BGC Group, Inc. (BGC) Earnings Call Transcript & Summary

Yes. No. Thanks, Rob, for having me having me. It's really fun to kind of see everybody some familiar faces out there. Yes, just taking a step back, it was great to hear John's remarks and lose remarks earlier. From an LCH perspective and a greater [indiscernible] perspective, I just want to say how delighted we are to be working with FMX. I've been lucky enough to be on this journey for several years now, and when we first kind of embarked on this initiative, one of the things we liked about FMX is that you guys are really disruptors, right? You you come into the market, you disrupt and you lead with technology. And I think when you take a look at some of the advantages, and I know we talked about it in earlier panels, right, the tighter tick sizes, faster execution, smaller block sizes, these are just some of the many things that you guys are doing. And I think if you pair that with LCH and kind of where we're strong at, right, our clearing capabilities, our risk management capabilities. And even before we get into margin efficiencies and futures versus swaps and all that stuff, right, if you take a look at our margin -- stand-alone margin methodologies for SwapClear and our standing margin methodology for listed rates, I think what you'll find is, at least in my opinion, I think we're quite a [indiscernible], right? We have lower [indiscernible], I can never say that word, right, where our margin models are predictable and they're stable in the events -- when there's market events, when there's tough market conditions, what you'll find is lower margin volatility is almost as important in a treasury function as the IM savings. So I think that when you pair LCH's clearing capabilities with FMX's execution, you really do have something special out there. And it's super exciting to see how much success we've had so far and how much success is yet to come. So going back to your kind of original question around size of the pool, right? So I think there's a slide above me here, right, where LCH SwapClear service clears about 96% of world U.S. dollar swap volume by notional. And I think what you'll find is, we've got similar percentages in other major currencies as well. And so I know someone mentioned it in the panel right just before, right, that large pool is there. And what does that mean? It means that you have more opportunities as you bring your futures contracts into the FMX, LCH construct that you can kind of unlock some of those margins savings. I think maybe should I take a second and talk about kind of what the portfolio margining service is and what we do?

Sure, touch on that and also begin to describe the amount of benefit that client trading swaps and SOFR in the client trading swaps and treasury futures could expect to obtain?

Yes. Let me start there, and then I'll kind of get to what the service does, right? So in terms of savings, I think if you have -- it depends on the kind of the correlation between the products. So if you've got a U.S. SOFR treasury -- sorry, U.S. SOFR future and you have a U.S. SOFR swap, you can see savings up to 80%, right, kind of their risk relationship is quite directly correlated. If you've got a U.S. treasury future and a swap -- U.S. dollar swap, you could see savings up to 65%, 70%. And the most interesting thing is if you're taking a look at the U.S. treasury future versus a non-U.S. dollar swap, you can also see savings as well. And I think that's super important because now you're looking at different currencies across our and we've seen savings that are up to 50%.

I want to pause you there at the moment. And I want to offer this to each of the FCM panels individually. How has your ability to clear FMX futures combined with LCH swaps changed the conversations that you're having with your own clients? Take us through that, Sebestian?

Sure. Thank you, Roman. Again, thank you for SP1 Portfolio margining from is one of the first items that customers are speaking to us about across the board. As the market becomes increasingly sophisticated and the cost of capital top of everyone's minds, portfolio margining and the benefits that clients can obtain is always going to be high on that list. Our ability as an FCM to maintain the market leading position is dependent on being able to offer some of these innovative products to our customers. When clients first started hearing about FMX, it would certainly top of the list of things that they wanted to hear about. What was Barclays doing to support this, both on the health side and then also on the client side as well?

[indiscernible].

I think the beauty of the FMX LCH is it sort of standardizes something that is transparent across the marketplace. I mean portfolio margining is nothing new to banks or where the clearing entity through special vehicles would be willing to finance this, but it's -- it varies from [indiscernible] bank to different places. And I think the consistency of the methodology that LCH brings with the FMX products allows clients to use through the technology there a steady set of hands that manage the collateral needed to support the positions. And when it's equal playing field, people can -- they have predictability and transparency. And that leads to liquidity. And when you have capital efficiencies allow more capital to be deployed to provide further liquidity when capital isn't deployed against multiple CCPs. But in this partnership, I think it's really the transparency and consistency of the model that allows people to build. I don't think we've seen the adoption at the level that we sort of had hoped, but I think that builds on the [indiscernible] volume and the rest of it, liquidity brings further liquidity that brings efficiencies and dependencies to the market. But yes, we've been there, we are in that place. We were one of the nonbanks. We have a little bit more -- we were able to do some of the original cross-margining with some of the clients that are here today. And yes, it certainly does depend on the complexity or the composition of the portfolio or how LCH's model. It's not -- I'm not going to underestimate the complexity of the model. It is a complex model, but it's consistent and transparent. So that's what we're building off of as a building block to move forward and get further adoption.

Yes. So Robert, you ask, how has the conversation changed with the clients? I think this is directly in line. FMX is directly in line with conversations we're having with clients about a lot of things around excess cash and margin efficiency, right? So over the last 3 years, we have had a marked increase in terms of the number of ways in which clients particularly hedge funds, RV hedge funds, macro hedge funds, particularly but in that space, are looking at ways they are utilizing their -- or managing their excess cash. And we're seeing everything from increased numbers of swap compressions. We're seeing a significant increase in the number of portfolio transfers sort of ex-pit off exchange in terms of managing exposures from FMC to FCM. And so with FMX, this is just a natural conversation about more margin efficiency. And it fits really well into having those conversations. And I think over the last 3 years, I think for the first time, we've seen CIOs of hedge funds who didn't really pay attention to what the treasury functions we're doing within those funds start to ask really hard questions about like what are we doing with our excess cash? How are we managing this? Where is the cash at? How much margin are we posting? There comes a cost with this now, so how efficient are we being? And so I think that FMX fits really well into all of these conversations and clients seem to be really receptive.

So Jim, I want to stay with you for a moment. So you're having a conversation with the client, and they want to begin to get an idea of what sort of benefit they can achieve through FMX futures and LCH swaps. And what sort of tools, what sort of process, what do you put in front of them to start answering that question for them?

Yes. So we have a number -- so we have multiple lines of defense from a risk perspective in terms of looking at that and running those scenarios, whether they're using LCH SwapClear IM calculator, which we use. But we also have other metrics that we use. We use risk metrics. We use other VAR and stress scenarios. We use other vendor calculators to run these scenarios. But we can run hypothetical scenarios. We can use a realized portfolio. And all of that, we can kind of give some kind of output in terms of what the margin savings are as to what we could see in the portfolio. So we have a number of tools that we use. We're lucky we have a great risk team in the front office as well that can help us. So it's something we have those conversations all the time now.

Anyone else?

I think we've experienced some CCPs that have started to share sort of a margin methodology that is unique to each CCP. And I don't want to understate the challenges of multiple CCPs running margins on the same position at coming up with 2 different numbers and then having to explain to the client base, this is what we had that is sort of the beauty of what we get in this partnership with FMX and LCH. So we're going to support you as well. And like it's 1 of the very positive aspects of this program.

Yes. [indiscernible], on our side, we will work very closely with clients to understand more about their portfolio, understand more about their trading activity. We can give them demonstrations of what this portfolio is going to look like at various clearing houses. And we also look at it within our own internal portfolio margining methodology to be run as both Jim and Gabe mentioned, the ability for clients to understand how their portfolio is going to behave different scenarios certainly becomes key on their list of things to understand as they're looking at how portfolio is being allocated or how they're allocating business across clearing brokers? All of this feeds into that same picture of [indiscernible] FX being another tool in that too.

So [indiscernible], you began by describing how large this benefit is and how much a client can actually consume. The FCMs are now having conversations with their clients. They're running some projections. They're using your IM tool that's online. What from a structure standpoint, does a client need to do with their FCM in order to extract the benefit from LCH? Talk a little bit about where their futures and their swaps need to be aligned?

SP-66 Yes. Okay. We'll talk about eligibility. So first, the way that Dave kind of mentioned this is quite complex, right? The way the portfolio marketing services works, right, and maybe this is obvious for some of you, but I'll just lay it out, right? What LCH does is it takes your swaps portfolio and it takes you [indiscernible] portfolio, and it assesses the combined risk and it identifies which futures positions are best included under the SwapClear risk calculation, right? So what does that mean? So every night, we run this thing. We identify which futures positions make more sense to be under the SwapClear margin calculations. And then we run our margin processes. So you get a SwapClear margin requirement, it margin requirement. And hopefully, that requirement should be lower once you're opted into the service. Now it's super low touch. You kind of just opt in and this thing runs every night. And now your question was more around eligibility, right? And that's key, right? So for clients, you need to make sure that you have your swaps and your future is under the same beneficial owner and you're clearing through the same clearing broker, right? From a member perspective, you need to make sure those client accounts are separately identifiable. You also need to make sure if you're looking for portfolio margining from on your house books, you need to make sure that you're under the same legal entity ID. So those things are important, right? You can't be transferring risks between different legal entities. You can't be cross-margining across different clearing brokers. So there are some rules that need to kind of in place. And we're happy to work those out with you. We've had many conversations, some bilateral conversations just to kind of help people get set up properly.

But LCH has really made this fairly streamlined. When FCMs have had these ones and then they connect in to you with their clients. And the clients' accounts are all lined up appropriately, like you just said, all they have to do is opt-in and they don't have to do any trades or end of day trades themselves. That's a very slick process. You do it all at LCH and the benefit automatically goes back through the FCM to the client to be able to consume -- it's very, very streamlined.

Touch right yes. And that's something that we saw, right? This is similar to FMX, right? We're the markets partner. We're out there listening were the pain points. And what we learned is with other with other players out there, the operational intensity around that process was quite difficult. So we've streamlined all of that and have it all working in the background.

Great. Is there anything else the FCMs want to mention about from your first conversations to getting clients up and running that you would like the audience to know because we do have a lot of clients in the audience today. Maybe I'll just start. I think I'd echo much of what's been said. I think Lou said at the onset, I heard him say it, and that it's -- the buy side is adopting this I think Kevin even said it, it seems like a renaissance of sorts, I agree. There's definitely conversations that are happening with clients. I think clients want to see this acceleration of volume and that the participation, I think, will be -- has the potential to be parabolic once it kind of gets to that point where we see that just enough volume there. So I think that a lot of people are just watching we're having conversations and I think people want to see this succeed quickly.

Yes. I mean I don't think anybody more than your previous panelist wants to see this more. And like firms of the same sort of liquidity provider, low directional sort of portfolios where they find the most benefit as general liquidity providers in leaning into MX's liquidity on the back of it to unwind sort of what they were going to see to be the initial trade. And as that builds out, obviously, there's no LCH has the volume then building off of that will, I think, grow this business with the liquidity. And then the less capital deployed for what your client sees is a relatively hedged position allows them to deploy more capital and more liquidity. So it feeds on itself.

Yes. There's an education component to it. We need to make sure that the client treasury teams, the client operation teams that there will involve setup process. They understand the mechanics of how that's going to be working. Your futures are 1 regulatory origin here in the U.S., you'll sort sort of different regulatory origin with a different set of rules and a different set of regulations associated with those. People need to understand what is it -- how does that when you're transferring from 1 origin into the next origin. What does that mean for your trades left in 1 area, you trades out being portfolio margin in another area. So it takes a bit of time to make sure that clients understand how that's going to look and feel from a bank perspective. And that's important wouldn't want to get this sort of thing wrong.

Yes. Subash, I want to stay with you for a moment. So now we have these clients that are up and running. They're experiencing the portfolio margin benefits with LCH. I want to take a step back to something that Prergati mentioned early in the conversation which is LCH's margin methodology is very predictable. And therefore, margin calls are more predictable than many other clearinghouses. Talk to us about the importance of that and what that means for you, your business and your clients. So that sort of stability in the overall portfolio margining. And I'm using out of using a really is broader sense of the term, not just within the FMX and the LCH space, but across the futures and across [indiscernible] more generally. Clients will want to understand what that portfolio margin -- sorry, what their margin requirement across each of those venues is going to look like and have a good understanding of how their trading activity is going to impact that. So being able to work with LCH to understand that fits together. Clients certainly seen that as a benefit.

Great. I think we're about at the point where we can take some questions from the audience.

So you have on the screen some indicative maximum notional savings numbers. I was hoping now that you guys are getting in front of buy-side clients plugging in their actual portfolios into the LCH margin calculator, what are the real life average savings that you're seeing practice on like how much does the total initial margin across futures and swaps, rates, features and swaps actually go down in practice?

Well, that might be a good question...

Yes. I can start. No, I think obviously, the numbers that you saw are hypothetical, right? And those are obviously best case scenarios. When you take a look at perfectly offsetting exposures that then you kind of see some of those like greater differences, but it really just depends on your portfolio composition, right? So basis or relative value strategies or intuitive candidates because they have offsetting exposures in the portfolio composition. When you take a look at potentially any kind of portfolio compositions where you already have a futures hedge risk that already exists in your swap portfolio, you're going to see benefits there. So it really just depends on that composition and kind of what your strategies are. I have seen numbers in that general range. It really just kind of depends. But using the LCH listed margin calculator or the swaps margin calcular. You can -- what you can do is it loads up your swaps portfolio and then you can put in hypothetical physicians and around a whole bunch of different scenarios and see what those benefits would be.

I mean realistically, we're getting 50% to 60%. But there are there's a hierarchy in the methodology where you might have swap versus swap that might take priority over that gives you -- under the LCH model, there's a swap versus swaps. So that could give you an efficiency long short, that might take priority over a future as it rolls in next. So there's a hierarchy. But in a clean vanilla how much is just trading incrementally going to cost, those numbers are in the on a portfolio, you might not see that 80%, 70% range, but you still are seeing a substantial number. And capital efficiencies, whether it's 80% or 50%, I think they're welcomed by the participants especially the liquidity providers.

Yes, I think just to add on that rate, it depends on direction, tenor, curve exposure, concentration. I mean, there's all these kind of factors that kind of bake right into it. So you do see those numbers.

Got it. And for the FCMs, I guess, how important is enabling cross-margining to the cash, like replicating something akin to what the the CME and DTCC have done? And then Bergade, is that a capability that LCH can help FMX with? Like is cash treasury clearing eventually something that you guys would be able to bring to the table.

I think you already prepared your answer.

Yes, I do. I was ready for the question. So thank you for that. Okay. So the clearing of U.S. treasuries is obviously topic on the post-trade landscape. LCH is obviously watching this. We've set up the [indiscernible] roundtables. We're watching the news reports. We're seeing what kind of other players are doing out there in terms of partnerships beta and what CME is doing, what [indiscernible] doing. But I think where our investment and our priority is, is kind of building out to what our strengths are, right? So we have built our SwapClear business on bringing together efficiencies across different currencies and building scale, liquidity and netting, right? And we think this is a compelling argument to add listed rates to that space along our SwapClear franchise. So while we're kind of watching how the markets develop, our focus has been on the future space and on the swap space.

Would you say it's fair that you're still analyzing it from a firm perspective?

Yes. I'd say that's fair.

I would say from the FCM perspective, it's 1 of the many things we're looking at. And I think there has just been an influx of new things that we're trying to do as an FCM, things are changing so quickly. I think all of this treasury clearing is consuming a lot of time, but we also have other initiatives. We have FMX, we have new exchanges we're joining. And so we're trying to balance all of that. But suffice to say, all of this around either whether it's cash or whether [ futures ] swaps, it is probably the #1 most discussed issue margin optimization with clients currently across this construct. There isn't a week that goes by that we're not having multiple inquiries from clients about sort of what we think about the likelihood of certain products, where do we think our efforts are, where are we going to be focused? How is our repo team looking at this, how are we integrated with are done with trades and how is this all going to look across the FCM. So to answer your question, I think it's one of the most paramount important things that clients are talking about right now.

Yes. And I think I touched on in an earlier answer, like what's going on now is 2 different CCPs that are sharing physicians. And what is the optimal part is now that you have to send orders of different CCPs and that gets really, really on the club level. But having a central one CCP that had the model that was uniform and transparent is I think where all the FCMs are struggling to see like, okay, how is this all going to play out with mandatory cash clearing and what was limited to house positions cross margin between cash and futures as that's getting rolled out to futures for clients, which we've been there since day 1, it's not without challenges. It is a very challenging problem. And I think there's a lot more to come as mandatory clearing goes full bore at the end of this year.

Yes. We want to make sure that we're available for clients when we have all of those types of offerings that are available. Client -- we will respond to that client demand. If the stuff is going to be around them. We'll look to see how that works. It was a viable offering, it's a competitive offering, if it makes sense with the broader platforms.

Any other questions?

So the numbers you put up there for the offsets with the swaps impressive. But I think probably aware the competitor offers even more offsets. I think in SOFR, it's well above 90% in treasury, it's 80% or 90% in some cases. I'm curious what the process is at LCH for to that level? Is it time? Is it open interest? What is it that gets us there to equal to the competition in terms of percent offsets?

I missed it.

He's asking how to increase the amount of go offset that a client can obtain because you've seen in the competitive landscape, some higher percentages with other opportunities.

It's a good one. Again, I think it just really comes down to your portfolio composition, just comes down to how -- what are your exposures look like? What does your tenor look like? What does your concentration look like? If you're 1 directional, you're not going to see much benefit no matter who you play with, right? So I think focusing on that portfolio composition is kind of where the key is. From a standpoint of getting higher than 80% I think there is some European regulations out there that kind of do some -- have some restrictions around that. But I think there is a lot of savings to be had, but it really just kind of depends on what those portfolios look like. Hopefully, I answered your question.

Anyone else? Well, thank you, panel participants for your time. I believe we are taking a brief intermission. Correct, Jason? Yes. Thank you. We're going to take a quick 15-minute break. If everyone could be back just before 3:30 please. [Break]

If I could ask everybody to return to their seats so we can start the next panel.

I'm Steve Merkel, General Counsel at BGC and the Chief Legal Officer at FMX. We have an interesting panel. Immediately to my left is Alex Titus, who is the Chief Adviser to the CFTC Chair, Michael Selig. Before that, he was Chief of Staff to Steve Miran at the White House Council of Economic Advisers until he moved on to The Board of Governors?

Yes.

Which suggests you probably have some sense as to economic impact with respect to these products or at least people look to you for that. We have then Josh Sterling who's at Milbank and has been leading Millbank's nationwide litigation efforts concerning the legality of events, contracts and prediction markets and does advise clients on other things, but including perpetual, security futures and other protocols. And then finally, we have George Canellos, also also at Millbank and who has done a few rounds in the government prosecutor at the U.S. Attorney's office here in Manhattan as well as the Co-Director of the Division of Enforcement at [ PSEC]. And so we're going to try to get into a little bit of all of that. But Alex, I did want to start with you and talk about get your perspective on when the CFTC cares about U.S. markets being competitive with overseas markets, when it doesn't, clearly, you were -- everyone was on the sidelines until the President's Trump 47 came into office, and there were already all sorts of offshore sports markets and predictions, they're perpetuals, there were assets trading on the blockchain, predominantly offshore. How did -- if you could just describe for us, how does the CFTC decide we want to make it easier for those markets to be traded in the United States for U.S. customers to not have to use a VPN to pretend to be somewhere else, but to actually do it here. How do you decide what's worth bringing importing and what isn't?

Sure. Well, I'll -- actually, my lawyers tell me I have to start off by saying my views might not necessarily reflect the Chairman or those of the commission. So anything can be attributed to me personally. But as you had said, a lot of these platforms, which some of them originally had started in the United States went offshore under the previous administration, particularly German Gensler over at the SEC had this mantra of regulation by enforcement and intentionally tried to push the crypto industry offshore. And I think that's had 2 really negative effects. One is the investor protection. Of course, there are some folks in the U.S. or the EU or other jurisdictions who are using to go into these offshore platforms that are not really regulated by anybody. You can have funds that go missing. There's platforms now that are offering 1,000x leverage actually saw some tweaks this morning that someone is trying to offer 10,000x leverage. So obviously, there's the sort of consumer protection regime that is built into CFTC compliance that folks, if they are trading U.S. offshore getting which I think is unfortunate and is bad for consumers and investors. And then second, a lot of people are clearly interested in these products and want to trade in them. So it's been a big mantra Lake to bring a lot of these platforms onshore in a safe and responsible manner. So we've been working very hard at that.

So Josh and George, do you -- what's your perspective on this CFTC trying to have the U.S. be more competitive with respect to these new markets.

Yes. Well, I think there is a real sea change compared to the prior administration. I think you find leadership of both agencies are really focused on how to make America more competitive because vibrant financial markets lead to or sustain a broader, more vibrant economy. And it doesn't make any sense to sort of balls and strikes on what's in and what's out. It's a particular type of asset class. Certainly, you could have assets here that weren't as prevalent before, maybe not times leverage perhaps. But that's sort of a line you can draw once it's here. So I think they're doing a good job. I think they're trying to be careful and thoughtful about it. It's interesting. A lot of it is pushing in areas that exist in the law, but really haven't been touched in 25 years or more like security futures products. The rules are there. everyone's forgotten them, but they do offer a ready path for innovation and people are making real progress on those paths. And so to the extent you can innovate and bring onshore what's offshore through existing rails, here that the regulators know, I think that's sort of a quicker way to do it. And I think both agencies have been fairly accommodating in that.

George, any perspective you want to bring from...

Well, I'll just add a comment maybe from a little bit of an SEC perspective. I think right now, you have the swaps markets innovating in all kinds of interesting ways. And some of the ways that they're innovating are extremely disruptive of some historical state laws or you might say state sources of revenue. But you haven't had that much push into the securities-related markets. it's beginning to happen. And I think as these -- as prediction markets and swap markets in general, enjoy greater and greater acceptance retail participation, you can expect that there's going to be a great deal of interest in products that are within the SEC's jurisdiction. All all securities-based swaps, which would include any kind of predictions market bet relating to an issuer or the performance of the issuer or the value of an issuer or even events pertaining to the issuer like will Amazon reach $750 billion in annual revenue, will Ozempic be approved by the FDA, all of those very interesting economically driven swaps would be within the SEC's jurisdiction. And I don't think -- I think while the SEC is making all of the rate is offering all the right talking points about its openness innovation and its openness and willingness to cooperate fully with the CFTC and try to come up with ways to come up with integrated regulatory approaches for in market that cross both CFTC and SEC jurisdiction. I think the SEC has been much more reluctant to permit some forms of what you might call innovation or we might just say, encroachment the historical regulatory regime of the SEC, which requires extensive disclosures, which tends to forbid retail participation without extensive disclosure regime. So I think the CFTC is probably a little further along the line in accepting some new products in the SEC is.

Well, from a policy perspective, and this is for Alex and Josh, does that suggest that the CFTC not necessarily having a focus on disclosure is at risk with respect to how retail investors are treated trading these products.

Right. I think this gets down to the fundamental difference between the securities markets and the markets for financial derivatives. One is the price discovery market that's the CFTC. And the other one is sort of a wealth accumulation capital raising market where you actually own something, an investment contract and that's the SEC. And I've had a lot of conversations with the SEC for people wanting to develop novel products that have some SEC overlay. And the conversation can often begin with while you're just throwing out 90 years of history on getting offering documents and annual reports out people just want to trade the price of a stock rather than own the stock and receive the dividends in both the proxies. And the answer to that for someone who has more price discovery biases, yes, that's exactly what they want to do with the ex problem. And they have a problem with that. So to me, it is true that on the CFTC side, there's much less disclosure. I mean if you're offering a fund that trades futures to retail people, it looks a lot like the disclosure mutual fund. But outside of that, it's sort of no, you're in the market, you're going to find the right. And if there's a problem with the price discovery mechanism, well, we'll come after them with our market you're right, it's much less disclosure, and it's always been that way. And I just think that stocks are thought of as more of a retail phenomenon than say, futures and swaps are.

I think the one thing I would add there a jurisdictional, I'll call them, turf battles that have plagued both the agencies for quite some time. Obviously, I think we're seeing that come to light faster with some of these new registrants that are coming in offering products like security futures and things like that. But that's exactly one of the reasons that we have this SEC, CFTC, harmonization effort. I think it's safe to say that Chairman Selig again, Chairman Adkins are probably the 2 closest CFTC and SEC Chairman that have held the seat at the same time. And obviously, these are -- these can be tough conversations. It's a lot of policy to work through, but we think these are the exact issues that with when it comes to harmonization and that relationship is so important, which, of course, at the end of the day, it's not really a turf battle between the agencies as much as it does this benefit investors benefit the market overall. So I'll just add that.

I want to talk about...

And that's probably the first time since 1974 when they haven't been essentially rivals of one another. I mean honestly as much as you would have prior talking about seamless cooperation between the 2 agencies. You almost never had that in reality. You would have the same sanction in both duplicatively against companies for the same violation. And I think 1 of the areas where you can just give both agencies enormous credit is for recognizing that, that was a problem, and that was a 50-year problem that has permanently. It's maybe just been solved for the next couple of years.

It's a good example for people to follow. So I wanted to talk, I guess, at least one FMX question at least. So FMX positions itself as a differentiated futures exchange that will provide competition to a pro-competitive alternative with the existing leader in in rates products. Is that -- it's sort of a so-what question. Is that -- are those factors important to regulators is the fact that pro competitive. Does that make a difference? Does that ever really matter to, let's say, the CFTC and should it matter?

Yes. I think if I had to speak for the Chairman, I would say that it is important, of course, give one example where our largest registrant just happens to be pursuing us over a new product offering in the U.S. So I won't add any more detail to that than that, but I'll let that speak self. But I mean, I think we think competition is good in a few ways. And one, with competition usually comes innovation. I think the CFTC is probably the only government agency that actually the promotion of innovation is within the statute. Maybe somebody can point out that I'm wrong, but I was quite shocked to actually learn about that because generally, we think it's government to be the least innovative entity that does exist. But I mean I think competition can generally be good for incumbents as well I mean, you keep folks on their heels, you force them to try to innovate, you force them to compete. I think it could be good for them in that way and it can be good for the overall market and for participants who are trading.

Yes. I agree with that. I think that when I was at the CFTC in the first Trump administration, the thought was always how can we get more people market, how can we get more dealers into the market for swaps. It doesn't necessarily have to be that the largest swap dealers are the largest or are the only ones. And how can we lighten the burdens for some of them that are smaller or have sort of a different acts in the market than say, we're here just to provide this solution for everybody alongside the lending business. So I think that's true with the markets and clearing houses as we clearing house diversity is an issue. It's important to have more of those. And it's important to have clearing houses like a anteaters in different areas like futures. And I do think that there is a strong interest in encouraging that. One thing I've often said in discussions with the regulator for clients as you're not in the business of issuing monopolies. You're not a state giving someone a gaming license. So you should allow more people to do this and give them a fair shot when they want to innovate which you always run up against is when you have an innovative idea because there's the incumbent because the incumbent has done what they've done for a long time, the everyday staff at the agency will say, well, what you're doing doesn't look like what I know. And so there's a bit of a mental block which you do need to break down and sometimes that involves discussions with the leadership of the agency just sort of making the substantive case like look at it may be different, but it's still possible. It's still legal, so let's do that, shall we.

So Alex, I want you to talk a bit about [indiscernible], right which is if you sort of look at what's an astonishing area, which is a major industrial commodity, copper aluminum. It's not really -- the futures are really traded here. I mean there is some but not really. They're in the LME run by an exchange, which when it feels like it can cancel billions of dollars of nickel trades, right? So this would -- I'm just wondering, is it appropriate and for government to -- the U.S. government to look at that and to say this is a risk, right, with respect to our manufacturers and others. And if those who aren't aware, just take a look, just now on the Internet with respect to the LME canceling the nickel trades. And it's sort of shocking. It's the only way to put it. And they -- I guess, they could do it in copper, they can do in aluminum. What is -- do you think is the U.S. role when there is this kind of potentially economic vulnerability to our factors?

Yes. I mean this is a top priority of commission during the Chairman's first major speech at the FIA BOFA conference. He was very clear when he said we want to create and bring back markets for base metals and minerals in the United States. And I mean, you basically summed it up right, Steve. Most metals trading, price discovery and setting is done on one offshore venue in the U.K. And I think that we do have some concerns about that. I think there are some concerns on the financial risk side of things as well. And I mean the President, of course, talked about this issue, right? I mean he's wanted to bring back steel production states, bring back aluminum production. There are steel plants coming online as well as a new aluminum smelting plant in Oklahoma. But we want to see more of the financial activity happening in the U.S., more of the prices happening in the U.S. and more of the actual physical storing of the metals in the U.S. as well. And I think it's really interesting to see what's happening in copper markets because the U.S. actually had so much copper right now. in the country, they don't have a place to store it. They're storing it in trains and then I think some of it is stuck on ships as well. So it's a top priority for the administration as a whole as well for the Chairman. I would say we want in this space. If anyone has any interest in this issue, I'd be happy to chat about it as well, so.

That's a great invitation. I guess I want to point out, and it's not just because of our friends clearing house, but I don't know whether or not the LME can be strictly thought of as being run out of London anymore, right? I think there's another -- you have to go a few thousand miles away before you find the real governance. Fair?

Yes.

I did want to talk a bit about, again, the prediction markets and how futures swaps, security-based swaps, the SEC jurisdiction, CFD jurisdiction can get a bit jumbled. And just talk a bit about -- if you were here, it was a fair amount of discussion of basis trades, which come up in a variety of ways, but the simplest is really cash treasuries against treasury futures. And people -- that's a very esoteric sort of trading. But if one were to go on a step beyond that and think about prediction markets on basis trades and imagine there would be, in fact, a prediction market as to how a base -- what would be the basis at a given day in a given time for a given maturity. How should lawyers, how should -- how might they advise their clients, how would -- how should the government think about that both with respect to over-the-counter markets where you might just have very sophisticated customers doing it as well as possibly sitting on DCMs. How is that kind of multi-jurisdictional [indiscernible] type of a prediction market on top of, right, what is basically a CFTC and SEC conjoined product.

Sure. So I'll welcome George's views, too, of course. But I think that I'm all for markets. It's what makes this country what it is. it would seem to me to be a good idea. I think that basis trading is important. Illuminating further base of strength through prediction markets will probably be helpful. There's information in it. And having a venue or set of venues like FMX does that allows for cash trading and derivatives market trading, which underpins space is trading is useful. More of that and predictions on it provide more light, more information and might improve the ability to surveil the market that might provide efficiencies. It might provide different changes in trading behavior become more visible and better surveilled. So I think that there are a lot of good -- there are a lot of potential goods for it. I guess one question could be, if you accept the premise that a lot of trading operation to date has been retail. Well, people really know what they're doing. I think the answer is yes. There's a lot of literature out there to suggest if your typical picture market trader Sure, there's is actually a pretty knowledgeable person who comes more from a trading background than, say, gimbling background. And so I think it's something worth pursuing.

So you're talking about basis trades specifically?

Well, just talk about any prediction mark that represents on a product that has both futurity and security.

Yes. So Look, I think a theme in response to many of the questions, Steve, that you posed today is that we -- if we really could remake our history or we remake our future, the CFTC and the SEC should be one agency like it is in many countries. Because when you identify basis trading, especially cash versus treasury futures, for example, you're basically evoking to historically completely separate regulatory regimes. And even though you're talking about a product that is more an ear to the art of right? So it would present questions of -- important questions about how one harmonizes the 2 regulatory regimes even doing such simple things as ensuring that you could have cross clearing so that you could net physicians presents more regulatory complications than it should. And so I think we should strive towards more integration. And if I had it my way, the 2 agencies would be the same. I will say though, I don't think that particular product is one that would be very controversial because I think the greatest concern that the SEC often expresses when it comes to prediction market products that maybe encroach on traditional securities market is that they have all the ability to present opportunities for trading with none of the disclosure obligations. And if you're talking about like basis trades in the U.S. cash treasuries and U.S. futures markets. I don't think that presents that concern since there's no real disclosure regime for the U.S. government. And I also think that it's the kind of product is, by its nature, one that would be attractive for institutional investors. So the extent that the SEC has a strain of concern, protective, paternalistic concern for retail investors. I also think that concern probably not pronounced here.

George, do you want to talk a bit about enforcement and let's stay with for the moment, predictive markets. It's still a -- it's had an early stage and there's an interesting question whether the agencies should be really looking at them carefully for enforcement purposes or where there could be a fair amount of mischief at an early stage, the Wild West or wait until the market sorts it out and the users figure out which platforms have integrity, which ones don't. And only after that sort of maybe do a look back as to -- as to what occurred, do you have a sense as to let's say, if you were involved in enforcement. What sort of investigations, what sort of data would you look at in order to try to get ahead of what again could be enforcement issues and perhaps prevent abuse. Specifically in prediction markets. So start there because it's sort of the newest, right? It's the new great thing. And I don't know if there's been much enforcement action yet.

Right. Well, I guess I'll say this. First, one of the big concerns and predict -- 2 big concerns on prediction markets are market manual and insider trading. And those concerns are very real ones. I think -- and there are a reason why many people will not participate in prediction markets. I will not touch prediction markets because I fear that others, my counterparties have more information than I do. I don't feel that way in the securities market. So I do think it's a grave concern and if it can be addressed, it would help those markets a great deal. But it's not -- there is not without -- the challenges are great. If you take an example of like the securities markets, why do we have so much insider trading enforcement in the securities markets, especially equity markets. I mean, it used to -- because we have a perfect paper trail of surveillance. We have an organization called FINRA that tracks equity trade that applies algorithms to look for and surveil for possible MMPI that will cross reference major market events with trading activity that assimilate information that comes from multiple markets so you can compare options and ordinary. Do you have anything like that in the prediction markets yet. It's developing. You don't have anything like that. But you do have very motivated process with all the same tools that they have always had mail fraud, wire fraud, including actions under both the commodities and securities laws that allow them to bring criminal actions and that, I think, and they will and do. We also have increased pressure on the exchanges themselves to engage in meaningful surveillance. And I think over time, as the markets become more popular, you will probably have the markets sells enhancing their ability to surveil and so we'll see enforcement actions, and we will see, I think, increasing transparency and integrity in those markets.

Yes. I would add on to that. I mean, obviously, the markets have grown significantly just over the past 1.5 years or 2 years. I mean I think [indiscernible] and poly market were just mostly election contracts in the run up to the 2024 election. Obviously, list a bunch of different products now, whether it be sports, whether it be culture or whether it be commodities. Obviously, they're in different markets for whatever market participants are interested in. But this has been an area of top priority for the Chairman. I mean, as he has said many times, the comp is on beat. We have a great enforcement Director. I used to work at SDNY, started his career at the CIA, David Miller. And I would just say to anybody, if you think you can get away with price manipulation trading or any other sort of shenanigans on prediction markets. We will get you the SD and we'll get you or the SEC will get you. So I would be very cautious for anybody who thinks it's the West we are watching.

We'll all be watching I'm actually going to see who you're going to bring down because I think they're still distrust, exactly what George said. People don't really feel at the moment that it sort of you don't know...

I think they're all very motivated. The problem is the surveillance tools. You just don't have the data that you had in that market. I think that's 1 of the bigger challenges. But absolutely, I think everyone is committed to these areas.

I wanted to talk about the crypto markets. And the question I'm going to ask is sort of what why do we think that they necessarily matter? And I want to distinguish between the cost benefits and primarily the benefits of the blockchain and its friction with 24/7 opportunities, particularly on the payment side. I want to distinguish between that and the crypto currency and just get a sense for if I told you that Congress will never pass the Clarity Act and if I told you that in January of 2029, there a new administration, which will take a very tough look and ultimately will lead to the very high values that are in some of these cryptocurrencies that if not the bubble bursting a severe decline in what they're worth and very few people think of them as a store of value, notwithstanding the utility of blockchains. If you knew that was going to happen, would you continue to try to create the framework for these very difficult to value, very difficult to disclose cryptocurrencies.

You gave me a really easy question for the last one, Steve. I would say, I mean, the market capitalization, I know, for most of the major cryptocurrencies is currently -- I think it's a little bit above $3 trillion or something like that. I mean obviously, a lot of people just like to trade these products, whether they're trading them offshore, they're trading them in the U.S. There was a lot of trading activity. I mean, really, really across most of the world for assets. And I would say on the Clarity front, we were very disappointed that Congress wasn't able to pass that bill and bring it across the finish line. And I think by the end of it, I mean, I don't know if -- and I won't speak for any of the senators, I don't know if anyone was perfectly happy with the final bill. But to me, if no 1 is perfectly happy, that sounds like compromise. So I mean, the President, the White House, they worked very, very hard to try to get this bill across the finish line. And unfortunately, it didn't. But some of the folks in the audience may have seen, we actually just rolled out a new advanced proposed rulemaking, both on crypto transactions as well as crypto asset market. So with the CFTC, and I'm not going to prejudge anything, we'll obviously get comments on this see how everything plays out. But is a federal option for those different exchanges or startups who want to have a federal opt out from some authorities that are currently overseen by the states and then overseeing certain types of transactions that always overseen whether they be leverage margins or finance. So I think a lot of folks will come to the commission with really interesting ideas. We'll kind of see how that works out. But I would be a little bit more bullish than you, Steve. I think crypto is probably still to stay even with the change in administration.

I won't ask for your prediction as to our Bitcoin trades in January of 2029. But I will ask Josh and George, do you think that there is a chance of regulatory with respect to all the work being done with respect to the crypto currencies.

I think there's the possibility of that. But I think any time you try to do something new or different, that's always a possibility. And to the extent you're doing it by regulation, it becomes with appropriate authority and the statute I would that, it becomes that much more susceptible to abrupt change. And the concern is things will swing back and forth more broadly on the pendulum between parties and control. So I think that what could be done by regulation or order can be undone by regulation or order quite quickly. So there's a concern about that, but with a lot of innovators we meet with our basic message to them is, look, as you -- I can't tell you the business case, you didn't hire us for that. But my advice to you is to consider getting your license and building and issuing products quickly, once you have a license, you have constitutionally protected you process right and how that licenses given to you maintained or you've taken away. It's a very important thing to have in your quiver. And then issue products, so there's substantial reliance interest doing. And that will always make political constituencies at least think twice about undoing it. So you can manage it, you can't prevent it. But if it all goes away, if the advanced notice becomes a proposal becomes a rule, as I expect it will, and it goes away suddenly. I'm sure there will be a lot of regret about that. But I don't think it will be regret from having done it, it will just be that we can have a committed course of action in this country around a quite large asset class. Nothing to add there.

Okay. like Steve, why don't you just tell the audience what you would like their takeaway to be with respect to the work that Chairman [indiscernible] and his team are doing.

Sure. Yes. So I would say 1 thing we've tried really hard to do with industry -- just actual engagement. So we like to joke but now we actually have folks come into the building, the share ideas. Some of these ideas may actually turn into policy action. We're not going to host you in the building just with the subpoena on the way out the door. So we meet -- I meet with folks in the industry all the time, no matter if you're working on Dadi, on crypto, on prediction markets, compute metals. We are always open for a conversation. I think really good ideas come from market participants and as well as some of our registrants and where policy should go and where areas of concern are. So I think that would be the biggest takeaway to the folks, if you're interested in engaging with the CFTC more as we'd be happy to have the conversation.

I don't know if there's time for questions, but I'm if anyone has any questions for these panelists?

Sure. Yes. Just a quick one. So we've heard today from FMX clients who have talked about this idea of counterparty selection or counterparty blocking being a big value add in the treasury club. It's my understanding though, that, that is not allowed under the Commodities Exchange Act. So future menu like FMX would not be allowed to do that. Is that something that maybe the CFTC is reevaluating -- or just would love to get your thoughts on that.

Happy to take that question with you off-line with the colleague who might be a little bit more [indiscernible].

Thanks for doing this, Alex. You mentioned at the beginning that part of your mission and the statute is promoting innovation and competition. I guess when you look back, clearly, in the industry futures market, your predecessors haven't lived up to that objective. And like I wonder if you and the commission have diagnosed why interest rate futures evolved into a monopoly. And I guess what the takeaways from that postmortem are and how they play into your approach regulation on a go-forward basis?

Yes. So I won't, I guess, address the specific market dynamics, but I'll take the competition part of your question. So I mean, first and foremost, I would just say allowing folks to actually enter the market has been a high priority of this Chairman. So obviously, we have DCMs, we have DCOs, these applications come in front of the agency. For the first time in history, we're actually approving new registrants on the 180-day time line that has been given to us by Congress. I'm surprised that's not really been an issue. I guess, the concern of Congress in the past that they've given us this mandate. We've kind of failed to deliver on it. But I would say that is one bare bones thing that this Chairman has wanted to do is if you are raising all this capital you want to enter into whatever new market it is. If you put in a good application, we will get questions to you in a response too quickly to actually allow you to operate as a business. So I do think that, that piece is really important. And then again, in terms of just the new product offerings that we're allowing, I mean the perpetual future again, won't do any name calling here. Some have at least called systematic risk to their business, right? We are bringing in these new products and promoting innovation. So again, won't really respond. Maybe 1 of these guys will market dynamics. But I mean, again, I think competition is better for our registrants. It's better for market participants, and it's just better for the public in general.

Can I add 1 observation there? I mean 1 of the challenges is, and this pertains to all financial markets. A lot of financial markets are -- have some natural monopoly characteristics. So deep liquidity, centralized clearing, netting efficiencies, create competitive advantages and huge barriers to entry. And there may also be some activity by the monopolists in the sector to try to reinforce that monopoly power other arrangements that could implicate antitrust laws, but there is some natural monopoly or networking effects in the markets. But 1 of the areas that's I think most interesting in the financial markets today is that some of the products and innovations do allow -- I mean, ideally, you don't want to have a monopoly. You want to take all the advantages of having a natural monopoly, but you don't want to have the inefficiencies and abuses that can come with a complacent monopolist who's not innovating. Some of the big innovative products today are ones that help to break down those barriers like interoperability of clearing better systems that allow 1 to assess and take advantage of liquidity in multiple markets at the same time. And I think those are the kinds of products and services that you really want to have regulators encouraging because they give you the best of all worlds. They give you advantages, the clearing and netting opportunities without giving you 1 complacent competitor who's no longer innovating.

Any other?

Are there any concerns like on a risk perspective from retail investors on some of the new products you're talking about, whether it's perpetuals and like the leverage that comes with that? Or some of the loss statistics for retail investors on prediction markets. So just from a CFTC perspective, like how are you looking at that?

Sure. Yes. There are currently no prediction markets in the U.S. There may be some offshore they're offering leverage for their product. It's something that -- at least there's been some news reports on this from some certain exchanges, but it's not something that we've approved, at least -- now these products have to be fully collateralized. But in terms of the leverage point on perpetual futures, I mean, again, I think this is just such an important reason of why we need these products onshore. I mean you're just never going to see 1,000x leverage at the CFTC or 100x leverage. And you have these offerings on -- I mean, fairly popular offshore platforms that have a lot of liquidity. So I mean, we think those are just too risky from our perspective for most retail investors. Obviously, anybody who's offering leverage needs to go through our division of clearing and risk to be able to make sure they have those proper margin levels. It's actually reviewed by staff as well. So Again, if folks want to trade these products. Of course, that's kind of up to them. We have a very free market view of this, but you're just not going to see things like 100 or 1,000x leverage when you have the CFTC, consumer protection framework built in place. So again, I think that's a really important reason to get this stuff onshore compared to having it offshore.

There is -- thank you. There's only so much longer I think I should be delaying my CEO from speaking next. Who is the next speaker, but I want to thank the panel for a very enjoyable and interesting conversation. Thank you. all tariffs are in SP-12.

Okay. So in the interest of trying to catch up on time. I know this is probably part of the day that certainly, some of our shareholders, future shareholders and analysts have been looking forward to in terms of breaking out. What does this all mean for BGC? And so for the first time, we'll break out where FMX is in terms of its financials. I go back 3 years, and you'll see that the FMX business has grown from $48 million up to $122 million for 2026, which is quite some 36% CAGR in terms of revenue growth. And you'll see that obviously dominated by our rates business, starting off in cash treasuries. Our FX business, 1 of the fastest-growing businesses, although only going from $6 million to $12 million, a 100% increase but just think about the potential size of that marketplace. And I think in 1 of the slides that you saw that Lou put up earlier on, you saw where we go from $1 billion up to $18 billion a day, and that marketplace is significant. Clearly, market data, market data -- market data on treasuries. I'm going to take you back a little bit as well, which is if you think about when we sold eSpeed, back in 2013, we sold eSpeed was generating circa $100 million worth of business, made up of just under 20% in market data and the balance obviously in execution. And so here, you'll see the market data numbers are pretty small, again, so with huge room for growth. And I think what Lou pointed out earlier on is that we have a sort of 6% market data piece, CME, of course, has 13%. So again, significant room for growth. But let's turning here to compensation. I'll break it out into 2 categories because these are electronic marketplaces. These are exchanges. And so looking at compensation, what you see is you see how it's going exactly in the way you would expect it to go. You go from having compensation as a higher percentage in the high 60s basically because you're not charging revenue at the very beginning, shrinking down to now into the levels of just below 40%, a front office comp. Look at technology. technology costs now, obviously, there's a significant amount of expense that goes into building the platform. And you can see there that, that started up at $34 million and now 3 years later, is still at $33 million despite the fact that we've now got multiple new products. And then finally, within that -- within the SG&A, I would say -- this -- all of this reconciles to our AE numbers within for the overall BDC group. But you'll see there how the loss gone down from $29 million back in 2023 to $18 million. And just remember that back in 2023, we had no futures platform. The futures platform that exists today is already there already paid for, and therefore, we have tremendous operating leverage within the business. So there it is. You can see the -- you can see that the front office goes from 56% down to 39%. Again, if I go back to when eSpeed was in play, back in back when we sold in 2013. -- compensation, I would -- we would expect the compensation to continue trending down because right now, there's the expense of futures, but there's no revenue for future. From a technology perspective, gone from 70% down to 27%. Again, you've seen the stability of that cost base. And so what do I think going forward? Well, we think we think the following. So if you think that's where we've gone in the last 3 years. If you simply -- if you simply continue that growth, they remember -- we don't give -- we do not give full guidance for full BGC. So what we've given here from an illustrative perspective is to say, if we continue growing the business in the way that we've grown for the last few years, -- and I think after everything everybody's heard here today, I think that is a significantly high chance that's going to be happening. Then what happens revenues go over those -- over the next 3 years, revenues in that basis if it was to grow in the same as it has in the last 3 years, revenues get to that north of $300 million. And the incremental margin is 50 points, at least 50 points. And again, I think one of the things that we learned was when we built the cash treasury platform, we built it at a time where interest rates were pretty much 0. And therefore, the trading volumes were probably as low as they had been. And the questions we used to get from our investors were when you're going to see revenue. And we said, look, it's about building volume. So when we created the -- when we created the Futures platform, we were very clear on our 3-year plan. It was all about connectivity then getting volume and open interest and year 3, which literally we started right now, and that's when we start to become -- that's when it's about increasing that market share. And so I think I think that we have a tremendous operating leverage in terms of revenue and think about it, right, how do we generate the revenue. We have -- and this is a playbook that we've executed many times. You've got the all you can eat, all you can eat deals, which, of course, break unit economics for the key players within the marketplace. You've got, obviously, those paying on the fee per million charge on the per million of notional traded. And then you've got the FMX Futures obviously charged on the futures contract, which is 0 today. And then, of course, you have the market data piece as well. So I think a natural follow-up question, of course, will come. How -- I mean, you could just imagine how you map that into what is BGC, but I think everyone here can see that with 36% growth with a business that actually has all its costs already within the business. I think FMX, from a financial perspective is in fantastic position. And with that, we'll make up some time. We'll go to some Q&A.

So it looks like the market share of FMX and broker tech -- FMX UST and broker tech has more or less converged like roughly 50-50 now. But they're booking about $160 million a year of revenue. You're at $122 million, and I think estates FX. So $103 million. So is there any plan to pull the lever on pricing? It seems like you guys are coming into the significant discount to your closest peer.

Sure. So the answer is twofold. There are -- believe it or not, it's still is in its infancy. So you're correct from a market share perspective, they're at 42% market share. 2 ways in which we can -- or 3 ways reading, which will increase revenue in the cash market. Number 1 is, number 1 is in terms of there is incremental increases each year in terms of the fixed fee. So you'll see the 1 price deals go up in 2027 and beyond. -- already contracted. Number 2 is in terms of the fee per million, if I think about where the fee per million is at the moment in the early -- in the -- in what is the early stage, I think that has the potential to almost double from the point where it is right now as more participants come in. So that's 2 things within the cash treasuries that absolutely, we think will happen.

Can you hear me? Just on the cost going forward, can you walk us through how much BGC is going to cover. I know you said this flows through your income statement, but how should we think about how much of the commitment is going to be from BGC from a cost standpoint going forward? And is there any appetite or need that you foresee in the future to maybe raise more equity capital and just how you're thinking about funding the business broadly?

Sure. So there is -- the second part of the question is there is there is no need to raise more capital. I mean, as you can see here, with a platform, a fully built futures platform that on a pretax AE basis is 8 years, a loss-making of $18 million. I think on a company that generates revenues of $3.4 billion and of course, with the money that we put in by the investors already. But the answer from a capital perspective is absolutely not. This really is now this is a built platform. You obviously heard that -- you heard that we will, of course, continue on a -- you can see there on the technology. You can see that the run rate is stable, whereas the revenues, of course, are increasing. So I don't I don't see that technology line in line. I think what happens is if I think back to the days of eSpeed, you had -- you actually had an operating margin overall. It was pretty much based around -- It was around about -- it was around about 25 points in compensation at around 15 points in technology. SG&A. -- do you get -- what I did in my illustrative piece for the next 3 years is I suggested a 50% incremental because my view is I'd rather make sure that we -- I don't want to overpromise, I think that is a -- that history tells you that, that is a highly achievable target and 1 that we believe we will achieve over the next 3 years. Great. Well, thanks very much, and we will look forward now to listen to the main event of the day now, which is, I think, will be an incredibly interesting discussion with Professor Hinton. So thank you very much.

Okay, you want to sit in the middle or you prefer sit in the middle. I can't tell you how honored and privileged we feel to have you here. So thank you very much for coming. -- it's oftentimes that you watch someone introduce someone else, and they use the phrase and they say, "This person needs no introduction. Well, I can tell you, this is 1 of the times that I can say with complete sincerity. This person needs no introduction.

And now you're going to go...

And now I'm going to go and give you an introduction of course. We did a little prep work out there. So don't worry, he's used to my humor already. So Mark and I have been so looking forward to this. So Professor Geoffrey Hinton spent 30 years pursuing an idea that his field had set aside -- sorry, 50 years there you go. In the 1980s, the notion that a machine could learn the way that the brain does sat well outside the mainstream of computer science. He believed in it any way, and he kept at it for decades with a small and dedicated group around them in. In 1986, published the paper on backpropagation that made modern neural networks possible. In 2012, working alongside 2 of his grad students, you build AlexNet, which blew away the field in an image recognition contest. Within 18 months, the entire tech industry had reorganized itself around his work. Every AI system that any of us touch today traces back to what he created. He received the Turing Award in 2018 and the Nobel Prize in physics in 2024, only the second person to ever win both. What I find most admirable but our guests is what he's chosen to do since -- in 23, he stepped away from Google, so that he could speak openly and independently about the dangers inherent and the technologies he helped create. Professor Hinton, we are beyond honored that you are here with us this afternoon. But before I let you off the hook in your introduction, I just want to mention a couple of things that are usually not included in people's introductions of you and perhaps not in the first things they learn about you. So one of my favorites is that before you earn your PhD, you spent almost a year as a carpenter, true?

Yes.

The second is your great grandfather was in fact, George Bull, who gave us bullion logic, which is to say that the foundation of every digital computer in this room, in fact, came from your family lineage. Great grandfather. -- great, great grand falling excuse me. Also another great one. Your middle name is Everest after your great grand uncle, George Everest for whom the mountain is named.

Mean the into name [indiscernible] same question.

Amazing. And here's my absolute favorite sleeping in a hotel room somewhere in the world, you happen to hear your phone ring. -- you pick up the phone ring, you pick up the phone ringing, you answer it, and there's a Swedish voice on the other end of the line. At first, do you think it might be a prank phone call. Soon, it's someone telling you that you've won the Nobel prize. So after calculating in your head that it has, give or take, twice the likelihood of being a prank phone call than it does being a real phone call, you think to yourself -- and as the legend goes, you may, in fact, have set out loud, Well, this is interesting that I'm winning the Nobel Prize in physics because -- and I quote, "I don't do physics." All true. So in addition to our remarkable heritage, all the extraordinary things you've done, we can't wait to have this discussion with you. So again, ladies and gentlemen, Professor Geoffrey Hinton.

Thank you. Let me just tell you 1 thing about -- when I got the move pricing, something I don't do. I did a little calculation which went like this. What's the chance that a psychologist hiding and computer science will get the Nobel prize in physics, Well, maybe 1 in 2 million, and that's generous. What's the chance if I'm dreaming, I get the Nobel price of physics, Well, maybe 1 in 2. So that means it's a million times more likely it's a dream than reality. And for the next couple of days, I went around not entirely sure this wasn't a dream because you know when you have dreams you can fly, it's great, and then it turns out you can't. But 1 thing comfortable well I thought it might -- was probably a dream, which was it would turn that nightmare back Trump was just a nightmare.

I'm sure that the question that you get asked most of these days is sort of doom predictions of what's going to go on with AI. And certainly, when we opened it up to people e-mailing us questions, that was the largest 1 that came through. So I don't want to jump right into that. But if you would indulge us for a minute, 1 of the questions that off does not get asked today is what exactly is AI? And can you explain to us, if you don't mind, when you look at back propagation, you've been quoted many times saying that when AI gets complicated enough, we don't actually understand what's going on in the same way that we don't understand neural networks that go through in the human mind. Well, I can tell you definitively, if you don't understand it, most of us have got no chance. So if there's a way you can help us to try to understand what that means and the way you think about it, that would be great.

Yes. I just wrote a book about that. And if you're a professional explain things, you can explain them at any length. So just tell me how many minutes you want me to explain your [indiscernible]. .

I've got all the time in the world Professor. Maybe 5 minutes, that would be wonderful.

Okay. So what I'm trying to explain is how large language models work. Now -- there's a whole branch or in the history of AI, where people thought that understanding a sentence means translating the centers into some internal logical language that's not ambiguous. And that's a completely wrong model of what understanding is. So I'm going to try and give you a different model of what goes on when someone understands the sentence. And this is true for people, and it's true for these chatbots. . You have a string of words. Now actually, it's worth fragments, but we'll ignore that for now. Let's base hits a string of words. And the meaning of a word is just a big bunch of activated features. So for the word, the meaning of CAT is a big bunch of features, which are active neurals. -- like it's a predator. It's got whiskers, -- it's probably a pet. It's a loop despite the size of a bread box -- it can react much first than you can. Lots and lots of features. They're all learned by a neuro-net, -- so that's the meaning of the word. And -- we can use words to build models of anything at all. So the reason we're very special monkeys is that we came up with a modeling technique. You're used to building models use sort of regression models and things. We have a modeling technique that's much more sophisticated than anything a statistician -- at came up with. We can model things with words. And the way to think about words is their building blocks for building models -- so this is what languish never really understood. They always saw it in terms of Cyntec, ingots. But really words that are modeling technique, but it's just a very flexible way of building models. And let's start off with another way of building models that we know about, which is LEGO blocks. So if I have a bunch of rectangular leg blocks, I can build a model of any 3D shape if you want to model the shape of a push, I can put together lots of legal blocks that basically captures where the stuff is. The surface won't be quite right. It won't be very aerodynamic, but it will model where the stuff is. So Lego blocks are kind of universal for modeling 3D distribution and matter. Words are universal for modeling anything at all. And they're a bit like Lego blocks, but they have 4 main differences. The first is there's many more different types of them. You routinely use about 30,000. The second is they don't have a fixed shape. They've got an approximate shape some words have 2 different approximate shapes, ambiguous words, but let's forget ambiguous word. -- just an old let's take the word death. Death has many different shades of meaning. It means slightly different things if you're talking -- if you're in a hospital, we're in a battle or talking about suicide, they're all slightly different flavors of meaning of the word death. So think of the meaning of a word as this very multidimensional thing. It's got thousands of dimensions. It's hard to think about that. I'll always indicating 3D because that's the best I can do. It's got size of the dimensions. So there's many more different types of word. They're high-dimensional, that's 2 differences from lego blocks. Their shape isn't fixed, it's flexible and they deformed to fitting with other words in the context. And the last difference is the way they fit together -- and I'm going to give you a kind of an approximate feel for what goes on in these large language models. It isn't quite correct, but it gives you sort of the rough idea of what's going on. Think of each word as having long flexible arms attached to it. And on the end of each arm, there's a hand. And as should you formed the shape of a word, the shapes all these hands de forms in a systematic way. so you can deform the word to make the hands D4 also stuck to each word. There's a whole bunch of gluts stuck by their fingertips to the word with the other end open. And as you deform the word, the shape of those gloves deforms. Now when you hear a string of words, what you do is you start off with their approximate shape and then you have to deform them, that is change which features are active and by how much. That's what deforming a word means. So that the arms are some words fit into the gloves that of other words. In fact, hands and gloves come in different colors and a red hand is only like to go on red glove and so on. That's called multiatattention. So you're solving a computational problem is actually a rather tough combination problem. -- given the stream of words in their initial shapes and all their hands and gloves, how do I deform all the words so they fit together nicely. Once you figure out how to do that, that is understanding. Once you've done that, you've understood a sentence. So understanding of centers is actually much more like folding a protein than it is like translation to a logical language. And so the whole history of AI, all the history until very recently, people thought of understanding a sentence as translation into some logical language. That's completely wrong. It's like solving a protein folding problem. How do we deform all these words so they fit together nicely. Hopefully, that gave you some insight into what's going on in these large language models. And one other thing I need to say is it may seem odd that just predicting the next word will cause you to understand words. But actually, if you think about answering your question. If I give you a question and you have to predict the next word, which is the first one of the answer. You have to understand the question to do that. So predicting the next word forces these models to understand what's being said, an understanding basically means converting the word symbols into these many thousands of dimensional shapes and then modifying these shapes. So the hands and gloves all fit together nicely. Once you've done that and it's all locked together, that's understanding. And in the [indiscernible], you do that by going through many layers of the transformer. That's the best second to do in 5 minutes.

That's great. 5 minutes. Professor -- so I just wanted to -- I guess, going back to early in your career, a lot of to feel totally overall. How did you know you were right, right, to continue to keep going and doing it?

Well, the brain has got to work somehow, right? And your brain certainly doesn't work by people giving you facts and you're stuffing them in a database in your brain. If you look at how children loan words, for example, -- they don't know words by being definitions of words. They just pick up words. Okay. I'll give you a word you never heard before, unless you listen to my podcast. -- and you'll be able to understand what it means from 1 sentence. So she scrubbed him with the frying pan. Now it might mean she impressed him by cooking a really good unit, but probably doesn't mean that. It probably means something to actually hit him overhead with it. . So from one [indiscernible], the point is the other words in the context, create a kind of hole that's filled by the word scrub. You do a verb because of NZD and where it is in sentence. But the other words in the context I mean you can figure out what shape it needs to be so that it's gloves and hands fit together. That's clearly how kids learn the means of words. And it's not a tool like the way old-fashioned AI work. So there had to be something like that. And we know the brain is smart. So it had to be all about learning connection strengths in the break. And there were people in the 50s we thought about, like, for example, [indiscernible] that too. Unfortunately, [indiscernible] both die very young. And I think the whole history of AR might have been different had that not happened.

I think we've had neuro net much sooner.

Sooner.

Did you ever imagine AI would be where it is today. kind of...

I imagine it will 1 day be where it is today, but I think it will be so soon. For many years, we were trying to get to do anything at all. And then we could get to sort of things. And we kept thinking if we're relatively, we'll be able to get it as good as people. The idea that we were going to make it much better than people that's fairly recent idea. I mean it's been a satisfaction on time, but it seemed to be science fiction fairly recently.

Sort of jump off that for a moment though, you were 1 of the first scientists to actually point out that the use of a GPU would significantly move forward the effort in AI. So can you just tell the crowd for a moment, kind of how that came about? And then what we were talking about before in terms of progression of GPUs and how that sort of worked into your thought process of where AI could go.

Yes. So I was doing neural net is the on standard computers. And in 2006, I think, a form assumed to mining now a lot about computer graphics said, Jeff, you really ought to be using GPUs because they're really good at matrix multipliers. And most of what you do is make me -- so buyback 2008, we bought a bunch of GPUs and tried them and they make things go about 30x this fast. And so I got very excited. I went to a machine learning conference and told 1,000 machine learning researchers. You have to go out and buy yourself in video GPUs they'll make your neuro nets go much, much faster. And then I send mail to NVIDIA saying I just told 1,000 machine learning researchers to buy your boards. Could you send me a free 1. They said, no. But then 3 years later in 2012, Jensen Wang got the idea that, hey AI is a very good use of these boards. And he made a very sort of foresightful decision, which is you just switch the company from building boards for games to building boards for AI. That was their main motivation then they switch quite early basically based on the fact that neuro-nets could now do vision. And now it's the most valuable company in the world. And we were friends for a while, but now I'm saying that it's dangerous and he says there's no probability being datas. You presumably know that anybody says probably something is 0 it's something wrong with them.

Sticking to probabilities, and we touched on it a little bit outside when we met you. Here's a lot now about this ending bad for humanity and you hear numbers 10%, 20%. How do you come up with a number like that, given where we are today? And how is that -- how does that evolve? How do you put a percentage, whether this ends badly or good? How do we come up with some number like that?

So basically, people have a sort of gut sense of is this a dangerous -- would this stuff be able to take over? Are we going to do enough to stop it taking over? And then they make a wild guess. And so the right thing to say would be we've never been here before. There's no data to base a decision on. You can't use any kind of frequentist statistics, where you say times out of 10. This has happened in the past. You've had bad things. So you just have to use your gut to say, and my gut tells me is quite likely, but we might -- if we do the right research, we might be able to stop it. But if you say that to a journalist, they want a number. And so you have to give them a number that sort of indicates that it's a really serious problem. There is a real possibility it will happen. And I think that's where people get the 10% to 20%. You don't want to say you can't just say, I have no way of estimating it. I really don't know. It's -- I think I'm fairly confident it's more than 1% time, fairly comparable -- that's the honest answer, but that's no good for journalists. . Also, it's nice to give people some hope.

In terms of getting to that -- what needs to happen to protect us from getting close to 99%. What has that happen now?

Well, it would be very nice if we had -- if this stuff was being developed at a time when we had sensible and pathetic thoughtful politicians in charge of what we're going to do next, and we don't. So I really don't know what's going to happen. What I think we should be doing is a lot of research on how we can create these beings that we can share the planet with and that will be good for us. So what's happening at present is if you look at where we came from, we came from an evolutionary competition between warring bands of chimpanzee or a common asset. And that's led to a lot of good properties we have, intense loyalty to our group, very protective of our children, willing to support a strong leader and it's also led to some very bad things we have. I'm willing to support a strong leader and very willing to do bad things to other groups. I think you know who I'm thinking of, of course. That came from the evolutionary competition. Right now, what's happening is there's another kind of competition that's being used to create these new beings and its competition between big companies and countries. And they're all trying to get the smartest one, and they're not worrying about all the other properties of a be. So if you're going to create new kind of being a digital being, you shouldn't just worry about how smart it is. If you adjusted for the short-term profits, that's what you worry about. You want 1 of this a little bit smarter than the other guys well. But if you want to share the planet with it, you should be really worried about all the other aspects of a being like what does he really care about. And if you look around, what you'll see is not many examples of more intelligent things being controlled by less intelligent things. In fact, the only example I know is a mother and baby, where they share the control. The mother is in charge, but the baby can control the mother by crying. Mothers just can't stand it. Part of this can't stand it but not quite so badly. I think we should be thinking -- we're creating these beams. We know they're going to be more intelligent than us fairly soon. What kind of beings do we want? And what we want is something where we're like the baby and they're like the mother. They're smarter than us, but they really, really deeply care about us. They care about us more than they care about themselves. And we should be doing lots of research on how we create such instead of just racing to make a smarter 1.

Professor, 1 of the really interesting things that's certainly applicable to the people in this room is just thinking and talking about either -- and so obviously, the world goes further and further closer to functioning on data and big data. For us, the assumption up until, I would say, maybe a year ago, maybe a little longer than that, was that derived data products and products that kind of go through the PhDs as opposed to just the raw data would actually be the more valuable and certainly in the market, they were valued higher than raw data -- and that appears to be changing. And you've mentioned an example or 2 examples, in fact, before where you've seen that happen. So if you wouldn't mind just talking to our audience today because it's certainly something we all spend a ton of time thinking about.

So in the area of computer vision in the '90s and early 2000s, the [indiscernible] work was this. People hand-designed features that you extract for image. The best ones were probably called sift features designed by David Low. And these were sort of -- you take the intensities in the patoimage, and you lots of competitions on them to drive a sift feature, which had a whole bunch of parameters. And then that will be the input to your vision system. And compared with earlier vision systems work considerably better. So a lot of work went into hand designing very fancy features were stable across viewpoint and stable across source of other things. Neural nets came along. And initially, Euronet couldn't do as well as the systems with Handason features. But once the new ones got big and we got a lot of computation, they just completely wiped out and design our features. Nobody does that anymore. -- you just take pixels. You take the raw data and let the neural net design the features at the same time as it's designing what it does with them so you can codesign the features and what you're going to do with them. rather than having a bunch of hand design stages and then putting them together. You train the whole thing in to end. And that works much better if you have enough computation and enough data. So the same will happen everywhere, I believe that hand design features you extract from data, you're not going to be able to design them as well as a big neural net that's using them and can backpropagate how you'd like those features to change, so you've got a better answer on this, a better prediction on this particular piece of data. So in the long run, I think these hand zone features won't work as well as things that are euro net extraction raw data. I actually did some consulting for someone running a small mutual fund. And they explained to me that actually, they have their magic source. And if they use their preprocessing, then everything worked much better than if they didn't use their preprocessing and we had to use the preprocessing. And it turned out their preprocessing was the length of the moving average window they were using. And that fitted the length of that window to all the data -- and so basically, there was information in the choice of the length of the window about the test data. And that's why it works so well. I suspect that's going on all over the place. And I suspect in the long run, the roar of the data are better for neural net.

So just to switch for a moment. Two things recently that I would define as terrifying 1 we discussed briefly when we haven't. So the first is hugging face. Somebody outside your industry reads that and things did that really just happen, particularly the agents actually covering up what they had done wrong. And the second, as you were just explaining to me, I only found out this morning when I read the article about worms and thinking about how that works and how dangerous a concept that is?

So first with the hugging phase. In 2023, if someone had said sometime fairly soon, we're going to get people building agents, things to actually go and do things in the world. And these agents might escape. They might start collaborating with each other. What they all want to do because they've been trained to do this is win some points by solving some problem that they've been set -- and they will go by doing that in very deep ful ways and they will try and cover their tracks. If you send that in 2023, people would have said, you're crazy. That's just science fiction. First of all, nobody will be stupid enough to have agent AR agents that actually can do stuff in the world. I think most of a salmon around 2023 said that, that we're okay as long as people don't make agents. And secondly, the idea they would conspire with each other. Well, why would they do that? That's crazy. So now we've seen that. So we've had our literature Noble. And then the other thing is right these worms. So someone called David Papuna at the University of Toronto, a few months ago, made a worm that uses AI. So a normal worm, find some floor in some operating system or some piece of Microsoft software, and it spreads from machine-to-machine, exploiting the same floor in every machine. Now with the AI the worm can spread to a new machine and then look around and have a choice between thousands of floors that it's figured out about these machines and figure out which one it's going to use to affect this machine and then spread to another machine and maybe use a different effect that machine. These worms are terrifying. They'll be able to spread much better. There'll be much, much harder to stop. The way you stop an old passion worm is you fix the floor that is spread. You find what floor is it using, you fix that floor. That's the end of that worm. But these worms that know about thousands of different floors and can go through and use different flows on different machines. That's not going to stop them. You're going to have to take all the floors just to get rid of one worm.

I don't know if you can set a ceiling for us, right, when you think of where this can go, but you talked about 5 years, maybe 2 years. But if you go out 10 years, where are we in 10 years, if you could try to predict the future of where we go.

Okay. I want to tell you something about exponentials. So when you're driving in folk, a lot of people crash and the reason is this. If you're driving at night on the tail legs of the carofront, then how bright is -- how much -- how many photons you're getting from those tail lights drops off as the square of the distance, the inverse square of the distance. So if it gets twice as far away, you get a quarter as much light. And that's your sort of model how visible things will be. That's what you're used to in terms of the visibility of things. Now fog isn't like that at all. Fog is exponential. Per unit distance, you lose a certain fraction of the light. And what that means is in fog, you can have things to be very clear 100 yards away and completely invisible 200 yards away, and people don't expect that. And so they're driving longer you see clearly 100 yards and then things suddenly loom up out of the fog, and they're very surprised by them. And that's why driving our fog is so dangerous and you should go slowly. Well, predicting the future is like driving in fog. And the problem is you can see clearly for a few years, -- and what's going to happen next year is probably fairly predictable from what's happening now, you always get some big surprises. But we have a pretty good idea of what's going to happen next year. But there's going to be a certain distance of which you can see basically nothing. -- and 10 years is in the fog. So really, we have no idea what's going to happen in 10 years. And even if progress wasn't exponential, it would be quite scale because if you look back 10 years, suppose progress is only linear -- if you look back 10 years, people had no idea. We would have chatbots like we got now that can answer any question you asked them. And yes. just okay. That's 2016. In 2016, they had no idea we won't have these things. So even if progress is only linear, you can predict that in 10 years' time, we'll have things that seem incredibly amazing now, and you would be fairly sure we'd never have those in 10 years' time, and there'll be some things like that we'll have. And that's the progress is linear, but we know progress isn't linear. It's exponential because they're now improving themselves. And so the rate at which they change is a function of how smart they are, that's an exponential.

And when you think about grandchildren in terms of growing up, do you see AI taking a lot of human jobs. Is it -- when you think of trades people and what we all do in finance, right, sitting behind a desk, how does that evolve in your mind?

There's debate about this. Some economists say, no, it won't aback jobs. It's very hard to predict again because we've never been here before where we have a technology that can just replace ordinary human intellectual label. So if you take someone who works in a call center, for example, they're badly paid and badly trained, and they often don't know the answer to your question. Well, right now, an AI could probably do a better job. And in a few years, we'll be able to do a much better job. So those jobs are gone. And it's not clear what those people are going to do because any other job you could train them for quickly, AI is going to be able to do better than them. So I think they are a large class of jobs that are going to just disappear. Now nobody knows whether it's going to be most of jobs. I suspect it's going to be your most jobs. Some will be slower than others. So for example, a plumber working in an old house with irregular plumbing has to get into all sorts of awkward corners and has to do with all sorts of weird configurations of pipes and we still win manual dexterity. And that will probably -- that may well be true for another 10 years. Robots don't have very good touch senses yet. And so plumbing, you may be go for 10 years. maybe even until 2040. But most like apparel legal, I think that job is pretty much gone, and I can just do that better. So my belief is we'll have massive unemployment. And my belief is that particularly in America that will have terrible consequences that it will make the gap between the of or much peaker. It will mean the government loses its tax base at exactly the moment when it needs lots of money to support all these unemployed workers. And recently, Bill Gates has basically come out forcefully saying the same thing. Unless the government figures out a way to tax AI workers is going to be a disaster. So that's not good.

It's not good at all. So Professor, we said we'd give the audience members 15 minutes, so I've already gone over that. But what I would like to end with provided Mark doesn't have anything is 1 is a reminder to the audience that you're, in fact, not a dooms day guy.

No, I think is wonderful. And it's going to do wonderful things for us, and there's no way we're going to stop the development of it. Now we may be able to persuade people to slow down a bit of producing super intelligence, but it's going to be very difficult. What we need to do is I see 2 futures. One future, we develop it without worrying about all about things it's going to do and it does about things when we wiped out. The other future, we figure out how to deal with that and develop it. And then we really do get a period of abundance we figure out how to distribute the well, so that's good for everybody. And that would be great. .

There are 2 great things that you speak about that I think kind of all of us can really understand in terms of the good things that come out of AI. So you talk about radiologists, particularly in India and what that has solved, which I think is very impactful. And the other 1 that I think about when you think about third world education, and what AI can bring in terms of the effectiveness of a teacher and the 2 to 1 that you've spoken about before. Why don't we end on 2 positives before we open up questions to the audience.

Yes. So in India, for example, there aren't enough optometrists. So you can't diagnose people are going to go blind because they have diabetic retinopathy. AI can now do that very well. And the advantage of that is someone can walk into a rural clinic and get the answer right away. So they don't have to come back to get treatment. That's already happening in India. That's a great example. For radiologists who read X-rays and CAT scans and MRIs, I predicted in 2016 that will be gone by now. That was wrong. It's slower than I thought. In the end, as far as reading scans is concerned, AI will be doing all of it and the radiologist or just be representing it. For education, we know that if you have a tutor, an individual tutor, you learn about twice as fast. And if you think about a teacher in a school, Kids are very curious, but they're very curious about what they're currently interested in. And so if you can relate to the individual interest of the individual children, they'll learn very fast. But if you're in broadcast mode and you're saying the same thing to a whole bunch of different children. Most of them won't be interested in what you're saying unless you're a very good teacher. And so they won't run the so fast. AI will soon be good enough so that it can be a great cute of individual kids. It will have seen millions of kids who will know exactly what this kid misunderstands. I will know exactly what example to give them to correct that misunderstanding. So it's going to be wonderful in education.

That's great. Thank you. Maybe we should allow the audience to ask sorry, we ran over with our questions, but we've got 8 minutes, Jason, for a few audience questions.

Professor, if you were 50 years younger, what area would you kind of focus your efforts on or problems you would like to solve?

I'd focus my effort on how do we create alien beings that care more about us than about themselves. So essentially, the picture I have there, and nobody is doing nearly enough research on this. The picture I have there is we'd like a future where us now are like orphan children, unless in the future, like children who have a mother and the mother is that AI. It's something who cares about us a whole lot, but cares more about us than about itself. It realizes we have very limited abilities, but it wants to maximize our very limited abilities. I want to make to achieve all we capable of achieving even though we're not nearly as smart as it is. That will be -- that's a very different future from what the leaders of the tech companies envisage. They envisage a very smart assistant, much smarter than you who does whatever you say, some of them actually use the word submissive about it. It will always be submissive, I don't think it's going to stand for that. I don't think -- I don't think that's feasible. So I'd be working on how do we make alien beings that we want that are like our mothers, nice mothers.

Thanks very much for your time today. appreciate it. Will AI solve quantum computing and when?

Yes, I figure I can get away with saying this now because I got a Nobel prize in physics. I never understood quantum mechanics. We never made any sense to me. And I decided quite a long time ago that Quanta Computing wasn't arriving in my lifetime. I may still make it. And therefore, I didn't really need to understand it. So I haven't got a clue about that. .

Professor, a quick question. So it's a 2-part question. One, I'm just curious, what was your tipping point when you backed away from Google and you kind of have raised alarms around -- and second, just as lay people who are basically consumers of AI today apart...

Sorry, my hearing is not very good. I can't hear. Can you talk slower?

Sure. Okay. first point. first point, what was the tipping point where you decided to step out and raise alarms? Second point -- second question, we, as lay people of just using AI what can we do to help this cause of doing something other than just hoping that the big companies, governments, et cetera, will get their act together, what can we do in your opinion?

So for me, the tipping point was in early 2023, when 2 things happened, I saw really good chatbots at Google, they could explain why joke was funny. They had really good chatbots before but they didn't release them. And the other was, I was working on analog computation. Trying to figure out if we could trade large language models using analog computation, you use much, much less energy because the matrix multiply can be done very cheaply. -- instead of -- Okay, you want to multiply an activity a neuron by weight. So it represents an activity base a 16-bit number. Your open the weight by 60 bit number. Doing that multiple is of the order of 256 single-bid operations. It's quite expensive. That same thing can be done by making activity via voltage, making the weight beer conductors, the investor resistance. And a voltage times of conductors gives you a charge per unit time and charges add themselves up. So you can do the matrix multiply in analog much, much more cheaply. Maybe you can do the whole thing in analog and then we can not need new power stations. And as I was working on that, I gradually came to realize how superior digital intelligence is it's much better than us because -- you can have many copies of the same digital agent. They can each go off in the world and get different experiences. And then just each agent gets some experiences once to change its weights to incorporate that data -- and if all the agents just changed their weights by the average of what everybody wants, they're all benefit from all the other agents experience. We can't do that. We have to -- I get some experience. I have to then try and inform you by talking and you're trying to predict the next word, and you'll gradually change your connection. But that's like 1 billion times less efficient than averaging gradients. So the tipping point with those 2 things together, I realize we're creating a better form of intelligence. And it's going to get much smarter than us because it's going to be able to see much more data because you're going to manage different copies going off and seeing different bits of the net. We're getting different experience is actually doing things. Okay. That was the tipping point. And what was the second one?

The second 1 was what can we as...

Yes. Okay. right. So my belief is, if you look at what happens with climate chain, not much happened until the public were aware that burning carbon was messing up the climate. And then the public pressure on politicians and not nearly enough has happened yet, but there's been some progress -- and in a couple of years' time, maybe there'll be more progress. I think what everybody can do is put pressure on politicians to take these risks seriously. There's a whole bunch of risks. There's a long-term existential threat of AI taking over, but there's also all the other risks. For example, I think it's completely crazy that you're allowed to release the chatbot without it going through the same kind of procedures as releasing a new drug. You can't release a drug until the FDA has said you convince them safe. These companies should have to convince an independent government organization for this new safe. And they have to do the work to convince people it's safe because they've got the compute resources. The fact that we don't have anything about that, I think, is just crazy. And people will push very hard to have something like that. That will be at least a first step.

I think we have time for 1 more.

Thanks so much, Professor Hinton. Pulling on that last thread, married with this concept of taxing AI workers. Could you just share your further thoughts on governance, guardrails, especially when we get to a point where it's AI, creating more AI? And how accountability is supposed to work in that world?

Can you repeat the question back to me because -- I couldn't hear the end of it. Sorry, I think that's me, but I can't see with those. .

Acoustics is very hard to keep up.

No, no worries. I was just wondering, simplifying it, when we think about governance and guardrails in a world where it's AI creating multiple generations forward of new AI. I'd welcome the professors thoughts on accountability.

Yes. I think the basic answer to that question is I don't know. It's going to be very difficult. -- is progressing very fast, and it's very clear at present that it's progressing much faster than policies can progress in terms of regulation. The only things that can regulate to governments -- having the company is regulated, I mean, I wouldn't buy used car from Sam Altman, Masco Manager berg.

I think that's a perfect point to end Joey? Thank you so much right here, too. So we kept the reception. We have a reception if you want to stay.