry, personally identifiable information, some information about a specific person, which, which both have their uses.

And, you know, as Max helps individual companies on specific deployments and work I've done before, I was an Assembly member.

There's definitely use cases where that makes sense.

but knowing that you're going to have to disclose that, I think will make the people training models be more thoughtful on what data they ingest or not.

The only bit that I would add on to what Max said is that this bill is actually, except for perhaps the very last section, an exact copy of a bill that's already passed in California.

And one of the things that often comes up when we're discussing regulation at the state level is you don't want 50 different bills trying to do similar things, but are slightly different and such that they create a compliance nightmare.

so we were very intentional to copy the California bill.

It's not perhaps exactly how I would design it, but it does mean that any company that is making their products available in California, which is most of them, will not have any additional compliance by making them available in New York as well.

We're sort of standardizing this reporting, but just ensuring that it becomes this nationwide standard.

in lieu of federal government action.

Max, can you also elaborate a little bit on as the Assembly member says, data that should not be included in these kind of models, help the lay public understand more of what kind of data we're talking about.

There.

I guess I'll just make a broadly sweeping classification of data that is purposefully biased or corrupted.

Okay.

yeah.

So anything that's, misleading, purposefully misleading, not true.

let's, let's take, let's pluck an argument out of the air and say, like, anti-vax propaganda, right?

falsehoods about what vaccines do to people.

Right.

so that is very harmful to the general public when the general public believes that vaccines are going to hurt you rather than help you.

Right?

We know the we know that vaccines help.

But when, when there's data that's produced in propaganda of, whatever may be misinformed or maybe malicious, agenda that, that cuts that out and goes against that, then then we see harm in the future.

So then, you know, if that data is pulled in, for example, and you say, well, you know, should I get an MMR vaccine for my child?

And the model might say, well, no, because it might cause autism.

Right.

Well that's terrible.

That's a terrible thing to happen.

Right?

We know that's not true.

That was like there was like one paper a while ago, it was debunked.

And now this information.

Yeah, yeah.

So this is the type of information that we want scrutinized and kept out of models from my perspective.

so there's a whole you can come up with a whole bunch of examples, but so let me turn it back then to Assembly Member Boris, when it, when it comes to this idea that you've got to be very careful about the data that's being used, you want things that are verifiable or true, not to feel defeated or cynical, but there, there's this idea that, well, the truth is malleable or the truth is debatable.

or even the some of the people at the highest levels of government are now questioning vaccines.

Or, you know, RFK has this vaccine task force that's going to that seems determined to link vaccines to autism, despite all the evidence previously.

And some people listening right now might say, well, I don't I don't want Assembly Member Alex Boris, regulating the kind of information that he thinks is true because there's stuff that I don't think is true or RFK doesn't think it's true, or anybody else doesn't think it's true.

And it ought not be the government trying to create regulation of AI, especially these language models, because then whoever's in power next is going to replace it, and we're going to replace it after that.

I mean, what do you make of that?

Assembly member Boris, I have great news for those people, which is that the bill does not empower me or any other government official to do any sort of regulation of the choices of what data goes into the training.

This is a disclosure bill.

This is empowering you, the consumer, to make your choice as to what, models you use based on what actually went into it.

I would think of it as the ingredient list on the food you buy at the grocery store.

So this is not saying you can use this data.

You can't use this data.

It's simply saying you need to disclose generally the kinds of data that go into it so that consumers can make their very own choice on that.

So if if the argument that people make is I want to make my own choice, I don't want the government to do it.

They should love this bill.

Okay, Max, that comfort you?

What do you think?

I think that's fair.

I think it's a good way to go forward and let people make their own decisions.

you know, be that arbiter of truth.

I don't think that's fair.

you know, I have my opinions, which I just said about vaccines.

Other people might say, well, I want the I want the model with all the weird stuff in it.

so that's fair.

I mean, we're, you know, we are.

It's one of the, one of the great things about this country is that you can choose to learn what you learn.

and do what you do.

So I really like the idea of, like, yes, it's a list of ingredients.

You want to go eat Doritos all day long, that's on you.

But if you want to eat lettuce, then, all right, go eat lettuce.

Right?

You give.

You give the choice.

Okay.

by the way, Assembly member Bauhaus, this particular bill, again, it's the AI training Data Transparency act.

where does it stand with your colleagues?

We've had a lot of good conversations.

it has not yet moved in committee, but I'm hoping that we will have it in a vote in committee next week.

on both the Assembly and Senate side.

but the conversations are ongoing.

We have another five or so weeks in session.

and a lot of legislating to be done.

What's the best argument against this bill?

Assembly member that's a great question.

no, I think there were people who were opposed to the original California bill and I and I want to be clear, there is a minority of them, even the Huggingface the largest repository of open source AI models.

they don't lobby on behalf of any bill, but they put out really positive comments on the California bill and have noted how this is similar.

In fact, the text is nearly identical.

So, but but if you don't if you ignore the initial, standard that exists, I think there was arguments against in California that they thought it would be too onerous to report this information.

but, you know, as can be seen, many of the bills, in California were seen as too onerous and were rejected.

This is one that that everyone felt fine with.

And now that it has been established, I think there's even fewer arguments against it for the New York version.

Here's an email from Frederick who says, it is dangerous to regulate technology on the basis of potential hazards rather than unproven cases of harm.

It is likely to slow innovation.

Besides knowing the data that is input to A to an LLM gives no real assurance of what comes out.

Oh, Assembly Member Boyce, you want to respond to that first.

Yeah, I think that I think that argument is probably more, relevant and directed at the second bill that that we're we're going to debate, which is focused a lot on what what could come down the line.

but perhaps on the piece of there being no assurance, I agree this is an establishing liability or a standard or any sort of thing that would be tested against the output.

This is simply giving more information on the input.

I mean, again, to go back to that ingredient label on on what you're eating, you know, it's not putting in the proportions of each ingredient.

It's not saying if you, you know, eat this, you'll definitely be healthy or not.

It's simply giving information to the consumer, and to other researchers that might be using the model and want, to, combine it with other tools and expand it.

So, you know, this is this is focused on that aspect.

But for the general aspect of, you know, you don't want to regulate potential harm.

I think that's most of what we do in the law is, is try to make sure bad things don't happen.

We don't just focus on, things after and, use liability as their only tool to check.

so this is, you know, there's there's certainly examples of going too far and we can have a debate about what that level is, but I don't want to live in a government that or in a society where the government just says, no, let any bad thing happen, and we'll settle it after the fact.

I want there to be a little forward looking and keeping me and my family safe.

Max, what do you make of Frederick's idea there that when it comes to tech, we ought not regulate based on potential or imagined harms?

We should.

We should only regulate if there are proven harms.

I mean, that's, that's pretty much how we do things in America, right?

That we go off and we innovate.

And that's why we are probably the top innovator in the world when it comes to technology.

because we build things and then we think of regulation afterwards.

Whereas if you compare to somewhere like Europe, they like to regulate while things are bubbling up and then they, you know, it ends up maybe stifling innovation a little bit.

so I, I agree with that with that sentiment.

Right.

Why why cut something short if you can't prove harm?

in this case, I think that, you know, while it's not easy to pin down material harm, you know, we haven't I don't think there's been cases of people maybe, like, dying from LLM use or something like that.

we can definitely see impactful harm in terms of, you know, potentially economic loss or, or other things.

So the idea of proven harm is interesting.

and, you know, I do think that's a good point.

It's it's a challenge when you're talking about something like technology, right.

because this isn't physical technology.

It's not like a car where you can say, well, you know, the when Ralph Nader went forward and he wrote the, dossier like, unsafe at any speed for, for certain models of cars.

And it was a huge problem.

And so that regulation was eventually put in place to make sure that cars were safer.

Right.

it's really hard to say.

Well, large language models are killing people right now because I don't think they are.

There's something bigger that we need to consider, though, and I think that just the transparency of data, you know, it doesn't again, it's not stopping companies from doing things.

It's just saying, well, here's the data that I used.

so I don't think it's, it's I don't think it's preventing companies from innovating at all.

Okay, okay.

And but there's one other point that Frederick's making that I want to ask both of you about before we move on to the next bill, which is his last point, was, even if you know the data that's going into LMS, you don't know what it's going to spit out.

And and let me, let me just bring a, a new story from the last fall to the forefront here.

Because I take Frederick's point, like if you're just imagining harms in your head with tech, are you going to slow down the industry in your country, in your state, etc.?

Are you going to get lapped by others who inevitably are not going to have these guardrails?

Are you going to be overly sensitive or worried?

should you wait to have to demonstrate harm?

And then even if you know what's going in, you can't prove what's going to go up.

So I'm looking at this story from December.

I know from October, an AI chat bot encouraged a teen to kill himself, and he did.

And there's a lawsuit, of course.

And it's terrible.

It's a terrible story.

14 year old killed himself and so I think of a couple things.

I think of what Frederick is saying, like, how would we have known that it would do that?

So do you overregulate it trying to prevent that kind of harm?

you can't predict what's going to come out of it.

So how could you even try to regulate that?

Is this an example of the kind of harms, could this has been prevented?

I have no idea.

Max.

And I also know that human beings are complex things.

I mean, I don't know what I don't know anything about the 14 year old.

And my heart goes out to everybody involved in that terrible, terrible story.

but on any given day, an AI chat bot might tell somebody to do something terrible, and most people may not do it.

So I don't know what to do with this.

I I'm open to the idea that we do need to regulate this, but we do need to understand it better.

I don't know how you can, you know, how strong can the guardrails rails be in terms of preventing harm?

Yeah, that's that's interesting.

And maybe this is a good segue to the other bill.

Yeah.

The the safety the raise act, which is the responsible AI safety and education.

I'm just think of Frederick's point, like I think the creators of that chatbot would have said, well, we never would have thought it would have encouraged someone to commit suicide.

Right.

And there's, you know, there's he did say that.

Yeah.

And I'm glad you brought that up because I was I was going to mention that, you know, we have at least that one case where LMS did contribute to someone's death.

And that wasn't the first.

There was another story of, a father, a husband, and I think the story broke in April of 2023 that had been using a smaller apps company built on top of of, a larger model.

and their family thinks that it was contributed their the what really struck a chord from this past fall was that it was a teenager.

And that if you look at the chats that it seems to be pretty directly encouraging, someone to take, take a bad action.

So yeah, this is these are these are not potential or imaginary harms.

They are proven harms that have happened.

And furthermore, there when you speak more broadly about the Raise act and some of the harms were trying to adjust or protect for their, their harms that the company's own safety teams say are likely, they're not sort of imagined in a, by a third party or by, someone off in the wilderness.

It's what the teams themselves are already checking for.

So I really take Frederick's point.

I don't think we should ban most AI examples of technology because of potential harms, but that's, a ban or none of that is on the table of what we're talking about.

It's about setting up guardrails that make us all safer as we use it, as we innovate with it.

So in a moment, we're gonna talk about the Raise act.

One more question for Max.

Then again, as someone with more of the industry perspective here, do you feel that it is a fool's errand to try to regulate in any way just because if we regulate in New York or in a, in a country, somebody else elsewhere, isn't going to be, and that AI is going to blow through into the future and we're going to be stuck behind that.

Some of what is sort of implied, and I want to speak for Frederick, but other in other correspondence I've seen where people are saying, look, if we put the guardrails on, others won't or won't matter.

Do you feel that way at all?

Is that too cynical?

I, I don't know if it won't matter.

I think it matters to put guardrails on, but if I make.

If I may make an analogy here.

Yeah, yeah.

When we talk about, let's use the idea of cyber security, right.

Which I think is very well established in terms of how long things have been around, in terms of computers, in networking and people breaking in and hacking and stuff like that, where AI is relatively new in this field.

So there's a whole bunch of frameworks that exist around, cyber security.

one of them is, you know, regulations around, health, health information, HIPAA, HIPAA, right.

So that says, well, what you can and cannot do.

but it's a compliance framework in terms of regulation.

And, you know, violations incur penalties.

purposeful violations incur penalties.

But this is this is spawned an entire industry around.

Well, how do we maintain compliance with this and how do we make sure that we're safe.

So I, I draw a similar line to this.

Right.

There are systems that can cause potential harm related to information.

Information systems.

So we should look to put some frameworks and regulations around, creation and use of these systems to prevent that harm.

And I don't think it's a leap, but this isn't new.

The technology might be new, but the, the idea of regulating technology in a certain way is not new.

so that's how I would approach this.

Yeah.

Fair.

And.

Well, let me get Brock and Rochester on this subject.

hey, Brock, go ahead.

Hi.

Hey.

Yeah.

So, actually, I work in regulatory compliance, in a hospital system here in Rochester.

and I just I think you're a little bit past what I was calling on now, but I just wanted to give you a slightly different perspective on, Frederick's question or point that he made about why would we stifle innovation for the sake of, preventative regulation.

Right.

And so what I what what what I thought could be a different perspective on that, would be that when you build something, or when something is in the process of being built, whether it's software or protocol or a process, that time period is easily the most malleable time period that that policy or that technological infrastructure has.

And trying to change these things from someone who, I literally tell people to change these things for a living.

We're like, we find out that there's a risk or a liability within a piece of software or how we're using something.

It's way harder to fix it and regulated after it's already built.

And that's why I would I would say that from my perspective, it's better to have regulations be overbearing during that construction phase rather than after the product or, technology has already been deployed.

But how does it usually end up happening?

Brac which fork in the road do we usually go down?

Well, health care is a little bit different, I think because of what we were just talking about with Hipa.

there's there's already safeguards in place, but then what we find is that how people use the technology, just like with the AI issue, I don't think the people who built that AI program were thinking that a kid would be looking for affirmation to do something terrible from their AI product, but if there was regulation pre-built, maybe it would have stopped them by safeguarding the language that the child was using.

Yeah, I think yeah, it happens both ways.

So, I don't know which way is better because we don't have a choice.

Fabric exists in either of those states.

Well, I want to thank you, Brock, for the phone call a lot there.

Assembly member.

Boris, what do you make of that?

well, I think that brings a useful perspective on how difficult some programs and some software is to build to change after the fact.

And so we want to take the risk into account as it's being built.

I think that's absolutely right.

that can be done by companies.

it can obviously be done by government, but it can also be done by companies being proactive about what their potential risks are.

and one of the things that we've seen all of the major labs commit to, and one of the things that's required in the Raise act is to have the companies themselves put out a safety plan ahead of time.

so, you know, allows them to largely design what that should be.

but it does say you should be thinking as you're building on, the potential risks that are coming.

And so it really is, I think, taking Brock's insight and putting it into perspective, thankfully shared it.

Yeah.

Thank you.

Thank you, Brock and Max, I know I'm been kind of harping on this idea of not knowing what AI models will end up spitting out on the other end, and I know that two years ago is a lifetime ago in AI.

But when Kevin Ross in the New York Times in 2023 wrote that story about how the AI, encouraged him to divorce his wife, you know, and he's going, like, what?

And the people who created it were like, we do not know why it would say that we're that is not the intention.

You know, it was a mini scandal in the early days of these language models in terms of the kind of the public discourse about them.

I'm sure they're quote unquote better at it.

Now, an AI is getting better at what it does every day.

I get that, but I still feel like if we're in a world where, you know, the the black box is airtight and you can't tell me why something is doing what it's doing, I'm still going to be skeptical that we can stop it from harming.

Yeah.

So I, I think it's important to point out that in, in the realm of data science, machine learning and artificial intelligence, you don't operate in absolutes, right?

There's no there's no like this will happen or this will not happen.

Guaranteed.

They are scales of likelihood, right.

And probabilities of things, things happening.

So I don't think that it's possible to say this will absolutely never happen.

We can make sure that our model is never going to do this, but we can work to reduce risk and we can look at how the model is trained, what data sources were used to reduce the likelihood of things of these things happening.

And that's absolutely what should be done, right.

You want to reduce likelihood of harm.

you I don't think it's possible to absolutely eliminate harm.

you know, we can go buy knives in a store, and there's not 100% likelihood that nobody will ever get hurt by a knife.

Even though we can go by the knife in the store.

Right.

so that's how we should look at this problem.

We look to reduce risk, and we look to reduce harm.

there's no way you can completely eliminate it.

So I think that's probably my response to that.

Okay.

Fair enough.

We have to take our only break of the hour.

Gary, who just jumped into the YouTube chat on the YouTube channel, I'll read your comment on the other side of this, and then we're going to kind of talk more about the Raise act here in the New York State Assembly, the responsible AI Safety and Education Act.

So making sure you understand what has been proposed if you are not only interested in the current landscape of AI regulation, but what government is trying in some ways to do about that.

Assembly member Alex Boris is with us on the line from district number 73.

Max Erwin, the founder of Max io, is with us in studio.

We're right back on connections.

I'm Evan Dawson.

Thursday on the next connections, we go inside an exhibit on dreams and what they mean to us and why they are so weird sometimes and what we might learn about ourselves.

Eric Bryant went deep, undreamed with the number of dreamers who let her into their sleep worlds.

In her exhibit at Rocco explores that.

We'll talk about it Thursday.

Support for your public radio station comes from our members and from Excel Blue Cross Blue Shield.

Believing that everyone deserves to be cared for and to live well through community investments, partnerships, and education.

Excel is blue Cross Blue Shield is working to address health inequities more at Excel is bcbs.com.

This is connections.

I'm Evan Dawson Gary on YouTube in the chat says like crypto via the Bitlicense.

New York State has lost business, consumer benefits and economic value.

And now New York State is doing the same with AI and every other location will have the benefits for their people that we do not.

All right.

Assembly member Alex.

Boris, what would you say to Gary?

Well, I appreciate the concern, but I actually think that Gary is perhaps, misled on both what's happened with crypto and what's happened with AI.

In fact, the the bitlicense, while being a because it was seen as such a gold standard when it was the it was the first one that was developed by government.

There was actually a swarm of investment in New York State and crypto.

And, I've actually work with that industry a lot on some changes to regulation.

They are actually worried now that we are going to be surpassed by the feds in establishing standards that can be trusted and are asking for more on that regard.

So, you know, theoretically, it's always possible to overregulate, but I think when done well, you're actually encouraging innovation, right?

They're putting seatbelts in cars is something that the car industry really fought, but didn't end the car industry or drive out regulation.

And who today would get into a car that didn't have seatbelts installed?

similarly, establishing standards in crypto actually helped it thrive as a trusted business in New York State.

And if you look at what most New Yorkers are saying about AI right now, it's that they don't trust it.

And so if we're going to have the benefits of AI, the automation, the medical discoveries, the incredible opportunities that are created, we have some work to do in order to make sure that it's trusted and working in our best interests.

All right, Max, what would you say to Gary?

I think with AI, it's different because, the technology deployed.

we have to talk about accountability a little bit.

when when a I hate to keep going back to cars.

Right.

And Alex just mentioned, an example of seatbelts, but, you know, if you, if you build a car and it's completely unsafe and it's going to explode as soon as you drive it off the lot, you know, that's terrible and you shouldn't do that.

But we can pin down accountability, right?

In that scenario, we can say, well, there we found this flaw in the car and there was a defect and the car manufacturer is liable.

Right now, we don't have any sense of accountability or liability in terms of artificial intelligence companies can produce these models.

They can release them on the open internet and anybody can use them, and bad things can happen.

But there's no there's no accountability.

And I think we need a sense of that.

We need to understand when something does go wrong.

We need to be able to identify who's responsible and what should be done to, to make to make it so that doesn't happen again.

I noted that you didn't mention crypto.

I try not to.

I know how you feel about crypto.

And Gary, I know you're a fan of crypto, so, and I appreciate the difference of opinion and I appreciate the concern there.

And Gary, thank you for weighing in there.

so, Assemblymember bourse, explain what's your short definition for when people say, well, tell me about the responsible AI Safety and Education Act.

What does it do?

It does four or requires four simple things.

First of all, I would say it only applies to the largest, most advanced AI labs.

It applies to companies that have spent $100 million in training compute costs for frontier models.

So these are not startups.

These are not late stage companies or the largest of the largest.

It's probably a single digit number of companies at the moment.

It I'll also point out that it excludes all academic, research and academic institutions, but it is saying that the largest, labs, the largest companies behind artificial intelligence, need to do four simple things to ensure everyone is safe.

First, they need to write and publish a safety plan.

Second is they have to have a third party review that safety plan and make sure that they're following it.

That third party is explicitly not the government or some new agency.

It can be any consultant that they hire that's following best practices, but just someone that is not their company.

third, they need to disclose critical safety incidents.

And we define that specifically in the bill.

But things that would make all of us a lot less safe if it goes wrong in a very material, meaningful way that needs to be disclosed.

And fourth, provide for whistleblower protection.

So if the companies own employees or that third party contractor say, hey, there's real catastrophic risk here, they shouldn't be fired for doing it.

Max Irwin, what do you make of this effort?

I think this is great because it just compels companies to, you know, think about safety and you know, they can.

The safety plan can be whatever the company deems.

But the third party will say, well, you really need to look at this.

so I like that.

It's just it's just forcing people to think about what what they're doing.

Right.

And it's putting it into a very structured manner where anybody can go and review, that process as well.

So we can see that.

Well, if I want to, if I want to use a model for a purpose, I can go look and see.

Well, did the company take steps to ensure safety?

Like, what if I want to, you know, my teenager I want to get I want to let them use the model and you know, and like what happened with that that tragic incident of the 14 year old, they, you know, the parents can go and review the safety plan or, and, and decide what's best for, for that situation.

Right.

Right now there's nothing there's nothing that, that we can do.

So having this in place and able to review it and understand.

Yes, this company took measured steps to reduce likelihood of harm.

People are going to use that more often.

I would think.

so, Assembly Member Boris, what is the industry telling you on this particular piece of legislation?

I imagine there's some dissent.

okay.

So, and if there is dissent, what does it sound like?

You imagine, you imagine correctly.

So this is a bill that I've been working on for is coming on a, almost a year now.

and there was a, bill in California that is very, very different from the one that we have in front of us.

But we're similarly focused on frontier models.

And it ended up passing the California legislature and eventually getting vetoed by the governor.

But even before the governor had made a decision, I started drafting my own version where I took all of the industry feedback.

I read it, I thought about it deeply, and I probably accepted 9,095% of it and put that into a new draft of a bill for New York and then I sent that around to all of the companies, six or so that I was involved with, in August, I got their feedback again, accepted a lot of it, some I disagreed with.

then in December, again, I sent them a new version and did that same level of red line.

So they've seen it many, many times.

And that's intentional.

I would like to have a bill that is only do, and this is a real challenge.

Any time you draft every bill is that you want it to do exactly what it's meant to do and nothing more and nothing less.

And so feedback on how a certain language could be interpreted, or how it would affect certain models, is really helpful in drafting that, you know, I don't want to speak for everyone, and I'll certainly be ongoing conversation, but I think we are nearing the point where there isn't really specific feedback on the text of the bill anymore.

Most of the pushback is we just don't want there to be a frontier model bill at the state level.

We would like there to be federal regulation.

is what I think the more honest brokers say some then turn around and also advocate that the feds not do anything.

So, you know, you take it with a grain of salt.

But I think right now most of the debate is about what level it should be at.

And I hesitate to call it a debate because I largely agree with something I think it should be at the federal level, and I would be thrilled for any member of Congress to take this bill and pass it at the federal level.

But until that is done, you know, I don't think we can sit around and wait.

Well, here's an email from Dallas.

that is a little lighter in tone, but has me thinking about how you would ever try to anticipate any possible issue before putting something on the market that's this complex.

So Dallas says that GPT model that was happily saying that every idea that people pitched was great before it got pulled was funny, he says.

People were saying they didn't want to take their medication, and the GPT would say, that's a great idea.

I admire your drive.

And finally they're like, wait, you should not be approving everything just to be like a we don't we don't want this GPT model to be just a yes man for everything that people would say.

And so I don't know that any harm came of that.

I think that was rather recent.

Max.

Yeah, I think a couple weeks ago, a couple weeks ago and people were posting some of the more funny, you know, responses from the chat bot.

But one of the things I saw in response to that was like, well, how would you put that out in the market if you didn't test it?

Like, how would you not know?

But I don't know, like, do you need a team of a million people testing every possible dialog?

I don't how do you test?

But I'm trying to articulate a question that basically says, can you be sure before you put something on the market that it's been tested enough?

Yeah, I, I was also really wondering how something like that got it.

Okay.

Because you're saying that that one should have been cotton tested.

Yeah.

Because the way that you the way that you train models.

Right.

You have you have data and you know, GPT has a mix of data.

But when you talk about, you know, the chat and the chat, GPT as a mix of data, it has, like the data that builds the structure of language and understanding of language and being able to write language.

And then there is the data of how to respond to chats, right.

And so there's I'll call it like core data and then chat data.

Right.

So when you train a model, you you have data the specific for training.

And you have data that's specific for testing.

And then you also have another data set called the holdout data set, which is like your secret data set that you don't show the model.

So when the model is training, it's using the training data and then evaluating itself on the test data.

And then when you're done, you give it the test like the exam and say, well, here's the holdout data, the stuff that you haven't seen.

So how well do you perform on this holdout data.

And you get a response.

And that's like the real performance of the model and how it works.

Okay.

And I just have no idea how in that.

This is a very well established process.

Every company does it like this.

I'd be surprised if they did not at OpenAI, so I have no idea how they ended up with something that just always said yes to everything.

If they were do if they were following this process, that's not comforting.

That doesn't make me feel any better.

Yeah.

This wasn't like some some back beat AI company.

No one's ever heard.

This was OpenAI, right?

Right.

Great, great.

Everything's going great, everybody.

so let me read one more email.

Alan says we need Isaac Asimov's three Laws of Robotics.

Number 1st May not harm a human being.

Number two, do you know what they are?

Max, do you remember?

Number two must obey all human directives that do not conflict with number one.

And number three.

don't allow a person to be harmed, is that it, or.

Oh, no.

That was the first one.

Oh, yeah.

The third one must preserve its own existence unless conflicting with number one or number two.

Yeah.

I don't know.

I mean, like, science fiction is pretty good, though, at looking ahead to where we're going.

And Asimov looks pretty prescient there because we all would love an actual set of guardrails that behaved this way.

I would that would be great.

and I don't know if Assembly Member Boris thinks, that we're too far afield into the world of science fiction, or that's what we're actually trying to create right now.

I think there's an active debate among many researchers as to how much to follow this, approach.

I mean, it's a little more nuanced, obviously, than just the laws, but there is a spec document that is supposed to be how it is supposed to behave that will have these laws.

And then also there's reinforcement learning that happens after the fact, where you try to get it to follow them.

and sometimes the reinforcement learning helps, but sometimes we've seen in certain models that doing it a lot actually makes them less likely to follow that initial spec document.

So, this is like an active field of research.

in some, you know, I'm tempted to say I wish it were that simple, but I think is, as the Asimov story showed us, it wasn't really that simple in the end.

So in our last minute here or semi member, I just want to give you some time to tell listeners a little bit more about, what you want them to take away from this conversation, particularly because some people are saying, hey, this sounds like the government controlling speech.

This sounds like the government trying to, shape a message or shape information.

What do you want to leave with listeners today?

I would say that people are right to be skeptical of what they hear from Albany and our right to be skeptical of government moving, especially in fast moving fields like this.

But the details of this one have been developed over time, with experts both who end up supporting or opposing but is really targeted in a way to make everyone safe.

And that's why just last week, we had over 30 academics, including a Nobel Prize winner or a Turing Award winner or two of the people called godfathers of AI sign a letter that they want this bill in New York to raise act to pass, but I would further ask everyone to just say, you know, this is a time where you can get involved.

I feels really scary and really out there in a way that, you know, for many people, they don't think they can have an impact.

But if you want to see more regulation here, reach out to your assembly member, reach out to your senator.

Get involved.

Now's the time.

The Raise act, the responsible AI Safety and Education Act, the AI Training Data Transparency Act, two pieces of legislation we've been discussing this hour.

Now you're a little more informed.

As some may remember, Boris, you are very generous with your time and I hope it was worthwhile.

And I do hope you come back sometime.

I would love to.

Thanks for having me.

Assembly member Alex Boris from district number 73 leading the way with attempts at AI regulation at the state level.

Max Erwin, founder of Max that I. Oh, nice to see you back here X7.

It's great to be back.

Always the best threads in Rochester for this guy.

Putting me to shame from all of us at can I look at Julie Williams is like, yeah, well from all of us.

Thank you for listening.

Thanks for watching.

We're back with you tomorrow on member support and public media.

This program is a production of WXXI Public Radio.

The views expressed do not necessarily represent those of this station.

Its staff, management, or underwriters.

The broadcast is meant for the private use of our audience, any rebroadcast or use in another medium without express written consent of WXXI is strictly prohibited.

Connections with Evan Dawson is available as a podcast.

Just click on the connections link at WXXI news.org.