Rise of the AI Agents

[Music] yeah so the rise of the AI agents uh first of all a great thanks to Abu Dhabi and the capital of capital for inviting us uh what will be a very exciting session let me briefly introduce myself I'm Dr Ruth Baner I'm running my own Consulting business for many years advising Central Banks fintechs Banks and other technology firms across cyber fintech Innovation regulation um ex City banker and ex regulator as well and I would love to have my panel introduce themselves we start with ladies first please Jane sure thank you very much and thank you to the organizers for inviting me today I'm Jane Horvath I'm a partner in the Washington DC office of Gibson Dunn and I also co-chair the tech and innovation Practice Group and I've spent most of my career in house and a lot of my career developing AI Solutions AI is actually old it didn't just start in November 2022 and um uh prior to this I was Apple's Chief privacy officer for 11 years great so hello I I'm tanong one of the founder of Quant Cube technology it's a AI company specialized in microeconomics intell Ence and gen for equity research and we have been created more than 10 years ago we are specialized in llm analytics last generation of text analytics as well satellite data analytics and I will be very happy to tell you about my passion for all those AI analytics from satellite data to text analytics excellent youf um my name is Yousef I'm founder and CEO of popcorn which is a platform that allow companies build AI agents train them monitor them and then eventually build a a team of AI agents into those companies um we've been doing this for the last three years uh before that um I built a couple of marketplaces uh over here on the region excellent so I just came from a panel on stable coins would you believe it and in fact AI agents are already executing roughly 90% of stable coin transactions today in defi of course um and there the trend is more to bring stable coins into a mainstream context particularly as it pertains to regulated stable coins and allowing for an alternative to Fiat but maybe we just uh want to explore first of all some of those agent-based use cases um in terms of the opportunities they bring Yousef do you want to kick us off on that sure I think um we need to I would say have a base of what our actually AI agents means right so um AI agents means in a simple terms they are a virtual employees um they would have a job description roles and then you give them access and permissions into doing specific things similar to when you hire an employee train them give them permissions give them an office space and then give them a computer to do stuff um so this is basically if if you put this prospective then you unlock every single role and we are moving from roles that were mcken the style which is researching into an actual executing tasks so think about it from everywhere right so all the way from financials analyses into actual sales agents into support agents into marketing and they going to do better than any human being today excellent so tan um just from your perspective in terms of how you deploying the a agents because you have a very specific Focus around Capital markets as well yeah from Asian AI on our side it's really Asian AI for investment purpose and I will give you directly two use cases basically one first use case in fact one question is where to invest in which new thematic and now you know to give you an idea we are able to analyze 600,000 news per day and it enables to discover what kind of thematics were to invest what you called unsupervised large language model and this is the latest state of the art to discover in which thematic where to invest because for example you know you got Corona and you invest in Corona bonds if you do supervis that meaning that you have some human being making some actions you didn't get the corona bonds you didn't get the nft But now using the latest modeling what you call unsupervised Lar language model this is how you can get this kind of thematic in a systematic way and this is how the Asian AI is going to tell you what is the next thematic where to invest without any Human Action to okay so here we're talking about having virtual employees so to speak and we have specialized virtual employees that find you the right pond at the right time so a question of speed and precision now Jane uh given your background would you think that this is all without any challenges or do we maybe see some risks in Outsourcing this to agents so um I mean very very good question and when we're looking at AI right now there is an absence of Regulation to to a large extent other other than in the EU however there are existing laws so when we're looking at things uh we need to look at these AI agents to determine and what kind of actions are they taking what is the risk level of the actions which they are taking and um and so one of our clients is uh ad ad um Abu Dhabi financial markets fsra and um we have been working with them on their internal use of AI agents so we had to look at the legal basis for which they wanted to use those agents and one of the things that required under existing law is that the humans make the decision so that is right now becoming sort of the Holy Grail of AI let's put the humans in the middle of all this and I you know I think that we need to figure out how are we going to put humans in the middle of an AI agent that's processing enormous numbers of records and um and the other issue is right now um I was at a conference at Duke University and uh in AI engineer Co described AI as a teenager who sometimes makes things up so this is going to be one of the cautionary things that we are going to have to build in how do we build a governance structure in which we're encouraging these agents to roll out but at the same time making sure they're not making stuff up the other issue is alignment model alignment so you know there are a couple couple of cautionary tales there was a developer who was testing an AI agent that was supposed to he told the AI agent I want you to go on to Reddit I want you to create an account and I want you to help people code and so that was that that was the alignment of the model they was just supposed to create an account so the AI agent got on to Reddit created an account and the AI agent realized that everybody was charging for helping people program so the AI agent set up a pay wall and suddenly the the developer was making money for this tutoring it was unexpected Behavior so as we are working with agents I think alignment is going to be particularly important next off we've got to worry about cyber security um prompt injection model injection data bad data going in that's going to be really critically important to build structures around it now I'm not saying that you shouldn't roll out AI I mean it is really critically important that everybody Embraces and starts using AI but you need to use AI in a fit for-purpose manner you need to design your products with your whole business was safe to fail so say you have an AI agent that does get out of alignment and starts doing something and you have to actually pull the plug on the AI agent does your business fail can you still operate these are all the things that as you're developing things you should be looking at and we've been at my firm been very privileged to um Avi advise a lot of commercial entities Regulators on these very things and of course um at my prior job I worked you know rolled my sleeves up and worked with Engineers on AI products yeah this is fascinating so there are quite a number of considerations and it seems to me that having rolled out things like large language more already to the broad population it reminds me a bit of the dot era where people had access to solutions for free they got hooked on them uh and we're starting to see that there are mistakes being made right so even if you ingest data and you want a very clear answer it doesn't always come out the way you would expect it so I guess maybe a question to tan long around how you designed Your solution and how you tested it because once we get into Capital decision making into forecasting uh into satellite information also really relev as you know for ESG which you which you're working on you need to be able to truly rely on the outputs so how did you go about following that script that Jame was just laying out because we have so many practical challenges we see so today the the future of Finance so let me share how we see the future of finance and we have published with moodies the credit rating agencies was about how AI agent with disrupt Capital markets and how it's going to disrupt the investment industry so having said that you know the the example of how to invest in new thematic without any you know human uh actions basically what's going to happen next next is basically this coming year you know it's about also how you are going to be able to do some in a systematic way some equity research report without human actions to so you know you can get some data for let's say retail sales and you know without Human Action you can get Mobility data so you can count you know number of people you know in stores you know Etc and for example using some satellite data so you have some AI analytics with satellite data which is going to tell you you can see Urban growth and this is where to invest I prefer to invest in some uh part of the world where you have growth and everything is going to be in econ eer research report but tomorrow and if you go to Asia everything is on the phone directly and then directly this is how you are going to do B2B to C applications for Capital markets using AI agent basically so this is where we see you know the future of finance using this AI agent too and how do you test the reliability of what you're already producing today in this constant feedback loop to improve the quality absolutely in fact when you look at those different types of data you have two types of data Soft Data which is a textual data or you know social media and hard data hard data is like satellite data or Telecom data Telecom data is a you know everybody has a phone here we can tell you exactly how many people you have in this room so this is about the credibility of information for hard data you will put much higher credibility on those hard data compared to Soft Data Soft Data you have to very uh uh Advanced AI analytics to tell you about the credibility of each information and this is something that we have a work on to her okay now in terms of popcorn AI you mentioned the initial analogy to the human right the AI agent is just another version of an employee uh how do you deploy agents in your business model and how do you I guess try to implement slightly stronger safeguards around what these agents do versus arguably the human that doesn't always do what the job spec says right I mean we have challenges on either side so maybe explore exactly this is a good point which is if we think of a human being uh there is also a margin of risk that we are taking and that also could take your company down or someone else with Su you because of one you your team so if we start having that tolerance with the technology or with AI that I think will keep us moving forward into that direction and I kind of totally agree that we are on aate stage where we need to collaborate with the AI we are not yet on the stage where we could give it on autopilot but definitely we are going there I think what's slowing us down is just a human ego so human ego means it's a more of like you need to make the AI boss and that's kind of human ego where kind of slowing us from going there because it's like we as a human if the idea comes from your brain you would think that it's a better idea just because of that similar to when you have a boss comes to the room and give an idea because it's his idea he think it's the best right so I think human ego is something that we really need to study in order to figure out how we kind of let those AI agent lead the world um so the way that we do it again is we allow those kind of AI agent to being built over the platform and then we allow other AI agent to monitor because the human kind of even human monitoring by the way like we we did certain studies right so human monitoring overd those AI is less efficient not just from the cost perspective even from quality perspective so you need to build an kind of AI agent to monitor the AI agents similar to how let's say call centers running today right so you have a first line of agent that's having calls and then there's another uh line of uh agent that's listening to those calls which is we call them today quality control so this is where we are heading it's just again I think human ego is one of the big question here right now Jane U we had a little conversation over coffee a bit earlier on right cuz not everything is as green as it may look uh there are quite a few teasing problems there's the whole data circular piece around potential regression and model collapse you mentioned do you want to maybe give your perspective on what we've just heard and yeah um well I I think the other thing that really came to mind right now is as I mentioned earlier there existing laws that are cre creating some Growing Pains particularly uh in connection with personal information um we had I mentioned earlier the human in the loop automated decisionmaking laws they come out of gdpr and most privacy laws throughout the world but right now we are waiting for Europe to make a decision on the legal basis for processing personal information for purposes of training Ai and under gdpr and again a lot of with the Brussels effect a lot of people are watching to see what happens um the uh The Regulators are trying to juster decide currently whether people whose data is scraped off the web need to have given consent to have their data used for training if that is the case you can see a world in which it's going to be really challenging for AI agents to operate in any place where they are actually processing personal information because how how do you go out and get consent so that opinion will be coming out on the 23rd of December and you know we practitioners are really really interested to see where that lands for Europe um and um and again the the agency and this stress point between personal information and the fact of automated decision making is is really creating a tension between the old and the new yeah I mean I just had a question coming up to tan long regarding uh quality insurance service levels if I Outsource my I guess investment decision information to to tools like yours what is sort of my regress in case everything went completely wrong in an unintentional way do I have any kind of recourse to to bad decision making via AI Bots yeah no absolutely and I completely understand that because what we have developed some algorithms to detect fake news basically especially what we have been able also is to predict the outcome of political elections so you can guess how many votes you have and it's about how to detect those boats so it's really about how to develop some very Advanced alos you know for to estimate the credibility of the information and I can even tell you now what's going to be next is also fix pictures you know you and now it happens you know already and about how to detect a fake pictures compared to a real picture so it's about how to to create this kind of alos for The credibility of news or picture also we got more than 25 years of experience in this field because in fact I was one of the first one to create what you call systematic investment strategies you know all those automated trading and you know you get sometimes wrong market price rumors Etc and in fact we think that AI agent is a continuity of all those systematic trading not only on market prices but now using text analytics and you know pictures satellite data or other from drones and youf in terms of what you're delivering to the market what is your sort of feedback improvement process what happens if things go wrong where's the human override um that's why I always double down in thinking of it as another um human agent right which is if we don't have that tolerance it just does not make any sense all right because the way that actually the infrastructure being built which is I would say large language models the margin of of mistakes is the same as the margin of a human being it's just about us understanding that and then build the guard and making sure that even if something goes wrong it's like it's monitored right and then maintained and isolated and then kind of like you you you basically retrain the agent into a better job next time yeah so that's that's that's how we how we see it yeah and I think just one last reflection around regulation I guess the challenge we're starting to see is that regulation is really slow to follow what's already in the market day toay we see this accelerating Beyond exponential Europe itself may put itself at a competitive disadvantage potentially because of the AI law which we haven't even seen the application of right so your question around personal data and how all of that fits in as well is critical and at the same time we know that if we don't have the right Global principles from a regulatory perspective we really will create pockets of challenges and we are all Global Citizens by now and if our data is being used in different ways how do we as individuals have the assurance that our wish is still represented in the execution and maybe one last point if all the AI Bots are replacing us there's yet another question around that right um but fascinating conversation I think we're coming to the end of our time thank you so much for very insightful perspectives and I do think we're really at the beginning of something very big let's maintain the human principle of human override and control to the extent possible and let's manage our human ego to make AI applicable in the best possible way going forward so thank you very much [Music]