2025 03 27 AI Roundtable Panel 01

Summary

The SEC's AI Roundtable opening panel explored the integration of AI in the financial industry, moderated by Rob Hegerty. The panel comprised experts from Citadel Securities, JP Morgan Chase, NASDAQ, American University, and BlackRock. Discussions revolved around the definitions, benefits, and challenges of AI in finance. Panelists shared insights into AI's operational, investment, and productivity applications while acknowledging inherent risks and the need for robust governance. They emphasized AI's rapidly evolving nature and advocated for ongoing dialogues between regulators and industry participants.

Highlights

• Panelists discussed AI's impact on operational efficiency, investment processes, and market transparency. 📊
• They emphasized the need for clear AI definitions and flexible regulatory frameworks. 📜
• AI's potential in financial services includes productivity enhancement and cost savings, but risks include scamming and biases. ⚙️
• Panelists agreed that cultural resistance and data complexity are significant barriers to AI adoption. 🚧
• The future of AI in finance lies in dynamic regulation and a deeper understanding of its operational impacts. 🔮

Key Takeaways

• The SEC held an AI Roundtable focused on AI's role in the financial sector, highlighting evolving uses and regulatory challenges. 📈
• Panelists stressed the importance of defining AI while allowing flexibility for its evolving nature. 🤖
• AI in finance aims to enhance operational efficiency, investor transparency, and market dynamics. 💹
• The discussion revealed that AI's primary barriers are culture, risk understanding, and evolving technology. ⛔
• Regulatory perspectives focused on balancing innovation and protection, urging continued dialogue. 🗨️

Overview

The U.S. Securities and Exchange Commission (SEC) initiated an AI Roundtable to deliberate on AI's growing role in the financial industry. The opening panel was moderated by Rob Hegerty, with participation from key industry leaders such as Citadel Securities, JP Morgan Chase, NASDAQ, American University, and BlackRock. Each panelist shared their unique insights into how AI technologies, ranging from machine learning to generative AI, are shaping financial practices.

Key conversations centered around the benefits of AI, including improved operational efficiency and enhanced transparency, as well as its challenges, particularly the complexity of data management and ensuring ethical use. Panelists debated the necessity of defining AI clearly while maintaining flexibility to accommodate its rapid advancement. Discussions also touched on AI's impact on investment products, operational tasks, and systemic risks.

Looking ahead, the panel stressed the importance of ongoing collaboration between regulators and the industry to navigate AI's evolving landscape. They highlighted the dual need for innovation and stringent regulatory oversight to protect investors while fostering technological growth. This dialogue underscored the critical role AI will continue to play in transforming financial markets, emphasizing both the opportunities and the mandate for careful governance.

Chapters

• 00:00 - 01:30: Introduction and Panel Overview The initial transcript for the chapter titled 'Introduction and Panel Overview' begins with a greeting and introduction for the opening panel of the SEC's AI round table. The session is titled 'The Benefits, Costs, and Uses of AI in the Financial Industry' and Rob Hegerty, the moderator and a senior policy adviser in trading and markets at the SEC, greets the audience. He humorously explains his appearance, noting his hat and unintentional nod to Boston roots, due to a recent medical procedure.
• 01:30 - 03:30: Panelist Introductions The chapter titled "Panelist Introductions" begins with an acknowledgment of the commission staff and commissioners for their relentless efforts. It emphasizes the recent efforts made over the last few months to adapt and incorporate AI into the commission's mission, which includes investor protection, capital formation, and maintaining fairness.
• 03:30 - 06:30: Panel Objectives and Recent AI Developments The chapter titled 'Panel Objectives and Recent AI Developments' opens with an acknowledgment of the remarkable advancements in maintaining orderly markets. The speaker expresses the honor in initiating the first of many anticipated panels and discussions involving the commission, the industry, and the public about AI's evolution in financial markets. A distinguished group of panelists from diverse sectors of financial markets is introduced, highlighting the variety in firm types, functions, and backgrounds, as well as the wide range of investors, clients, and participants they represent.
• 06:30 - 17:00: Defining Artificial Intelligence The chapter "Defining Artificial Intelligence" begins with an introduction and brief background of the panelists, starting with Greg Berman, who works at Citadel Securities in the legal and compliance department. He leads a quantitative shop focused on applying data and analytics to inform on rules and regulations. Berman has been with Citadel Securities for over eight years.
• 17:00 - 35:00: Use Cases of AI in the Industry The chapter discusses the uses of AI in the industry, featuring insights from professionals at major financial and regulatory institutions. It highlights the role of AI in governance, research, and engineering within organizations like the SEC, J.P. Morgan Chase, and NASDAQ. Each organization has dedicated teams and systems for the deployment and maintenance of AI models to enhance their operations and decision-making processes.
• 35:00 - 43:00: Cost Considerations and Returns on AI Investments The chapter discusses the financial aspects related to investing in AI, focusing on cost considerations and returns. It starts with strategies to manage markets and enhance productivity and efficiency within businesses. Hiliary Allen, a law professor specialized in financial regulation and new technologies, introduces her perspective, mentioning her book 'Driverless Finance: FinTech's Impact on Financial Stability.' Daniel Petero from BlackRock shares insights on driving AI adoption by collaborating with the COO and deputy COO, stressing the importance of integrating AI to achieve business goals.
• 43:00 - 54:00: Challenges and Barriers to AI Adoption The chapter outlines the primary aims of a panel discussion on AI adoption, emphasizing the importance of setting a foundational understanding for further discussions throughout the day. The initial focus is on learning, engaging in meaningful debates, and exchanging ideas with other panelists and industry experts. The setup is attributed to the excellent groundwork laid by the commissioners, indicating a structured environment conducive to exploring the challenges and barriers in adopting AI technologies.
• 54:00 - 58:00: AI and Investor Benefits The chapter 'AI and Investor Benefits' features a panel discussion focusing on the role of AI in the industry. The main purpose is to spark discussions, raise pertinent questions, and exchange views to uncover insights into AI's implications for investors. Panelists are encouraged to interact, debate, and present diverse perspectives to enrich the conversation. The chapter underscores the importance of dialogue in exploring AI's potential and impact on investment strategies.
• 58:00 - 62:00: Regulatory Considerations for AI The chapter titled 'Regulatory Considerations for AI' begins with an introduction to recent developments in AI technology. It highlights Nvidia's launch of the DGX Spark, the world's smallest AI computer with 128 GB of memory, and the announcement by the Chinese firm BYU of two new models: Ernie 4.5, which surpasses GPT 4.5 in performance, and Ernie X1. These technological advancements are significant as they may influence regulatory frameworks and considerations in the AI sector due to their cutting-edge features and capabilities.
• 62:00 - 69:00: Future Developments and Conclusion This chapter discusses recent advancements and future potentials in AI technology. It highlights significant events in the AI field, including the introduction of a more affordable reasoning model, a groundbreaking medical AI application in the UK, and an impressive display of robotics technology by Boston Dynamics. These developments illustrate the rapid progress in AI and robotics, suggesting an exciting trajectory for future innovations. The chapter concludes with a call to witness the remarkable capabilities demonstrated by these technologies.

2025 03 27 AI Roundtable Panel 01 Transcription

• 00:00 - 00:30 good morning everybody and welcome to the opening panel of the SEC's AI round table the benefits costs and uses of AI in the financial industry I'm your moderator Rob Hegerty I'm a senior policy adviser in trading and markets here at the SEC Um first let me say that I'm not making a fashion statement up here with my hat or leaning into my Boston roots Uh I did have a poorly timed medical thing done last week So uh
• 00:30 - 01:00 it's uh it's staying away from the cameras Um let me uh let me start by first thanking my fellow commission staff and the commissioners for their tireless efforts over the last several years and particularly recently just in the last several months uh the work that's been done uh to adopt to adapt to to incorporate AI into the commission's mission of investor protection capital formation and maintaining fair and
• 01:00 - 01:30 orderly markets has been astounding Uh it's been a pleasure to be a part of I'm honored to kick off the first of what's expected to be many panels and discussions between the commission the industry at large and the public on the evolution of AI in our financial markets I'm also delighted to have such a distinguished group of panelists with us today representing a diverse cross-section of financial markets both in the type of firm the functions they represent the backgrounds of each person as well as the wider array of investors clients and participants they serve So
• 01:30 - 02:00 with each panelist please introduce yourself by name title and organization and your area of focus Thank you Rob Uh my name is Greg Berman Um as says on the plaque I'm at Citadel Securities I'm in the legal and compliance department but I run a quantitative shop supporting uh rules regulations anything that you can apply data and analytics to give it an informed thought about I've been at Citadel Securities for about eight and a half years And prior to that I was
• 02:00 - 02:30 actually right here at the SEC well not necessarily in this exact spot but here at the SEC for about five and a half years Thank you All right Mike Kelly J Morgan Chase I've been there for 21 years and I run our AI governance and enablement team Hey folks Doug Hamilton I head up AI research and engineering at NASDAQ where I'm responsible for the build deployment and uh maintenance of various models and AI systems in production that both help uh you know drive our uh drive our
• 02:30 - 03:00 business forward uh manage our markets and of course uh drive productivity and efficiency into the business itself Hi I'm Hillilary Allen I'm a law professor at American University up there somewhere Um I specialize in financial regulation and new technologies and I'm the author of the book Driverless Finance: FinTech's Impact on Financial Stability Hi everyone Uh Daniel Petero joining from BlackRock I work with our COO our deputy COO and business partners across BlackRock uh to drive AI adoption across
• 03:00 - 03:30 the firm Great Thanks everybody So let me just lay out the objectives of the panel Um they're manifold actually Uh the first thing is we want to set the table for the rest of the day here You're going to hear from uh many other panelists and industry experts and our job here is to sort of level set and I think the commissioners did a great job of setting up what we're here to do today which is learn Um it's also to uh engage in a spirit of discussion and debate uh among
• 03:30 - 04:00 the panelists on how AI is used in the industry Um the goal today is not to provide all of the answers but it is to get the questions out there It is to get the topics into discussion and provide answers where we can Um I encourage all the panelists to uh engage with each other to interact uh to debate to provide alternative views and ask each other questions Uh it makes things more interesting if everybody's uh in interactive So I've encouraged the panelists to do that
• 04:00 - 04:30 So with those remarks and um just to get the ball rolling I'm actually going to list uh just a few recent developments in AI Um and these uh these are just uh very recent So first Nvidia revealed the world's smallest AI computer um the DGX Spark with 128 GB of memory Chinese firm BYU announced two new models Ernie 4.5 which outperforms GPT 4.5 Ernie X1 which
• 04:30 - 05:00 matches DeepSk's R1 model at half the price further fueling the race and reasoning models Uh third thing was a UK hospital was the first first in the world to use an AI tool with a camera attached to a mobile phone to make clinical decisions on skin cancer diagnosis The fourth thing is Boston Dynamics released a video of their Atlas robot showing it crawling cartwheeling and break dancing You really should check out that video Um and lastly
• 05:00 - 05:30 researchers released Arvar weather significantly enhancing weather forecasting speed and accuracy with greatly reduced computing power So maybe now meteorologists won't always be wrong anymore The most interesting thing about those developments and I said they were recent that was all just in the last week So if there's any doubt about how fast this space is moving um that's all you need to look at Those are very significant developments just in the last week in this in this field So with that we're going to get
• 05:30 - 06:00 going on the panel Um we did hear Commissioner Krenshaw to talk about getting a level set on the definition of AI And so having said that we're not here to get to all the answers I thought we'd start with maybe the hardest question of all How do we and should we define AI and does it even matter i would say it does but let's talk about the uh the definition of AI Um why don't we start with Hillary
• 06:00 - 06:30 sure Well I mean this is a challenging thing to do in part because of all the hype around AI and you know the SEC has been great about pursuing AI washing cases people saying something is AI when it's not I think the most helpful framing I've seen actually came from an interview in the Financial Times with Ted Shiang and he said the term artificial intelligence is really misleading because the things that these models do they're not intelligent in the way a human is intelligent They don't reason They don't understand They're not
• 06:30 - 07:00 going for accuracy because they don't understand really what accuracy is So what we have instead is applied statistics And I found that to be a very useful framing What we have are algorithms looking for statistical patterns in data sets and then using those patterns to perform an assigned task Now the machine learning versions of these tools which have been in use for a long time um they are intended to perform tasks like predictions or
• 07:00 - 07:30 classifications and then you get to I think 2022 and then you have chat GPT and everyone's very excited in about generative AI and it sort of retroactively charges interest in the machine learning stuff as well but there's often a lot of confusion about where is that what dividing line between machine learning learning and generative AI And so generative AI works in the same way in many respects It looks for patterns in data sets in order to use to
• 07:30 - 08:00 perform a an assigned task The difference really is that the data sets are just so much bigger that it can sort of learn more things from them and can use those to generate new content be it video um you know text which is probably very relevant for this audience but also there's talk of music and things like that Um so I think it's important to realize that a lot of the the buzz that we have around AI now is actually hearkening back to machine learning models that have been in in um place for
• 08:00 - 08:30 some time I think they're very much worthy of the SEC's attention Um but the Genai hype is sort of what's what's carrying us along in many respects Great Thanks Hillary Dan can uh can we get your view uh and uh how you define AI and uh the importance of it Sure And first off just a thank you to Rob for chairing Thank you to my fellow panelists Thank you to everyone that's joining into the SEC for hosting today's event Looking forward to this conversation that we're going to have I would go back to some of the comments
• 08:30 - 09:00 that Professor Allen made and some of the comments that the prof the acting chairman made as well AI isn't new as an academic field It's been around since the 1950s In practice machine learning has helped us through a number of different events in the financial markets over recent history Looking forward the idea of building out a common taxonomy defining what is what isn't generative AI AI would be helpful uh could be helpful in assigning and
• 09:00 - 09:30 defining clear principles to help guide us forward What I would suggest that we think about is making sure that such definitions have sufficient flexibility such that we can adapt and evolve to changing capabilities that are moving at a rapid pace within this space Okay great Thanks Dan Greg get your shot at uh defining AI and how you look at it Sure Um thanks Rob Um as I mentioned
• 09:30 - 10:00 before I was at Sidel Securities I was here at at the SEC Um I joined in 2009 In 2010 the SEC had a concept release that they issued on equity markets And in that they noted this new thing called highfrequency trading There's this brand new thing called high frequency trading and I sat either some combination of here or out in the audience or in the multi-purpose room panel after panel how do we define highfrequency trading so
• 10:00 - 10:30 the question is to in my mind is not what the definition of AI is but um does it matter do we have to define this and this is the exact same question that the commission not only in the US but also security commissions everywhere really struggled with because they knew knew that high frequency trading was a new thing They had to figure out how do we regulate this what do we do about it and how do we define it and maybe just sort of cut to the chase In the end it was
• 10:30 - 11:00 determined that not only is it not possible to really define or create both a flexible definition that also does not thwart innovation but it's actually completely unnecessary And the reason why it was unnecessary is because the commission in my opinion did some pretty brilliant things at that time to try to get their heads around how do we encompass this So for example there was a concern that if you had blackbox algorithms that were just running a muck etc that you might be able to send tons
• 11:00 - 11:30 and tons of orders to the market at the same time So what did the commission do commission said you know we actually have a regulation here you have to be a broker dealer to send things to the market So they created a market access rule and that rule basically says that the broker dealer is 100% responsible for every single order that goes out their door and goes to an exchange Um if you violate that you're in trouble If you violate that because you did a a fat
• 11:30 - 12:00 finger and you put the wrong keyboard that's a human error That's a problem If you violate that because you used an algorithm and you say "I don't know what was in the algorithm." Well guess what that's your problem And if you violate it because you gave your pipes to somebody else to use then that's a problem That rule by itself has proven incredibly resilient and incred incredibly effective at preventing the types of things that people were concerned about but it also prevented a lot of the things that people hadn't even thought of And you will not find
• 12:00 - 12:30 any definition of HFT in any of those rules And I can speak we have had entire panels on this topic but I think the exact same lessons that were learned 15 years ago at this stage are probably just as applicable today as we struggle with what the concept of AI is That's great Greg Thanks very much A lot to think about in terms of comparing and contrasting to prior technological evolutions too And I think we can draw a lot on that Um Mike how do you look at
• 12:30 - 13:00 it in defining AI and and does it matter yeah also a note of thanks Um and I pick up on some of the points that Greg and some others made around how fast it's changing Um you talked about just the technology landscape that's changing The definition will change quite frequently So I agree that the definition itself isn't overly important You need to have one and we'll have one There's two other things that I would put onto that though One is the risk lens um and being able to kind of quantify what makes something risky and and then you know be able to focus in on those areas whether it's the
• 13:00 - 13:30 commission or organizations And then the other that I'm finding very important is the inventory aspect of that These aren't necessarily new to a lot of the focus points You've got a model you have a lot of data you have some technology you have some business process and tying all that together with a definition with some risk aspect with the inventory to know what you have and whether they're using that is is even more important than getting to that fine definition Okay great Doug All right so I'm going to start off with excuse me how I would not define AI and I how I would not
• 13:30 - 14:00 define AI is however the EU has done it which is so broad and all-encompassing it might include human thought which is really fun Um it's kind of meaningless and uh my only thing I like about is it guarantees American leadership in this space for the foreseeable future Um when I think about AI I think a lot in the ways that Mike and Greg uh describe that that is through this risk lens And many of the risks that emerge that are particularly unique to AI have to do with the way that it is trained and designed in that artificial intelligence
• 14:00 - 14:30 tests many more strategies many of which are idiosyncratic um for achieving some goal than you could programming something by hand As a result of the scale of decisions the scale of data and the scale of strategies being considered you find interesting ways to solve things And some of those interesting ways maybe don't really work when rubber meets the road And so when I think about that risk lens to me artificial intelligence is
• 14:30 - 15:00 really defined in the modern sense as any program or function I'm give this an attempt and I'll just preface this by saying that this has been a problem ever since you know the 1950s when I think it was Minsky described AI as anything a computer does that we don't typically think of it as doing Um but back to the matter at hand AI for me in the modern sense is any program or function which is designed by another program containing a training loop A training
• 15:00 - 15:30 loop which is defined with four elements data ingest a learning structure a loss function as a reference point for accuracy and lastly a convergence criteria And that at the end of the day you kind of throw away that code that designs the program and you just keep the program it designs So if that's what you're looking at to me that's AI because that's something where the person wrote the code the code wrote the program and the program is the thing you care about Um and all of the kind of
• 15:30 - 16:00 idiosyncratic uh ele strategies tested various uh nuances or novelties in the data all gets picked up by that design program and are things that we don't usually have to deal with in just writing you know a T-wap or VWOP by hand So that's my attempt at it Okay Okay I'm going to throw an ad hoc question out here to the panel around this definition one because I I do think that striking a balance of understanding um the evolution of a new
• 16:00 - 16:30 technology and I'm I'm going to draw on Greg's comparison to high frequency technology I sat on the CFTC's subcommittee for highfrequency trading way back in the day and the job the first job was to define it and the definitions while I think at the end of the day they uh they were less important maybe even not important the exercise of defining it was actually very important Is there anything that any of your firms
• 16:30 - 17:00 are doing that's actually going through that definitional process in order to differentiate between the types because in my view artificial intelligence has taken on a marketing uh view of it right we those of us in the industry know it means machine learning deep learning generative AI large language models there are subgroups to AI that in my view I think we're just going to keep tacking on to are there exercises within your firms that are that are about defining AI and how it differs
• 17:00 - 17:30 from how you knew it before launch of large language models I could jump in So I I I do think that while I'm critical of a precise definition in terms of thinking about how one would apply a rule in a regulation because then the definition really really matters um that classifying different constructs to try to understand how you want to make you might want to have policies and procedures around them does become
• 17:30 - 18:00 important So if you include in the construct of AI a machine learning algorithm you're going to have to go back to the late 1500s and say everything before the to the 1500s was AI because there were constructs that have been used for many many many hundreds of years Just the plain vanilla linear regression would be would fall under that definition So people who are responsible at a broker dealer or an exchange wherever for writing models they are responsible for writing the
• 18:00 - 18:30 model and understanding how it operates in all circumstances whether or not they use machine learning or not What changed is basically with chat GPT um coming on the scene and generative AI creative video etc that creates a class of basically a class of of new types of applications and those applications um regardless of the technology creates some very interesting new things that people have to think about When you use a machine learning technique you took your data you put it in and you were
• 18:30 - 19:00 done with that If you use chat GPT there's nothing inherently wrong with with making use of that But you know there's privacy rules when you're at a broker dealer about using information And if you take information and you put it into a regression model in Excel something as simple as possible it's still within your boundaries Um people might not realize that if you type in queries into an online platform it now knows something about you that it did not know before And other people may then learn something about you that you didn't necessarily want them to use
• 19:00 - 19:30 Nothing to do with AI nothing to do with the technology It has to do with way it's been deployed So as things are used you want to identify those types of risks and those edge cases which are not particular to AI but are particular to how they're deployed So I think that at NASDAQ when we think about how we define the different sorts of AI and I know that we are in a room with a lot of attorneys so I'm going to paraphrase a court case We kind of take a I know it when I see it approach Um
• 19:30 - 20:00 and uh a lot of that falls into how companies can or maybe even should think about their own internal AI governance structures Um to Greg's point we don't want to uh call linear regression AI because we kind of know that that's not really what we mean when we say AI We usually mean there's something about its nonlinearity something about scale something about its scope Um so certainly we think about it in terms of four very classic types of artificial
• 20:00 - 20:30 intelligence right is this uh is this a predictive regressive system is this a discriminative classification system is it a uh generative system uh you know a joint learning generative system or is this a reinforcement learning system and those are the ways that we typically think about it because each one of those is very unique um testing standards that have emerged in both academic literature and in industry over the last 10 years and and uh what's actually particularly remarkable about the generative AI
• 20:30 - 21:00 systems that are you know multimodal and semantic is even over just the last two years uh from white papers from publications uh and from various uh academics who are now investigating how these systems work there is a set of standards that has emerged as to how to test even those very very new sorts of systems In fact reasoning systems today which reasoning systems today which have only existed for a few months have a wide range uh wide set of literature
• 21:00 - 21:30 that are really defining uh very good standards very good benchmarks um that firms can use if firms can use and you know regulators can reference when seeing if AI systems are uh appropriately tested uh and safe for consumption by uh by end users I love it So J Morgan to answer your question Rob we're definitely doing a lot of efforts around the definition We kind of have to Doug mentioned the EU AI act um that being quite new we didn't have inventories of categorizations of
• 21:30 - 22:00 things to just go filter and trigger and then go address those And so we had to go on that search to understand what we have and apply definitions But it does come back to the the sort of inventory piece as well because when you can apply that across the existing things that we've had for a long time some of your new gen AI things then you start to get into the vendor aspects and you've got these very large providers that do a little bit of everything You've got your cloud services your model providers Now there's a Gentic and it starts to get complex really fast So we're trying to break those down into its simplest parts
• 22:00 - 22:30 a little bit to like Doug said you know it when you see it Yeah I think it's interesting this there's clearly work to be done Uh I like the way so when Doug was talking about this just in the evolution right that we started talking about his me machine learning and there was supervised machine learning unsupervised machine learning and reinforcement learning and it it's interesting if I get this right Doug that you split out reinforcement learning as sort of its own now what's the logic behind that yeah so the logic behind that is twofold
• 22:30 - 23:00 uh one is reinforcement learning works a little different from traditional supervised learning in um how it learns Uh it employs much more of what's called an explore exploit paradigm in how it learns its kind of decision structure Um and by the way super important thing explore exploit it's awesome If you're not doing it in your own life highly recommend it's it makes it much more interesting Um but the other aspect of it is that uh if you look at say a
• 23:00 - 23:30 traditional classification model supervised classification model like XG like an ensemble boosted model most of them have basically the same loss function that is the same thing they're trying to optimize for whereas reinforcement learning functions or reinforcement learning systems are trying to optimize or learn for pretty novel stuff and because they're trying to learn for pretty novel stuff Um they require a little more care a little more attention uh a little more oversight
• 23:30 - 24:00 frankly from from um whoever is governing them And you know that's why we split them out And I think you know we saw that actually a lot when we filed for dynamic mellow which is we put out this 30-page white paper that went in excruciating detail about how the system was designed what the loss function was what what the inputs were what all the various um you know what the what tail risks if any we saw were how we were ailarating those and all these things become very important for reinforcement learning which is making this like
• 24:00 - 24:30 explicit decision versus for uh these classification systems which more make which typically make inputs that are then used by a person or another system to make a decision Can I can I just just uh I don't want to push back on that but just at least to reflect on on how does that so this is a conference at the SEC why why are we doing this at the SEC I think we heard from from the chair and the commissioners um there's rules and regulations about how to use this I
• 24:30 - 25:00 think an important classification a critical classification is how do we define constructs that are useful for business that are useful for AI development that are useful for firms terms to think about versus constructs that are going to be applied in a rule or regulation Those are very very different definitions and very different reasons why you'd want to do a definition So if you're doing reinforced learning it obviously involves a different type of technology and it may
• 25:00 - 25:30 give some different answers But if you were going to define that and then say we're going to have a rule and regulation that makes that split I think that becomes extremely problematic because you say so what the risks that you have when you aren't using reinforced learning are gargantuan We've seen those issues already in the marketplace and therefore we have rules and regulations that govern how people have to have well first of all you have to be registered to send orders to the market you have to be a registered
• 25:30 - 26:00 investment advisor in order to provide advice and things like that If you're using that technology then the risk of running a foul of those existing rules might be higher because you haven't really taken into account everything But that shouldn't mean that you would use that definition as a regulatory definition as opposed to an internal risk assessment definition I think that's pretty fair Greg When when I think about this from a regulator I if I were in say a policy and analytics group I might be very
• 26:00 - 26:30 concerned with these definitions as I'm say reviewing or auditing something to ensure that you know the underlying the underlying methodology aligns with the underlying testing the underlying disclosures etc Whereas yeah to your point the the thing you're actually going to write down needn't necessarily make those exact distinctions themselves Okay great If there's nobody else on the definition I would I would submit that we've answered the first question that it does matter Um I would also say that uh the good news is everybody's working
• 26:30 - 27:00 on this and understanding that there are differences uh in the types of AI and uh definition I think is going to be important at least at the at the outset and I again I like the the analogy to prior technological innovations where even if we wind up not uh having definitions that become important and rules down the road the conversation to get us there is uh critically important All right So we're going to get to some some meat of the panel in terms of how AI is
• 27:00 - 27:30 being used A big part of this panel is to give the audience an understanding of how AI is used in the industry And uh it's that's important because the the variation of uses is almost unlimited Uh we have five people up on the panel There are five million ways that AI can and is being used in the industry Um so we're going to talk a little bit about specific use cases and we're going to look at it by segment So whether it's investment management broker dealer exchange um your clients um uh and your
• 27:30 - 28:00 participants And we're also going to look at it by function the investing function and research trading operations customer interactions um and development and engineering which we can't forget the impact that AI has had on uh the increase uh efficiency of software development uh can't be understated So with that um Doug I'm going to start with you um in NASDAQ if you could talk about how NASDAQ is using
• 28:00 - 28:30 AI and uh some of the use cases that you have Yeah absolutely So I think there's there's four use cases I'd like to highlight The first one is definitely um around internal productivity and I think that's a theme we're seeing a lot with generative AI in um in industry right now Um you know generative AI as a productivity tool is like any other tool It's intensely useful It's the kind of thing that's going to to allow us to do more with the people and technology resources we have today and uh you know of course more with the people and
• 28:30 - 29:00 technology resources we'll have tomorrow Um and there we're really uh trying to understand how to safely apply or safely roll out AI tools and technologies to almost any area of the company uh whether that's you know a natural starting point like engineering and technology or even places that are you know maybe and I don't mean to pick on my attorney friends here but maybe not as technosavvy like our our legal and regulatory function as well I'm going to
• 29:00 - 29:30 pause that for a moment though and talk more about our uh in the product side of things which is uh there's three areas where we found where we've been using AI now for some time that have proven to be really really interesting really useful The first is in our index business Um we've been using AI in our index business since like 2018 to you know create minimum volatility port portfolios uh that become ETF products to create to use patents to define what companies are and are not particularly
• 29:30 - 30:00 exposed to different technology themes etc etc And there's obviously going to be more that are coming And within that area what we found is we're able to at kind of very low cost provide novel sorts of uh risk exposing products to many many many more types of investors that typically would be locked up in you know either more expensive actively managed funds or behind you know the the walls of uh being an accredited investor itself So in some ways what we're seeing
• 30:00 - 30:30 is that AI is allowing us to create uh more products for especially the retail investor to make sure that they can invest whether it's in their values or their beliefs about various secular trends We've also been using AI since 2022 to help us list strikes on our options markets Um if you're familiar with options markets and everybody in this room certainly is you know that most of the contracts on options markets like trade zero times or one time Basically not at all There's a huge amount of noise out
• 30:30 - 31:00 there effectively And uh you know our market makers have to cover that They have to cover very quickly and very diligently And so since 2022 we've been engaged in an effort to use uh artificial intelligence within the various bounds and rules set set by regulators and industry um to make sure that the set of product that exists on our options markets maximally aligns with the anticipated demand for the that product And then lastly because it's my baby and I want to talk about it all the
• 31:00 - 31:30 time um uh Dynamic Mellow which launched last year is much more of a pure institutional grade product This is an expansion of a midpoint extended life order that uses AI to make a very finicky sort of market structure decision about how long certain orders have to sit uh before they're allowed to be traded with the goal of improving uh outcomes for these institutional side loads in two areas that are traditionally thought of as in um you know opposed to one another So these are
• 31:30 - 32:00 three areas that we've been uh areas that we've been very active in and we think that there's opportunity to do more in how we manage our options markets how we bring uh new novel uh risk exposure and uh secular trend exposure products to investors as well as how we can improve and make market structure more dynamic overall All right that's great Thank you Doc I'd I'd encourage the panelists too as we're talking about this it's occurring to me that to the extent that we can talk about artificial intelligence the
• 32:00 - 32:30 specific type that they were that we're using that would be helpful just so that sure so for those three really quickly uh in index we're largely using largecale Monte Carlo style optimization and natural language processing in the world of op of options we're largely using kind of regressive and classification systems to make those uh strike listing decisions and then demount Dynamic Melo itself is a reinforcement learning system based on uh double deep Q learning
• 32:30 - 33:00 Perfect Thanks so much Doug Um Dan representing uh the investment management side of the world can you tell us how you guys are using uh AI sure Uh there are many applications but similar to Douglas I'll focus in on three First would be improving transparency Sorry Dan can you move the mic a little closer yes of course Can you hear me better now perfect Great Thanks Um I'll focus in on three specific areas First is the way that it's helping us improve transparency Uh an example that I would give is uh
• 33:00 - 33:30 algorithmic pricing which is helping uh with liquidity and also helping to reduce cost Second is acting as an investment process augmentation tool So there I would think through the ways that it's helping us achieve optimal trading strategies which assists with achieving best execution as well as reducing transaction costs And third is probably an area that doesn't sound exciting at first but is some of the areas where we're finding the most
• 33:30 - 34:00 interesting uh and applied uses of artificial intelligence including generative capabilities is in a focus on driving operational efficiency So on my way down to DC I spent the day yesterday with our operations teams in Wilmington Delaware And I try to get down there once every week once every two weeks Uh and these are the types of functions that you'd expect Reconciliations corporate actions uh the processes that live further downstream from the portfolio managers and the traders And
• 34:00 - 34:30 the way in which those teams have been challenged to use artificial intelligence to drive process transformation is some of the areas where we're seeing the greatest results And what's really exciting to me is two things One as we're having these conversations during this AI moment it's actually bringing to light not only AI solutions but also non-AI solutions that when strung together are achieving greater results for our teams and ultimately our firms and our clients uh and the way in which we're approaching
• 34:30 - 35:00 it is looking at the processes from an endto-end perspective What we try to avoid is making sure that we're focusing what we're trying to do is make sure that we don't focus in on any one specific area in isolation but looking at how the pieces in the overall investment management process can can work together to achieve results in totality Uh so those two things stringing together AI and nonAI solutions as well as looking at it as process transformations and making sure that those process transformations are businessdriven and enabled through technology as opposed to technology
• 35:00 - 35:30 initiatives uh that are delivered to the business is uh something that's particularly exciting for us Okay great Thanks Dan And uh don't knock the benefits of AI on the operational side of things Uh I will just relay prior to joining the SEC uh more than four years ago I worked at a AI platform company and the biggest use case there that resulted in a $300 million savings over five years for a large global bank was in the operations world where it was
• 35:30 - 36:00 an operations analyst that realized that they could be more efficient in moving money around the world instead of moving you know yen from Tokyo to Zurich on Monday only to have to bring it back on Thursday Uh they realized that by reducing the number of movements and being predictive about where they where they were going to need the cash around the world resulted in massive savings both in terms of available cash to lend and uh and uh fees transfer fees and so
• 36:00 - 36:30 forth And that was in the capital markets business believe it or not Um so don't knock the uh the operational benefits on on AI Um Mike a little bit about uh JP Morgan how you're using uh AI and what the specific uh uses are Yeah absolutely And where Dan went and where you picked up on is the operations is a big area And so um that that's been one that we've had for the longest standing time whether it's sort of the back office fraud detection credit decisioning regulatory compliance It's not as exciting as some of the Gen AI things but those to your point have had
• 36:30 - 37:00 the largest impacts and they've been uh sort of longstanding Genai is still newer It's only been about two years that we've had that So that's still kind of coming online There are some things we can see a little bit less in in maybe the trading areas but like helping customers summarize documents trying to uh write nice emails We have a firmwide product called LM Suite that is in use by 200,000 employees to just take a lot of the very simple basic tasks and put that right at their hands There's a mobile version of that that's been from an impact one of the largest ones Um
• 37:00 - 37:30 some of the others around again jai just answering client questions some of the marketing aspects of that as well So the operations is the largest though When you talk about the one that's rolled out to 200,000 is that was that a bespoke developed by JP Morgan or is that uh did you use a base i'm not trying to get into names and and numbers here but I'm I'm curious is that something that evolved with the commercial availability of of uh Genai or Yeah we partnered um you with Azure and ChatGpt which is a very common one and had to go through
• 37:30 - 38:00 some extra things so we could get comfortable with it from a data protection and privacy and all the things you might expect Um and so that was a long journey to get there and and then we feel comfortable and there we're continuing to build upon those capabilities and it's just got a lot of uses for that particular one So yeah Okay great I appreciate that Um Hillary sort of um pivoting a little bit to the end investors and middle market participants and sort of the uses outside We have a lot of institutional knowledge here um can you uh talk a little bit about some of the use cases
• 38:00 - 38:30 that you're seeing out there whether it's institutional or you know retail side or what have you yeah So I thought I'd talk a little bit about sort of the impact on institutional usage for investors because I think you know a lot of what I'm I'm hearing from the panel is that we're expecting cost savings through particularly operational efficiencies that hopefully will then be passed on to investors making it more easy for them to invest at cheaper prices and and you know that's been the pitch behind um a lot of sort of u
• 38:30 - 39:00 machine learning based portfolio management tools for example that instead of having to sort of pay a human to to construct things for you it can be sort of outsourced to an algorithm and there are potentially some some cost savings there Um a couple of things to watch out for though as well Um one is that you know we've talked about sort of training these algorithms as if it's sort of a passive process and it's really not you know the people who are training these algorithms have lots of points for intervention whether that's in the data curation stage or tuning the
• 39:00 - 39:30 algorithm itself and so we need to be wary of the fact that these sort of black boxes in many respects can be tuned in a way that perpetuates the problems that we see in traditional conflicts of interest in the provision of financial advice um or brokers dealers etc So um I think you know to the point earlier about defining um AI I'm all for the you know future proofing
• 39:30 - 40:00 regulation Don't tie it to a particular technology that's going to be obsolete but but these are long-standing concerns about conflicts of interest and just you know we need to be clear that the AI is not going to eliminate them So there may be some cost savings um in terms of portfolio construction but what are the trade-offs there in terms of lack of clarity about how conflicts of interest are actually being sort of filtered through the model Um another thing I want to note about sort of the the
• 40:00 - 40:30 operational facil um uh efficiencies that we're we're hearing about um in the context of say you know back office is cheaper because they're using Gen AI for summarization and things like that I think we need to be very cognizant of the fact that we're in the blitz scaling phase of Gen AI And if you're not familiar with that that's sort of the venture capital term for um making things available at very low subsidized cost to encourage adoption And then once adoption has been encouraged that's when
• 40:30 - 41:00 the price goes back up And really no one knows how much it actually costs the big gen AI providers to operate their models and respond to to queries So we don't actually know what that cost is going to look like but currently it seems like they're all losing money on every single prompt they respond to So when they start needing to make money those prices are going to be increased significantly And so I just want to be a little wary of the fact that you know we're
• 41:00 - 41:30 expecting these operational efficiencies to filter down to investors but in fact use of Genai could get a whole lot more expensive um perhaps more expensive even than having a human and and what does that mean for business models and the end flow on costs um for investors um in terms of actual Gen AI interactions um between investors and you know and institutions I mean the obvious sort of use case here is and some some financial
• 41:30 - 42:00 firms have tried it is the chatbot basically as a uh replacement of of um customer service Now the problem here is the hallucination problem Um the research indicates that hallucinations will never be eliminated in any significant way So you're basically as a institution trying to decide whether to use chatbots or not having to make a costbenefit analysis How comfortable am I with the fact that I can't predict what the chatbot will say and it could
• 42:00 - 42:30 quite conceivably be wrong Um Air Canada got into trouble um with this it it had a chatbot misrepresent its policies Air Canada tried to say it wasn't us it was the chatbot The Canadian tribunal said "Nice try." Um and so I think there's been some some weariness in using these in financial services where the stakes are so high with money Um and then also there's just the fact that people hate chat bots Um so Clara um the buy now pay later company went all in on having sort
• 42:30 - 43:00 of Gen AI as its customer service And then I thought it was very sweet On Valentine's this year it sent a love letter to humanity saying "We goofed Actually people want to uh interface with real humans and we're going to work on that." Um so I guess um there are potentially some cost efficiencies that can help out investors in the long run but they come with a lot of risks I didn't even get into my my hobby horse which I think will be dealt with later which is basically the systemic risks
• 43:00 - 43:30 that come from having more and more decisions made by the same kind of algorithms based on the same kind of data And even though there are hopes to create more data so so we have a basically all right I said I wasn't going to get into it I'm clearly doing it Um you know we we um we really only have one timeline for the market that's not a big data situation um to think about how different market institutions and prices um interact So we really
• 43:30 - 44:00 don't have big data There's a lot of interest in creating new data synthetic data with Gen AI but that's really just an echo chamber because that data you're creating is based on data you already had So I think there are systemic risks but I'm going to try and stick to the investor protection uh beat today and say that you know a lot of the the efficiencies come with tradeoffs Okay great Um thank you very much I think um I don't know if there's any
• 44:00 - 44:30 other use cases that uh anybody wants to bring up If not I sort of Greg let me talk about a particular use case Um and then we'll spend four hours talking about what what professor just talked about Um so for those of you who are young people who are just about to go into quantitative analysis uh and finance and are all excited about it uh please lower the volume so you don't hear what I'm about to say Not everything in quantitative finance is exciting
• 44:30 - 45:00 There are a lot of things that are tedious that are boring that are absolutely critical and that you're going to have to do regardless of what level in the firm you are Well maybe at some level you won't have to do it anymore So um in almost every field we have to consume data And of course you have to consume data when you are in uh quantitative finance whether it's a time series whether it's a spreadsheets of 10ks and 10 qs it could be an infinite array of data um in spite of the fact
• 45:00 - 45:30 that data and data science has been around for for many many years there are as many formats for data as there are the number of of data points that exist In fact there are more formats for for how data works The basic format that people have used for so many years has been a simple file separated by commas Everybody in quantitative finance has at one point or another has taken a file and said I have to bring this file in I'm really excited I'm going to analyze this And then you bring the file in and
• 45:30 - 46:00 it doesn't work And you're like well I can't do my work till I get this data file working correctly So what do you do you create a small program That program basically takes the data and it puts it into columns so that you can then move it and do the actual work that you that you think you're excited about doing So you write a little program and the program doesn't work Why does the program does uh not work because you were do you were consuming survey information that had in the first column the names of people and in the second
• 46:00 - 46:30 and 12th columns all sorts of data And in those names there was someone who had a junior or a senior So they put their name in John Smith Jr Well if you put a comma in a commaepparated file you've completely destroyed the entire file and you're nodding because every single person has done this So you take the next half an hour and you find out at row 4,236,000 you have finally figured that out You write a new program to do that And the way you do that is you put quotes around all of the names Put the
• 46:30 - 47:00 quotes around the names You're going to protect against the commas And you run it and it doesn't work again Why because in row 46 million someone entered their name as William Bill Smith and they put quotes in their who has quotes in their own name but they put quotes in their own name So you get dirty data and you write another script So you are constantly writing these small little programs It would be awesome if I can hire somebody who did absolutely nothing but sit there and write little programs for me all day Not a very exciting thing
• 47:00 - 47:30 to do You know who can do that genai can do that So you can use Gen AI to write these little scripts Do these scripts create systemic risk i'm not quite sure Do these scripts um create conflicts or interest well for the person that had the quote in the name and I took it out sorry about that But it does make incredible leaps in efficiency And if I'm going to spend an hour on working on a model I'd rather spend 59 minutes working on the model and one minute on the data instead of 59 minutes on the
• 47:30 - 48:00 data and only one minute on the model Greg giving me flashbacks man This sorry little PTSD it's oh god So I I will say this too that um so often you go in to solve the problem and you try to use machine learning or genai or whatever it is and you realize that 90% of the problem is the data I will say that AI and particularly Gen AI has come a long way in terms of cleaning that data up and making it so that the data doesn't
• 48:00 - 48:30 need to be pristine Um I I would say you know there was a day when it was my data is just not good enough I can't run machine learning or Gen AI against it I think those days are are numbered because the AI is actually helping to solve the problem of not having pristine data So uh with its errors of course So uh we we'll leave that there I want to circle back a little bit to what Hillary was talking about in terms of cost And I want to dig in on that a little bit just
• 48:30 - 49:00 to give the audience a sense of how are how are the firms that are using this how are you looking at cost are you looking at ROI how are you doing that i'm you know again not looking for hard numbers If you want to provide those terrific But more interested in the in the approach and the methodology in the in how important is that at the moment because uh there's no doubt that we are at the early stages where cost is less important in from a venture standpoint right they are plowing money into this
• 49:00 - 49:30 um I'm more interested in how the industry is looking at the use of AI machine learning gen AI and the cost of it and are you looking at ROI and secondarily Where do we think those costs are going to be born i think Hillary touched on that but I'd like to open that up to the panel in terms of how are you actually looking at the cost of AI i if that's something Dan do you want to take a shot at that is that an area that you're looking at at Black Rockck yeah Uh so
• 49:30 - 50:00 looking at ROI is definitely something that we're doing Sometimes it's more challenging in certain areas than it is in others Uh the way that we're approaching it is by looking at ROI through a few key lenses First would be going back to what we've talked about amongst the panel operational efficiency and productivity really as a means for achieving scale Second would be alpha generation and protection Third would be revenue And fourth would be operational
• 50:00 - 50:30 risk reduction Those are the real lenses that we're looking through in terms of how we're looking at uh the overall ROI framework Uh and what I would say is is that there's also some inorganic ways of looking at the ROI too Uh in some of those conversations I had yesterday with our end users just thinking about the time that they're getting back to face off with external clients the time that they're getting back to spend on more complicated challenges that would otherwise would have been spent on level
• 50:30 - 51:00 one type tasks Those aren't perfectly articulated in scale metrics So thinking about how we can also think through capturing those types of metrics is very much on our mind as well I can add a little bit So I think we are looking at this as well You kind of have to um and it also wasn't new There was investment governance processes that we had for technology that would tie back to some business problem we were trying to solve and benefits You know this is all technology at its heart and so it just kind of builds on top of that platform Now where it gets hard a little
• 51:00 - 51:30 bit to what Dan just said is that some of these aspects are time you know fraud loss well how do you kind of so it's a little bit less precise and I think the the aspect to get into around is just that you've built it more in the culture if you over mechanize it that can be a challenge itself it's just built into the culture that you have a business leader who's thinking about it and you've got these enabling things over here and you're using one of the methods that Dan talked about or other ones and it's allowing you to steer the ship Um that's what's really key And then you know as far as the outcomes and that
• 51:30 - 52:00 it'll always have to be back tested because it's going to change And so it's not a static thing Yeah So we obviously are thinking about ROI as well I think if you're in industry and you're not thinking about ROI first and foremost you're doing your shareholders a grievous disservice Um so on the revenue accretive side and I love revenue accretion because it's a lot easier to deal with Uh it's pretty easy Does the product make money yes Cool ROI On the other side a lot of what we're talking about to Dan's point I think is
• 52:00 - 52:30 uh those scale metrics in particular cost avoidance And there I think what we've seen is uh we're getting a lot of mileage obviously out of uh AB testing Uh which is great Uh but we're getting a lot more mileage than I think anybody wants to admit out of work journals and stopwatches just looking at how long things took and how long things take um you know that that can get you further than than you'd think in trying to make good estimates about the sorts of value you're driving with uh new tools and
• 52:30 - 53:00 technologies It's not perfect because it's an estimate Nothing's ever really perfect but I think it gets you close enough that you can drive confidence that uh you know the the relatively low costs at least that we're experiencing today while venture capitalists pay us to use their systems effectively um because they're losing money on every prompt I'll clarify that is uh is worth it I might weigh in for a sec if I can Um so first of all I want to agree with
• 53:00 - 53:30 Michael that the I think fraud and sort of moneyaundering detection using machine learning is an A+ use of this technology that is really you know pretty uniformly beneficial Um with regards to sort of trying to to estimate the return on investment in terms of employee productivity I want to point out that there's some conflicting survey evidence on this So there really seems to be very different views depending sort of on the the level within an institution at which you sit
• 53:30 - 54:00 So a lot of um surveys of actual workers have said that they they really don't feel that AI tools increase their productivity And in fact it's challenging because their bosses assume that their productivity should increase but in fact using the tools and sort of weeding out the hallucinations etc is actually more timeconuming often than just doing the work themselves And so there is sort of a disconnect in perception I think depending on where you sit in an organization about how um
• 54:00 - 54:30 you know the return on investment these tools are and you know I think it's sort of been borne out um you know the standalone Gemini and co-pilot um products for Google and Microsoft really didn't get much uptake They actually had to bundle them into their other um services because people didn't want to buy them on an individual basis So I just don't want to overstate the productivity gains from these Gen AI tools Sometimes they're actually more time consuming than just doing it
• 54:30 - 55:00 yourself the first time So I don't want to sound like a broken record but I I do like to go back to history especially when you only need to go back a few years for any given topic and just see how things played out So analogous conversation that we had exactly like this cloud computing what's the ROI for cloud computing is your firm using cloud computing how are you thinking about cloud computing what about the people who are going to get displaced because they're no longer maintaining servers but you're in the cloud what about the privacy issues so a
• 55:00 - 55:30 lot of the things that we're discussing now we discussed as recently as five years in panels on cloud computing So what lessons were learned from that lessons in my mind the primary lesson was that it's not static and that there's no uniform solution to this A lot of firms went out and they said we're going to move 100% of everything into the cloud because that's the way to go for the future Um and for some firms that might be And then other firms said this was a terrible idea Now that we've experimented we're going to move
• 55:30 - 56:00 everything back in That's also pretty extreme Not a lot of firms went went all the way out and then all the way back in But what a lot of firms did find is that there are certain things where this actually really works well and there are other things where it doesn't really work well because there's methods that if you do internally are more cost-effective and if you do them externally in the cloud they're more cost effective So my guess is that if we fast forward three years from now we're going to find that there are things that today's technologies are particularly good at and um and enhance productivity
• 56:00 - 56:30 And there are things that people tried and experimented and it turned out that maybe it wasn't as valuable and they actually wound up pulling back This is the explorate the explorative side So exploring all of those things is the only way to figure out what the answer is going to be unless you're going to create an AI that's going to sort of tell us the answer in advance Yeah this really is the fun time in this uh in the evolution of uh AI and and back to Hillary's point too about uh
• 56:30 - 57:00 fraud being custommade for one of these It that's where I sort of find the separation between the types of AI that you're using reinforcement learning is is perfectly suited for doing fraud detection taking something that it doesn't know exactly what it's looking for and you know having reinforcement learning try to figure that out So um you know to Greg's point this is really the fun time of this The exploration and the and the experimentation about what we're doing is uh is important and that's going to cost some money and it's
• 57:00 - 57:30 going to be important to see where you're getting return on those investments Um let me pivot a little bit now and I want to talk a little bit about the barriers to adoption of AI because we are at that stage where um there are you know there are uh failures um you know any any sign of a good technologies you've had a few failures in it and you've been able to um you know figure out how that's going to change your strategy So what are some of the the barriers to adoption that you're seeing whether they're technical barriers or non-technical barriers
• 57:30 - 58:00 business um including data is that we talked a little bit about that but is that really a barrier to the implementation of AI Mike we start with you yeah and I'll I'll bring us back to actually something uh Hillary said and then Greg to your point um there's a bunch of aspects around you know policy frameworks and data and the technology but honestly at the heart of that is the human and just like the cloud journey It was easy to go move your technology to cloud and some folks did that But what a lot of times folks got tripped up on is
• 58:00 - 58:30 the cultural change of the humans on doing something very differently and building cloudnative technology and doing that And I think both on the employee usage aspect and then just that journey of it you can give us the tools and there are these barriers and those are hard but then it's that changing that culture of the humans often where they may feel uh threatened or displaced uh you need them to help on that transformation journey That can be a very tough pill to swallow depending on their perspective and fear in that So that's the one that honestly at the very top I've seen both back in the cloud journey Greg that is deja vu and I'm
• 58:30 - 59:00 seeing kind of play out again right now Um I think the other aspect is you sort of talked about technical non-technical and back to cloud we often reference things of business business and this is technology and data started to bring those together when you thought about the ways you could develop and low code and no code and so you could have business doing technology type things Well now that barrier has been completely broken because it isn't as distinct as it once was And so that what I'll call artificial barrier of well this has to be a tech thing or a business thing This very much sits in between Um so kind of tying it back you
• 59:00 - 59:30 know having that cultural breakthrough and and getting that in place uh the data aspects then you're able to get into what I'll call the model explanability and a lot of the ethics in that And so we've got uh we've got groups that look very much at that but the the human cultural journey is the hardest Yeah really great points I I like how you you you talk about that uh sitting in the middle too It's it's a really important point that this is breaking down those artificial barriers that we've had in place Greg you and I had a
• 59:30 - 60:00 little bit of a discussion around the data in the in the prep for this call Um whether it's data or some other uh you know uh impediments that you're dealing with in terms of implementing AI Anything that comes to mind for you uh or top of mind I guess So not not so much the impediments from from looking at the um AI itself but I think more broadly uh and this is something that that I think Hillary had had mentioned when when we think about the blackboxy
• 60:00 - 60:30 nature of of models I think we want to be careful about how we um we define that and also how we're concerned about We've had models that were considered blackbox for many many many years And there are solutions to that The solution to that is to open up the box And when you open up the box that doesn't mean that you understand every single thing in a box You only have to understand enough of it and enough of what the boundaries are in order to make use of
• 60:30 - 61:00 that safely We do that in our everyday lives There's no reason why we can't do that in in finance etc Um if you look at some of the models that are used today whether or not they're just playing machine learning whether using reinforced learning ultimately it's a statistical model that's seeking patterns and seeking to optimize something Those were originally done through um what's called mean variance um optimization where you use a return projection and you use a correlation
• 61:00 - 61:30 matrix and you put those into a framework and you come out with an answer But a human has to digest that answer And if a human is digesting that answer and doesn't recognize that I think either I put the wrong data in because it told me to sell everything and go short one exact stock and sell my house probably not a good recommendation to do that But if the results look reasonable and it's within the frame of what you had expected but it's optimized in a slightly different way than you otherwise would have that gives you some
• 61:30 - 62:00 shy that it's doing what it's supposed to do When you get comfortable with that then you get a very good handle on whether or not the machine is doing what you expect The danger about using any new technology is to just get complacent and say you know what this thing is perfect they don't have to look at it wrong You have to look at it and you always have to look at it So it does become much more of a cultural phenomenon to just don't get complacent Keep using it with the same diligence that you did the day before you use the product
• 62:00 - 62:30 Great Thanks Greg I want to um I want to go to Doug next and after that I want to go to Hillary to talk about some of the um you know some of the barriers to adoption that we're seeing overall in the industry too But first Doug if you want to talk about some of those barriers to adoption Yeah So I actually think um so I'll give you my not hot take and then I'll give you my hot take that uh OGC and comms always cringe whenever I say Um but I'm still here They haven't fired me yet so I'm fine Um so uh the less hot take is one of the
• 62:30 - 63:00 biggest barriers I think we see to adoption is that AI development is just intrinsically different from traditional software and other technology developments How you test it how you build it how you think about it And we're still uh frankly we as well as as almost everybody else on planet Earth who's not maybe named Google are still really working on how we staff up and uh cross trainin people to kind of develop those years and years and years and years of quantitative acumen needed to
• 63:00 - 63:30 really be designers and developers of these systems Um I think the other part of it that's a real challenge is that for people who are either product managers or users there's this huge sea change that occurs and this is my hot take uh because AI in many ways is much more explainable and much more auditable than the traditional ways that people make decisions When an AI system makes a decision we know what data was available to it We know what decision it made uh it can't after the fact say make up
• 63:30 - 64:00 things like you know I crashed my car when I was 16 and after the fact there were all sorts of reasons it happened but the reality is I was just playing with my radio Um AI can't do that and because we measure everything with AI in ways that we don't in our usual life people get like really have really weird responses to or expectations about what it should do that are not aligned with reality I remember one of my early experiences de delivering um a decision
• 64:00 - 64:30 system um when I worked for an aerospace manufacturer in the Seattle area and I was talking to a group and said well you know what what kind of accuracy do you need think you need to be at here and they said 100% Okay so you're at 100% accuracy now yeah you never have a disagreement about anything Well no we have disagreements all the time okay about what things mean So okay it's not 100% You've never had an error Well you know we have errors
• 64:30 - 65:00 You know we we have errors Okay it's not 100% What is it actually and it turns out the reality is that if we got like above 60% accuracy we were delivering a lot of value But because we're actually measuring a thing that means that we get to say out loud a number that maybe before we didn't want to say out loud And that can be a huge impediment if all of a sudden it comes to light that this large aerospace manufacturer in the greater Seattle area had a particular process that was correct 55% of the time
• 65:00 - 65:30 It wasn't a it wasn't not a critical process Didn't involve doors Uh and it should have been right And being right 60% of the time was very good when in our human experience 60% is a Dminus Uh and you get yelled at by your parents Right So I think that that aspect that you're measuring everything uh that you can't make that you can't kind of like throw somebody under the bus afterwards These are actually real human like human factors impediments to a uh AI adoption that um I think if you are either a regulator or a practitioner you have to understand how to uh overcome artfully
• 65:30 - 66:00 So you know listening to to Greg and Douglas reminded me of one of my my favorite little AI anecdotes which is um it's in the medical field and the this is this is not Gen AI this predates Gen AI Um but basically the the output of the um the AI was that having asthma made you at lower risk for pneumonia um of death from pneumonia And the reason why it had concluded that it
• 66:00 - 66:30 turned out was because people with asthma are at such high risk for pneum dying of pneumonia They get more attentive medical care and therefore are less likely to die But if you just took the results of the the AI it would suggest that you didn't need to pay attention to people with asthma if they had pneumonia to the same degree as others And of course the doctors saw that and said "No that doesn't make sense." And so I think this really illustrates the point that both Greg and
• 66:30 - 67:00 Doug have been making which is in the best world you have smart people who know what they're doing who use the AI tools and they are enhanced by the AI tools because they know when it's off base They know when it's outside the realms of plausibility and it might you know highlight a different perspective that they hadn't thought of That's that's the best case scenario And that's um Corey Doctoro who's a a tech blogger I think calls that being a centaur right you have a you have a machine that helps
• 67:00 - 67:30 you do your human stuff better But the problem is when you get a reverse centaur right where basically a human becomes worse because of how they become dependent on the machine And I think the challenge with AI is that really to have the sort of the the confidence to say the machine is wrong because we have this these automation biases They're well documented We tend to think that anything spit out of a computer is better than what we come up with
• 67:30 - 68:00 ourselves It takes a lot to be able to say no the machine is wrong And so that's a real challenge in terms of adoption because you have may have people deferring to stuff when they shouldn't And sort of to that end um you know I agree with the points that have been making that you know made that say you're the SEC and you're trying to enforce your investor protection rules saying you know you haven't acted in the best interests of your clients when making recommendations Now we have data because
• 68:00 - 68:30 it's been fed into a machine but you need to have the resources at the SEC to be able to take that apart to understand it So it's really um you know on a sort of writ large society what we're asking for in order for AI to make us better is for everyone to who uses AI to be smarter than their AI or know more about the area than AI And I'm not sure how realistic a proposition that is So we need to be mindful that there are a lot of
• 68:30 - 69:00 instances where these can lead people astray So that's that's one um barrier to adoption I wanted to discuss Another is in Gen AI the hallucination problem It's sort of been spoken about for years that we'll get rid of it with scaling If we have more data more compute these will get more accurate It's become increasingly clear from the research that we've sort of hit more or less a dead end in that regard There might be marginal improvements but these hallucinations are not going to be eliminated So to the points I was just
• 69:00 - 69:30 making you need someone who knows enough about an area to know when that you know response is a hallucination And the thing that I worry about is that how do people learn enough to know when a machine is hallucinating well they have to do those low-level boring tasks in the first place But now we're talking about supplanting those low-level boring tasks Now the AI is
• 69:30 - 70:00 going to do those So how do you learn enough about your field to be able to point out a hallucination and go against the machine and so that I think is a real challenge to adoption as well Right I think so I I agree with everything that you had just said about those concerns but um I do think we should level set both about expectations and what um uh what we really should demand of any technology Just show of
• 70:00 - 70:30 show of hands How many people have ever dialed customer service for anything in any category ever one person didn't raise their hand You're a very very lucky person Um and and of those of you who have called customer service how many people have gotten 100% the correct answer by the ends of that customer service and you exactly knew how to reboot your router in the way that they told you to get your cable back on or fix your flight and get your seat aligned You you have
• 70:30 - 71:00 been given hallucinations for many many years You have been given bad advice and wrong advice I got very bad advice just calling an airline the other day and I had to call them again Just is there anybody else that I can talk to there thank you very much Type of thing So um what are we asking a Gen AI program to do are we saying you need to not be worse you need to be equal or you need to be better than what you otherwise would get So if you're calling customer
• 71:00 - 71:30 service or you're calling your doctor for advice that always has imposed some sort of responsibility on the user of that The question is are people saying yes but if it's done with AI I don't have to do that anymore Um I hope they're not saying that because there would be no reason to say that Maybe you go 100 years in this you know Star Trek future we can get there But as long as you are as cognizant as you would be if it was a human then that is a really good guidepost to start with
• 71:30 - 72:00 Yeah Yeah I know such a good point too that um the one thing I keep thinking of so those of us that are in AI because it's in the zeitgeist now I I know I get asked you know how do I do this why do you use that why do I need it and I enjoy showing people how to use claude or chat GPT this is all my personal life make that clear um and in doing that I it's kind of funny I take particular pleasure in letting Claude or Chad GPT
• 72:00 - 72:30 know that they just hallucinated and saying that's a hallucination and you get an immediate apology back They always say yes you're correct I'm sorry for my mistake And then it it comes back and when that really happens and I caught it in like a triple hallucination last week and I kept saying "No that's a hallucin." And they were dead It was dead wrong So the importance is to just keep practicing Like keep using it so that you can detect a little better when that seems a
• 72:30 - 73:00 little bit off So I I just it it stressing the importance of that that we're used to this We've been hallucinated to forever Uh absolutely true and uh it's going to continue Um let me uh the next question is is one that I think I'm going to start with Dan and then and then go down the line I think So this one is um is AI good for investors and what are those benefits to investors um and their advisers um from
• 73:00 - 73:30 the perspective of the industry and what I mean by that is when we look at this holistically and the use of machine learning and gen AI and reasoning and agentic AI you know some of the the uh new methods and techniques is this is it good for investors and if so what are those benefits why is it good dan great So I might take this in a little bit of a different direction than some of our other panelists One of the areas that I
• 73:30 - 74:00 would highlight and suggest that we focus on uh partially would be the investment opportunities that it does create for the individual investors in terms of the investment vehicles and the investment products themselves So the AI uh the AI moment that we're living in right now opens up significant investment opportunity for individual and institutional investors across ETFs mutual funds and private investments throughout the different components of
• 74:00 - 74:30 the overall AI tech stack So a few areas that I would highlight are the infrastructure investment opportunities that are created think chips think power think cloud infrastructure Uh I'd also highlight the intelligence uh that sits within the tech stack So the models themselves and then also all of the applications and services that live around those capabilities are offering opportunities from an investment perspective to individuals that are packaged in all those different ways
• 74:30 - 75:00 that I mentioned before And certainly from the perspective of BlackRock making sure that those opportunities are available to our clients is important to us And one of the areas that we're focused on is making sure that those products are available to our clients through the different different ranges of products that we offer out Um that that that'd be one element that I would highlight and I'd turn it over to my panelists to talk about some of the other angles Great Hillary So maybe I come at this from a slightly different um angle than than Daniel but
• 75:00 - 75:30 um I you know as an investor would be a little bit wary potentially about investing in this space There's been a lot of talk of a bubble um in particular Microsoft has been cancelling data center capacity and so this might not you know you never want to invest at the top of a bubble So just a you know word of caution on that respect with regards to the impact of in on investors um in other ways I've already talked a little bit about potential cost savings and the trade-offs The one other thing I wanted to note is that one thing that geni Gen
• 75:30 - 76:00 AI has been excellent at is um scamming Um it's created all kinds of new ways of um you know doing video scams or very um personalized texts etc So um I think you know one thing that the SEC can really concentrate on is investor education in that respect in terms of protecting themselves from generated scams because that's definitely on the rise So uh my answer better be yes Uh
• 76:00 - 76:30 because I've dedicated my life to this I think um this technology is definitely well you know there's obviously areas for concern uh definitely good for investors both today and in the future And today it's good for investors because to as Dan said we're able to offer re kind of democratize many of the tools of quantitative finance through the ETF space ETP space we're able to create uh dynamic market structure that's able to kind of evolve as the markets evolve to make sure that to maximize the efficiency of price
• 76:30 - 77:00 discovery in those markets in the future One of the areas that and this is my opinion not NASDAQs um that I'm most excited about is um and it's related to actually first I'll give you my most controversial take uh human trafficking is bad And a lot of what we do to prevent people engaged in human trafficking from getting access to financial ecosystems is uh know your customer And it's very good at preventing us from you know accidentally financing Boo Haram That's great It's also really good at cutting off capital
• 77:00 - 77:30 flows from small business owners in you know emerging markets and failed economies at debanking people in uh poorer parts of developed economies etc etc AI allows us to take very different approaches to how we think about the world of anti-moneyaundering and and anti- fraud that might allow us to in a more nuanced way make sure that we can get the small business owner access to deeply liquid capital markets to make sure that somebody who is getting their first job after making some mistakes in
• 77:30 - 78:00 life are able to have banking facilities for that while still keeping those bad actors out And that to me is actually something that's really exciting that I I hope folks at the SEC will consider Mike on your end Denny yeah Also yes I mean um I'm a big believer in innovation I think there's a reason we're not running Punch Card and mainframes in our infrastructure anymore so far as I know Um and and this continues to drive things and allows us to provide better products and services to our customers
• 78:00 - 78:30 Okay great Um we're running up a little bit We have about 5 minutes left and I want to take this opportunity to to ask two questions and sort of a lightning round of the of the uh panel Um the first one gets to what chairman UEA had had referenced uh in his earlier remarks and that's um if you could pick one thing in fact all the commissioners mentioned this If you could pick one thing that the commission could or should do when it comes to artificial
• 78:30 - 79:00 intelligence what would that be so floor is yours If there's something that we should be doing as a commission uh around artificial intelligence uh what would that be Doug yeah So I think we hit on this actually a little bit in in the DMO um approval process but I think the one thing that the uh commission should really start thinking about is the different sorts of protections that a B2B product might need versus a B TOC
• 79:00 - 79:30 product Um you know with all due respect to my my friends at Black Rockck if if I were to sell them something that was not fit for purpose that was uh going to rip them off they would very very quickly stop using it They have armies of quants They have huge armies of analysts They you know they know their business incredibly well They think about it 100 hours a week in ways that you know my you know my aunt Edna doesn't necessarily think about when she's investing in you know what investing in a product or you know getting on Erade or whatever it might be So I think
• 79:30 - 80:00 making sure that we that there's some distinction between the the professional investor and how they interact with AI versus say the retail investor uh is something I'd love to see the SEC uh codified Great Thanks Why don't we just uh jump to Dan since his name was invoked there i'd keep it simple I'd say this right here these types of engagements bringing together stakeholders those of us on the panel as well as those of you in the audience and uh watching in uh this is an evolving space and continuing to talk
• 80:00 - 80:30 as things evolve continuing to align on best practices uh and common principles super important So very much appreciate this opportunity to speak today and look forward to more types of engagements like this going forward Great Thanks Dan Uh Mike yeah I would just echo that and you know at the beginning it was stated a few different times of avoiding sort of these checklist kind of you know uh technology itself being regulated versus that risk based approach So I think that tone's already been set
• 80:30 - 81:00 Great Hillary I mean I know it's challenging times but staff up on technologists if you can Okay I think we heard that Um I think the commission should take advantage of the um educational aspect So uh today we live in a regulatory environment that creates an enormous amount of protections for individual investors If you're an individual investor and you're working with a broker dealer or individual investment working with a registered investment advisor investment company there are an incredible panoply
• 81:00 - 81:30 of rules that are designed to put protections on you How do you know as an individual investor that you're actually working with one of those parties that's where AI um when applied directly to the individual investors um provides an opportunity for bad actors to to pose as a broker dealer or investment advisor Um I know the SEC and FINRA really hammer on the idea get the the identification number of the party that you're looking Go to the website e the SEC website FIN
• 81:30 - 82:00 website type their number in They should come up with their name their history the firm that they're at If you're presented as an individual investor with advice on the web um regardless um you should ask for that information Can I have your CRD number um I'd like to look you up If the thing comes back and just hangs up on you then you probably were not dealing with a registered party So the more you know about how SEC rules and regulations protect you the more as an individual investor you'll be aware
• 82:00 - 82:30 of something is a foul here I don't think I'm necessarily dealing with with a registered party Great Thanks And then uh we have a couple of minutes left Um final question and uh hopefully this sets up a little bit for our last panel tonight which is the future of uh of of artificial intelligence over the next 12 months 24 pick whatever you like Um what's the most significant development that we're going to see in AI in the financial markets uh over the next 12 24 months
• 82:30 - 83:00 i'd like to see 12 but if you need to go out 24 we'll do that too Uh why don't we start at the end Dan and work our way back Yeah I think what's interesting is is that I might give you a different answer within that 12-month period Uh but two that I'd highlight right now are agentic frameworks as well as uh complementing large language models with small and medium-sized language models I'd say that for AI and for humans it is hard to use historical data to make
• 83:00 - 83:30 predictions and therefore I will not make one Well said fixing a few very very visible uh let's say kurfuffles that occur from vibe coders creating AI AI agents I think continuing to mature in the gen piece that's still relatively new but then also to Dan's point the agentic is going to be a game changer my guess would be if we held this panel in 12 months from now we would all say we had absolutely no idea when we talked
• 83:30 - 84:00 12 months ago that six months after that this thing would have happened and that's everything that we'd be talking about So I think it's the unpredictability and the excitement of new things that are coming out Well said I'd say that's why we're all here today and why it's a topic of interest Um with that we're out of time Please uh join me in a round of applause for the panelists Thank you all