The Unstoppable Rise of AI - Regulation, Innovation and Litigation

[Music] morning everyone thanks for joining good morning um introducing uh my co my panelists here to speak on the Unstoppable rise of AI regulation Innovation and litigation to my left I have Jane Horvath from the law firm Gibson Dunn I have in the middle uh Baghdad gas the chief data officer of madad holding a diversified conglomerate based in the UAE and at the end I have cfir Goodrich the managing director and CIO of Black Rock who focuses on Innovation so thank you for joining me uh I'm excited to to share everyone's insights I think we have a diverse uh group here which is on theme with uh abui Financial week uh to start I want to lay the groundwork and and make sure that uh you know folks have a grounding on this idea that uh the UAE has stated um that it intends to be a global leader in the development of artificial intelligence I'm sure many are aware of uh not only the Investments that uh the UAE and its its relevant uh asset managers have made investments in leading Frontier models developers but also the work that local companies are doing here which is included tii uh and g42 and their work in uh pushing the edge and and uh bleeding edge foundations in technology that are being recognized globally um I want to start uh by talking about though uh how AI we read all these promises and claims names how it's already affecting the industry I think folks here are aware of uh software development people have been using codexes which are uh tools that are that automate software and uh software development as well um customer service uh response agents people have read about AI agents and Bots um uh I'm going to start with you Baghdad can you speak a little bit about this idea of about empowering Workforce and how AI enables uh companies and individuals to meet their Ambitions and in particular within uh the region the the Ambitions for Workforce empowerment uh without necessarily leading to a displacement what are your thoughts on how you balance those those twoh I don't think they can hear you you have to you want to try again hey oh I think there f sorry so um so basically uh we are talking about probably uh digital Workforce okay so the idea is mostly to empower employees to um have a a co-pilot right I think that's what everyone is using right now is an assistant that uh have the AIM to just take over those repetitive task and fastidious task that no one is really interested to do uh by doing this basically uh most of the employees are able to focus on more I would say added value uh uh task and and and focus on more reasoning capabilities than just repetitive task um on the other hand and that's mostly maybe for the teris sector from uh a manufacturing standpoint and more industrial side there is the idea of embodied Ai and that's more about how can we have within the UA for example um something that we can call autonomous manufacturing units so those are mostly highly robotic intensive manufacturing unit that operate mostly uh with robot obviously that will bring or I'll say reduce some of the dependency that the UA might have with uh blue color um but obviously that will create also jobs on the engineering side for the uh you know local population that uh will be more involved on added value task thank you uh you know there was an MIT Economist uh named David otter which pointed out that 60% of today's jobs didn't exist just 80 years ago so there's this Rapid Cycling and development as you point out Baghdad about new opportunities but I want to focus a little bit more on displacement and in particular regions and Outsourcing so if you look at countries like Bangladesh and South others in Southeast Asia which you know have large Outsourcing practices focused on call centers um and and the idea that as you automate or create these uh quick heuristics Ro whether it's robotic process automation or intelligent agents that these manual tasks could disappear and what happens uh to that impact and that labor um kafir I think you have you have some experience on this as well the idea of automating tasks within your organization can can you share with some of your concerns and how you balance this idea of empowering the right people without necessarily disenfranchising or displacing uh the the human capital well I think um for the last maybe five to seven years you see this doomed articles all over the world they are saying you know 100,000 people are going to lose their job in asset management and you know all kind of you know uh um statements like this and these are you know statements that some people have some opinions um but they they are very irrelevant because we live in a times that things are changing so fast that an artical like this become uh irrelevant U you know with the speed of light I'll give you some examples uh BD that mentioned you co-pilots right U so beginning of the year uh starting actually almost like last year uh there was U data published uh for people developing on GitHub and saying you know uh 40% of developers efficiency uh you know productivity is getting through uh through the co-pilot there um couple of uh weeks ago I've been on a stage with um um you know one of the executives of GitHub and they mention already 55% productivity so from the beginning of the year already that changes it was not even published at 55% improvement with co-pilot that's awesome what does it mean it means that somebody is doing some things on GitHub and he more is more productive what does it mean for the company he's working for well um the number there if you're going to look into it it's maybe 10% how can it be such discrepancy between the numbers well because Developers in um financial institution or any kind of place that they need to learn something from somebody in order to code they need to have some meetings so about 50% of their time is spend in these meetings Okay then if yesterday uh I was doing uh development and Baghdad knows me and has some experience with my work when I'm coming for him to QC to Quality look into my work he will say you know trusting me pretty much and he will go very fast on the code and he will give me approvement approve to go into production now um I'm telling him that I did that with co-pilot and AI he is going to spend the whole weekend to look at every line of code and every explainability around it so here is how time for an institution is changing versus the number published so productivity suddenly is different so yes you can have Improvement in development of the code you can have Improvement in documentation and quality of documentation with these kind of tasks and then you mentioned these things about you mentioned in one uh Bunch you know this Ai and the process Automation and these are two different things but they are somehow connected and this is something that hyperscaler even understood that in the last year or so and they are doing efforts of that why because we're doing co-pilot we're doing agents right now like you saw the Beautiful song that was put together by Rachel which is uh you know everybody can do that you go to any app and you can do something like that it's awesome but can you cook with that in work to create productivity well not so much you need to take one agent that knows to do just that with another agent that knows to do something else and start interconnecting them how can you interconnect them through some kind of process automation so that's um in some places and some Publications people mention that almost like the uh you know like the difference between atoms and molecules you know you have sodium uh it's awesome you have something poisonous like chlorine but you put them together you create salt that is required in any Kitchen in the world so why this is important because you need something to interconnect these agents in order to create workflow in order to create value in order to get to productivity so we're right now in this run that we're going between co-pilots with agents the next thing is going to be We Believe 2025 the story of co-workers which they are going to be more integrated things to your point you know me on how to get value out of it so this is where we are right now in the journey uh if it makes sense that's very helpful and I and I think uh you pointed out that sometimes mentioning AI leads to more human review I was speaking to a friend who uh leads uh an AI company uh actually based in the region and they've actually their salespeople have dropped mentioning that AI is in their solution because they found that the minute they mention AI is there people worry about hallucinations and start to want to check the homework too much even though they're using machine learning yeah but as a friend you immediately told him no no you need to have this disclaimer develop with AI iwise they going have a leg issue well heading that's a very good segue into the regulation um and and and legal issues so Jane I'd like to start with you and and uh I'm hoping you can share some insights into how different jurisdictions are approaching AI regulations are you seeing Global themes or are there significant differences based on each country's uh perspectives and Ambitions sure and I want to thank the organizers for having me here today um I it's we're in a very very interesting spot with respect to AI regulation right now I have spent my career as a privacy lawyer I spent 11 years as Apple's Chief privacy officer so I have been in the trenches on privacy regulation AI is also not new AI has been along around for a long time I had a team dedicated to working in AI at Apple long before the headlines in November 2022 so um the interesting thing that we're seeing right now and um for many of you in the room maybe familiar with the Brussels effect um the Brussels effect was very very evident with gdpr gdpr was promulgated throughout the world um you know some might argue that the basis of gdpr was the Fair information practice principles so it was easy to make laws across the world with AI we're seeing a real hesitation on regulation right now so the EU prior to November 2022 had been drafting an AI regulation and then once uh once the announcement of chat gbt put AI on the front page they really sped up and they wanted to get this law out anticipating that they would have the Brussels effect that they would be the first and in fact when the EU AI Act was was passed um one of the EU Commissioners texted the EU is the first in regulation waiting to see what would come but what we've really seen is a hesitation worldwide on passing laws the EU AI act which we'll probably discuss a little bit more later is a risk-based approach but it is it it is a law it contains uh significant penalties and what we're seeing throughout the world right now now is is really a hesitation in principles Singapore has put out best practice principles recommendations Etc the US has an executive order but we understand that the president um the new newly elected president president Trump will most likely be cancelling that executive order on AI so we're really in a moment here right now from a regulatory standpoint with one footnote there are existing laws out there that can be used to enforce against the most egregious risk of AI and so I think that that is the you know the one caution I want to have is people think oh AI is not regulated I can do anything well that's actually not in fact true there are so many different laws out there including copyright which you already discussed that will come into play as we're looking at the propagation of AI yeah and I think it's important to note you you highlighted it but uh for for the benefit of the audience uh President Biden in October of 2023 passed uh the executive order which was meant to be a start really in the in the US uh it it amounted to uh what I'd call maybe very light regulation on some of the Frontier Model developers to show what went into or or to verify some aspects of training uh but then it was largely a homework exercise for the federal government so not to give you all a a a civics lesson in uh US law but an executive order only has the authority to control the federal go government and so much of what the executive order did was controlled anybody who sold to the federal government you know these were requirements for entities that worked with the federal government and then it tasked different parts of the government to conduct studies and then as you point out though as as some of you all may have heard there was recently an election uh in the US um and we anticipate uh a number of changes uh one of which the the incoming Administration has been very um uh expressive about which was undoing what they thought was a was a u poorly drafted I will note it is the largest in terms of volume executive order um ever promulgated in the US 111 Pages well over 100 uh pages and uh the Trump Administration has essentially said that they're going to be undoing uh large parts of that the reason I I want to focus on that a little bit is as you mentioned the Brussels effect how one regulation can have broader effects um I think it would be a misreading uh to say that all regulation of AI in the United States where a lot of the large uh language models are trained is going to disappear uh I I think you and I probably will agree on this idea that California will come back to the fold so maybe you know being uh formerly from Apple can can you speak a little bit about what California almost did didn't do and now what they might do absolutely um so we also we have a federal system which is the overarching federal government we also have the states and um we fully anticipate that we have blue states red States Democrat Republican controlled States we very very much anticipate that California and a lot of the Democrat controlled States will fill the vacuum left by the federal government they've already done that in privacy so California has already passed a privacy act they have within that Privacy Law which is common in most privacy laws rules on automated decision making and you know if if a process is purely made by in automated decision making human beings have certain rights and so California has just publish proposed regulations on automated decision-making under that act I am you know I I would encourage everybody to read it and comment if it impacts you because it is very broad the definition of automated decision-making incredibly Broad and I would argue actually covers a lot of things that are not actually automated decision-making the other thing California did is they proposed an AI safety act and um many many uh providers and AI developers said that that would have essentially killed opsource AI it uh required a kill switch on all AI being developed it also required Downstream liability if you if your AI is distributed um which if you're an open source provider in the truest sense of Open Source it's very difficult if you're liable for the downstream distribution so I do think very similar to privacy in the United States which right now uh is 27 States or 20s something states have their own privacy laws that are all not consistent we're going to see that in the states Colorado has passed an AI act um I think that's where the vacuum will be filled in the States so yes to everything you said I also want to point out interestingly though that uh State build 1047 which which you refer to uh with the kill switch has has one very uh notable proponent uh which is Elon Musk and uh I I think that raises the question of when they're signaling that you're going to tear up the executive order and there's going to be less regulation uh we we have uh president Trump's uh what we'll call maybe Chief technological advisor and a lot of other things uh who was a a noted proponent of State Bill 1047 so uh to be written around regulation I want to take it back though on on the idea of Regulation to what exactly is it that we're regulating this idea of artificial intelligence which really started as a philosophy of creating machines that emulate the way we thought humans think um you know we we have uh Regulators all over the world who are not necessarily very proficient in technology uh who are trying to regulate something that even basic researchers right now this week uh there's a huge conference uh called NPS uh going on I bet you if you wanted a precise Definition of artificial intelligence and ask 10 researchers there you'd get 10 different answers so uh with that I'd like to ask our co-panelist uh if here in in Baghdad um you know what exactly are The Regulators regulating do do you think they really understand what it is they're regulating and in particular I want to highlight cfir you you pointed out when I mentioned RPA and and automation the distinction between uh what we call deterministic and non-deterministic uh and and this idea as Jane pointed out machine learning has existed for a while what's so different technologically between you know generative AI this new type of artificial intelligence and in your views both of your views Baghdad and cir do you think that should be regulated whether generative this newer strain should be regulated with the same light and view as the classic machine learning or is that going to cause more confusion so um indeed um AI it's about 68 years old so um but but um November 22 was something that changed a lot of stuff so the story of deterministic and non deterministic and why this is important so it is crazy that in the legal Forum I need to talk about uh uh you know um the theory of systems but but this is core to u to to uh lawmakers it's uh almost I think um like the idea for those of you who read all kind of fiction about the x makina that got into the machine so that's kind of exactly what it is you have a box and uh the box is predictive you put something at the entrance of the box and you get something at the other side of the box and you know exactly every time what you're putting inside what you're going to get out of it that's awesome how about when you're going to put something in that box and every time you're going to get something else um it's kind of going to be a problem for the people that they are overseeing uh this whole process happening so lawmakers they cannot live so well with these kind of environments when you're suddenly putting a box uh you know and have every time different results and um you know to give examples like this you know you guys are familiar know pricing it's something that is kind of deterministic you go and you create a price and you understand how the price is getting done it's formula uh we have you know models we have model committees and financial systems every kind of model you need to put because in our case we're you know fiduciary we need to explain everything explainability is very important Regulators every everyone is looking the bank Regulators are looking at all of your algorithms abut you have to show your homework everything every Super so therefore you have model committee and it's via doosa you go through the model committee doesn't matter what kind of brilliant scientist you are you're going to be asked all the questions and it's going to be very uh tooro now um you know markets are a little bit more problematic it's almost like uh markets are uh you know um you know the the the the the technological twin I'm not saying digital technological twin of network channels where you have all kind of things happening there uh and uh therefore you have this definition of signals to noise ratio snls right so kind of understanding what is real and what is not so much now um I think the best understanding of um what's the danger with this kind of a nondeterministic think um it came in the middle of the year and the magazine nature I'm not sure if you guys seen that this idea that with this AI we're going to generate a lot of data that is um uh kind of uh machine data uh it's kind of machine generated data and you start using these things and when scientist starting doing that more and more with this synthetic data something happened to the models and the models they suddenly they don't work so well they be behave weird so there is this notion called Model collapse I I think that's the most interesting thing for lawyers and for people that they try to do regulation uh to to look into it look um I I I had the opportunity to have a chat with you know people that they work uh you know uh in uh uh the area of U you know regulating things or taking care uh of things to be right into uh uh AI companies and um suddenly they got fired they say we don't need you guys anymore the machine knows to do it better than you do and these people told me that's true the machine they are doing a way better job than uh what we do for now later we'll see because things are changing so what these people are doing right now believe it or not they are advising to the EU committee to the all kind of low things in in order to understand what to regulate so I just had last week a discussion with the panel of uh uh lawyers that they were looking into what's coming out towards us and U you you're seeing you know all kind of uh people in the area uh of uh credit for example that they use you know customers data without the approval of the customers plugging that into the AI models and that that's kind of not exactly right thing to do so what they are doing they are putting a human in the middle in order to deal with that and basically oh you know I did not let the machine use that data I actually had somebody in the mill so this whole idea that once we called immersive Computing it's suddenly becoming a requirement in order to deal with the low so that's kind of some I think some cool angles of the complexity that you guys talking about well there's a there's a further level of complexity and this intersection of laws that are purporting to regulate a technology and so Jane I want to focus uh on on your expertise in privacy and understanding privacy's role in all of this uh kfer mentioned that this I idea uh of of Technologies there's the European the gdpr notion of the right to be forgotten which assumes that when people collect your information you as a as a as a data subject can simply request of them uh to delete your data that's a very sensible uh request and idea the challenge is when you're building uh large language models these huge AI models uh they are not simply databases you can't simply go to a cell and delete uh delete a a a data uh row or element and have that reflected in the output and I think the regulator sus already that people basically companies forgot to delete or that storage volumes they were kind of lost in that you delete I I think this is the technology it's disaggregated you can delete the training data and it has no effect on the way the model is able to work so Jane how how are The Regulators thinking about this what what's going on how do you how do you comply with the law in a world in which the technology has moved Beyond well this is this is the number one question right now I mean more fundamentally there is a very large conflict with gdpr and training data so we know that AI as we've talked about needs huge amounts of data to be trained and the best way to do that is to go out and scrape the entire web to train your model well right now the group of regulators uh the European data protection board which is made up of all the European Regulators are considering what legal basis is required for taking that personal data scraping it off the web because obviously when you scrape the web you're going to pick up some personal information so what legal basis did these model providers and developers have to process that data to train their models because under gdpr you have to have a legal basis one of the legal basis is consent the other which we would argue would be the better basis would be legitimate interest where you do a balancing test based on the potential risks to the individual to and the benefit to the company you balance it your balance always needs to be that the risk to the individual is lower um the edpb right now is considering whether consent is going to be required for process processing this personal data to train a model should they come out and the opinion comes out on the 23rd of December which is an interesting date right before the um the holidays um should they determine its consent one would argue it would be next to impossible to train a model in Europe so then that takes us once you've crossed the Rubicon to having a legal basis to process data um you have to then then comply with gdpr one of the requirements are these user rights the the right of the user to edit their data delete their data see the data that you hold against them or about them and open AI this is very public has been under investigation by the guarante which is the regulator in Italy and this was one of the initial claims that they violated user rights when they were training and and early users and even now you if you query chat gbt about yourself you may find and it produces something different every time because it's a it's basically looking at logic you may find that it produces erroneous data about you I did I did ask it to write my bio and it was a very well know well written piece of fiction um so I say okay I have user rights to fix this it's wrong but because they've trained the model the model is only looking to see what is logical in its output there's really no way to correct your data so what open AI has done is basically told the model not to produce data about a potential individual but if you go down this Rat Hole how do I tell open I AI that I'm the Jane Horvath that doesn't want data produced about me I mean my my name is rare but it's not unique I'd have to start providing huge amounts of data I'm the J Horvath that lives in you know blah blah blah blah blah in order to actually figure out to tell the output of chat gbt not to produce output about me so we are in absolutely a a very difficult situation legally gdpr comes with very large fines 4% of global annual turnover and I would say that you know a lot of jurisdictions have copied gdpr so this is not just going to be a gdpr problem but it's it's going to be one that's going to need to be worked out from a regulatory standpoint so we have personal data as one of the issues in in AI model training Baghdad uh I want to ask you you know I know you're familiar with um how how uh non-english language models have been trained can you uh set the table and and explain maybe some other issues that come to play with uh training whether it's you know Arabic language model or just model training in general outside of personal data well I'm not sure there are some specific challenges fac on the language small on the can you hear it's more tied on the volume of data that you can have access obviously English Corp are way more I would say available but at the end of the day training a model if you want is like a little bit like doing product development okay it start with the problem statements C da identify breaking down into I will say sub problem in which you're going to start to look at what data do I need to collect where is the data how can I enhance the quality of the data and obviously when you are in an English uh uh Corpus those data sets sometimes already exist in other languages not necessary so you do a lot of work related to data set collection and and data preparation and that obviously impact the uh accuracy of the model and that's why uh we are more to what we call now a data cetric approach to AI That's why somehow the European Union is trying to push into this direction of having highly created data set because they can prove that you can lose you can use way less data and still have very good model rather than going you know everywhere and take a lot of data and just one anecdote on that maybe you guys are familiar with the interview of mamachi the former CTU of open with the Wall Street Journal she's being asked like you know Sora is doing great what kind of data did you use and she was oh you know public available data what does that mean public available data right and then she was very very confused and people you know make this meme do not ask a man his salary a woman her age and an AI engineer what kind of data they use like it's it's like a dirty little secret that you know and it's so sensitive I yes uh you know I as someone who was there in the spring of 2022 and was helping curate and train these Foundation models I I would say there was there was a a um there was an understanding amongst the AI research community that if it was available and there were sets you know whether it was lion 5B or other data sets that were being available for training uh the researchers which were really mathematicians uh found these data sets and said we're just going to start training I think uh lawyers and others got involved a little bit you know later in the process but but I think you're right it was this idea uh that we can train and I think those those of you all that were here uh for my introductory remarks there is is a concept in many copyright regimes about using facts and ideas and basic data but there is a tension with copyright law um uh cir I want to ask you because there's really only I don't know I'm just going to call it 10 companies that are fundamentally training very large language models so this idea of like scraping the whole internet while intellectually interesting to us really only applies to a small number of companies 99.99% of companies and entities are going to be uh using these models to make their own organization smarter in this process called fine-tuning taking your own data and making your own uh version of a model smarter can you talk a little bit both about this idea of fine-tuning but also how to think about the data that you can and should be using you you mentioned you know a cognizance of that so yeah I think um you know made good point you know it's about um the effectiveness of the whole process that's what matters at the end of the day and uh large language models they are not necessarily the best thing that you need it's very cool for writing a song uh like this but at the end of the day um you know for example um we used u in a certain application you know open large language models and they were awesome and they were bringing us to about 50% uh Effectiveness for the process that we wanted to use I cannot do any kind of there is no value for me 50% Effectiveness yeah if Jane and I by the way answer our clients and say with a disclaimer I'm only going to be 50% right we'd be 100% fired exactly so then you know you start using closed language models and suddenly um suddenly you're you're jumping you know 30 40% in your Effectiveness or it's bringing you into the uh range of possible but then you know you mix data that it's a little bit more private data a little bit more things that they are Irrelevant for for you and things start you know it's you you're in the you're in the right directions there but I think for a lot of applications in particular for things that you know uh you know a million years ago which in reality was just five years ago people were talking about age Computing for these kind of applications they are a little more remote a little bit more segregated medium um you know um language models you know might better or even small um you know small language models might be uh better that companies like Mistral understood these things and others and basically utilizing uh data sets that they are a little bit more relevant to you can can bring you more value for For What and there is the there is obviously the cost of the computer oh but of course also in the moment you know we are currently very focused on developing on large language models you know Jensen hang me uh you know we need to have ai million time faster better than what we have right now in order to get there let's let's let's bring this to reality uh we need to get to a GPT 4 pretty much to a gpt7 to get closer to an AGI uh you know uh Direction you know but so therefore everybody is talking about um you know training models big models let's train very few people talking about inference and that's what it means that this is why currently we are very little uh into actual production and generating real you know money with with AI you know I think my hypothesis is because no one has explained the technology so I want to go back to this theme I mean c raises a really good point people got excited about the capabilities we saw essentially public demon demos demonstrations of chat GPT back in 20203 and it and and no one explained the architecture of AI models in today's cloud computing and and general SAS world uh you you spend an enormous amount of time building these systems but relatively speaking once you've built them deployment is rather low cost in terms of energy and other things um AI doesn't work that way at all you get no discount in the deployment uh in some ways inference is more expenses so as much as we've talked about the entire Corpus of the web being scraped and how much energy that takes to train a model I mean these systems will run for weeks months at a time with 10 203 now Elon has 100,000 uh gpus hooked together right these systems are running constantly in order to train a model I think what cafir is pointing to is there is no discount for inference which means every time I ask chat GPT a question or any of you all ask it to uh Sora is a great example today I think they released Sora today the video model from chat GPT it is enormously expensive in terms of of compute cost to generate a video I can ask it to create a video of uh a Mastadon roaming the planes in Africa and it will do that for me but enormously expensive uh in terms of inference cost so that that's one of the things where I think once we get to a global glossery and everyone understands the terms we can address these challenges better I I think it's also because the way we've been discussing about AI is mostly tied to compute so every time like we have the biggest GPU farm so we are the most efficient and in the other other side I would say of the of the world uh in China for example there are building AI model that are more frugal because they don't have access well they've been subject to export control laws exactly yes and that bring us to the conversation the export control because when Regulators try to regulate AI the first pieces that seems to be obvious to regulate is the hardware right you know where it is you can track it there is a clear supply chain uh a limited number of supplier and I would say an extreme concentration of supplier and therefore going into the different layer which is like compute data and model the obvious one is is is compute and that's the reason why I will say companies like open AI with the deal that they have with Microsoft and other is okay let's secure the biggest amount of of gpus because that's going to guarantee my lead in Ai and turns out the efficiency of those model very recently started to plateau so when we enter into reasoning capabilities it's not about having the biggest amount of data and the biggest amount of of compute now we are getting a little bit more into reasoning capabilities because it makes sense how can we build intelligence out of strength right intelligence should come in a very I would say refined and crafted Innovative way of building those algorithms uh and on that note of the machines needing to better Network and better conference that we raise here at uh a a global Conference of ourselves I'd like to thank uh the panel for joining me in this uh exciting discussion we're available for questions afterwards should anyone have any thank you everyone [Music]