From the MIT GenAI Summit: Building Strong Business with Generative AI

thank you all so much for joining um this is going to be an amazing panel hard to follow what we just listened to but uh we're very excited we have four amazing people we're going to talk about how to build strong businesses with generative AI um I'll just start by introducing our amazing panelists uh we have Paul here Paul karamov is a partner software architect at Microsoft he's working directly on integrating large language models primarily into the Microsoft Office Suite so Paul has touched you know products that have millions of users and he's really at the Forefront of some of the implementation around gen AI so welcome Paul thank you for coming next up we got Christian Christian cos is uh amazing he founded and he's been leading the generative AI for ads group at meta and his amazing team you know is comprised of Engineers scientists working on multiple modes of gen AI text video image audio multi modal to invent the future of digital advertising but most importantly Christian really wants you to know that he doesn't host an award-winning podcast he doesn't run marathons in his spare time and nobody asked him to be on the Forbes 30 under 30. foreign Delphine is up next Delphine zukia is a senior partner at McKinsey and Company here in Boston she's one of the leaders and I think recently leading now the generative AI practice for uh Health Tech I think if that's correct um so Delphine has a sort of Life Sciences Health Tech med tech specialization and she's got a lot of generative AI experience she's been counseling companies in that space she's done an AIML deployment in the space and she's an MIT Alum so welcome back Delphine and last but not least we have Armin mccridgeon who is a senior principal at Flagship pioneering where he's leading the pioneering intelligence initiative so Armin is expanding the use of AI in portfolio companies and he's applying AI to healthcare biotech and a lot more and most of you may know Flagship as the the place where moderna was born so hopefully we get to hear something around the flagship model in a little bit so Armin gets to see startups at their earliest stages and we're really excited to touch on some of those topics here welcome Armin my name is Sid sriram I'm a second year MBA student also part of the organizing committee for the summit great to see this turnout thank you all for joining uh so we're going to talk today about a lot of wonderful things but over the last six to 12 months a lot has changed in the world of generative AI we've had some amazing Technologies come to the Forefront chatgpt Dolly but today we're really going to discuss how to build strong businesses out of those Technologies uh we'll talk about what success looks like what the risks are we'll talk about startups big Tech and a lot more so I'd like to start by asking each of you how the last six to 12 months have kind of changed in the context of your AI work so Paul can you kick us off it's 12 months I feel like there was like a Cambrian explosion of players in the space everybody started racing towards uh you know productizing open AI models I work a lot with GPT family models so that's what I think about first and foremost but lots and lots of startups as well as now large companies getting into the game and it feels like uh like a large race and then in terms of just my own work we we have products that have tens and some cases hundreds of millions of users and understanding how to impact those users in positive ways becomes really interesting and it's not uh you know it's not a straightforward uh solution to that so it's a huge race it's it feels like it's going pretty quick I'm very excited for what's going to happen over the next six to 12 months I think there's going to be uh some more really interesting announcements and revolutionary advances both in science and in the business applications of it thank you Paul go ahead Christian uh well Kimber and explosion is such a good I'm going to steal that uh that's that's that's very good um I I think three things have changed uh for me the first one is just the sheer amount of attention that this area is getting um you know I spent about three years at meta evangelizing generative Ai and one of the things that you realize when you're kind of in this space is how um this room is above average nerdy and most people that you talk to uh potential customers they're not and so there's been a lot of fear and trepidation and uncertainty about oh you know is AI mature enough um is this really the you know what we should be investing in uh chat GPT and stable diffusion those two things have really put AI out there and created far more attention even inside Facebook for this area than I could have ever done alone though I will claim the credit later on and the the second thing that's changed is I think the the speed at which certain problems that I expected to be hard have been solved in large part thanks to the open source Community things like how do you do photorealistic uh Renditions of people with a stable diffusion type model how do you preserve the proportions of a product very important in ads um how do you preserve perspective when you're doing sort of digital digital ads all these things turned out to be solved by the thousands of people that that are working on these things in the open source Community now and I think finally the third thing that's changed is just the huge amount of fomo that's happening in other Industries outside of my my modest little field where banks are thinking about this life sciences are thinking about this and you know maybe it was myopic on my part but I never thought of the field of application being quite that large yeah it's very interesting and I think probably Delphine is going to touch on a little bit of that in a bit but go ahead Delphine about how your life has changed over the last few years so actually I'll start on the personal front you mentioned I was I was an undergrad here uh way back when uh when you could actually draw draw a neural network on a piece of paper and I try to predict the stock market at the time and it didn't quite work and I was very frustrated so I moved on to Medical Imaging and I was just incredibly excited about what it could do for Physicians but I I realized AI is just not ready to be adopted which is you know one reason I went into business and what I've noticed over the last year is that finally the business world is willing to hear about AI in a way that it's not just a cool technology but there are all kinds of questions and Eric um mentioned them around the adoption of AI and how you should have some design thinking around it and so you know I used to have a lot of AI conversations with cios and heads of r d and now you have them with board members now you have them with CEOs who actually want to understand how it works which you know makes it very fun for and I know many of you are excited to talk about technology and finally we have now a platform but there's a lot of questions and I think you know for us it's very important to think about the responsibility of making sure it's being rolled out the right way links and Armin go ahead great hi everyone Armin mccirchen it's a pleasure being here I was telling Delphine that I think last time I was here probably was 12 years ago and I'm still taking classes in this in this building so it's a pleasure being back um I think I'm going to Echo a lot of the things that I think my co-panelists said but uh just to kind of maybe emphasize a few things I think the level of excitement probably about AI probably has been the highest I've seen in my lifetime maybe I'm just going to use an analogy maybe on all of you since I have a four-year-old he was asking me today about Orchestra and positioning of the different instruments um if you if you for a second imagine that the in an orchestra all of the different instruments are the tools that companies have digital tools to address different problems uh maybe for an average company AI would be French horns nothing against French or horns but what's happening now I think is a lot of companies are actually either just bringing too many French horns they may not need it but they are or pushing some of his French horns to the front of the line even though violins are supposed to be front of the line again nothing nothing bad about it except that's I think what's also happening at the same time I think just the kind of the level of excitement also is lowering the barrier for some of the companies to adopt ml tools well they generally probably would not adopt those tools which is a great thing and then and then the last thing is I think there are problems as maybe Delfino saying there are problems that at least in life sciences now we are seeing what we can solve that before we'd not even think of tackling and mostly due to generative AI I think those are probably the three things that at least we have been seeing in the last six to 12 months Christian and I are French so this is going to be about friends not China I I just learned about this French Roots common French Roots yesterday so if they start spontaneously French people in Disguise I think we're artificial intelligence was invented right yeah wonderful um so Christian if if you can I think everybody um is kind of wondering this so and curious about this can you can you tell us about a couple of uh maybe one or two cool products or features um that your team has been working on um and you know what kind of Market needed solving and why your team decided to work on that thank you uh that's uh that's a that's a big question but and I have to be a little bit cagey in how I answer it for for obvious reasons but let me let me maybe start by what are the big problems in digital advertising today and sort of how generative AI is going to lead us but also other companies like Google Amazon tiktok probably to want to solve the same problems there's essentially two things um that are really really hard in digital ads today one is a purely workload kind of problem and the other one is how do you personalize digital advertising and they're they're largely intertwined today if you ask any marketer or any creative strategist what one of their major pain points are they will tell you it's just the sheer amount of stuff that I need to create because you and I will think um I want to advertise a bottle of water so I'll just take a picture and sort of throw that out there at no because in fact every time you want to create one ad you need to create different versions of it for you know Facebook for Instagram for Instagram stories Facebook feed different things like that so in fact it's never just one asset it's something like 47. and then your manager reads an article on LinkedIn about a b testing so now it's like oh well now it's 47 times 2 because we want to make sure that we have the best one um and then you learn that oh but ads have a very short lifetime so you need to redo everything three times a month and then by the way it's not just Facebook you're also on Tick Tock you're also on all these other things so your life is horrible and generative AI is um generative AI the promise of it is really the ability to sort of do what what uh sort of in a Java kind of way you know you make it once and then you kind of run it anywhere because we can do things like General generative outpainting we can create variations on a theme automatically all of that is what we're building towards in the very long run what you want to do is create ads that are not only liquid across these different systems but also ads that are personalized to different people so obviously all these large companies including meta but also Google know a lot about us as consumers so how do I use that knowledge to generate an ad that is pertinent to you right off the bat and so in the very long term those are the kinds of problems that you should expect these companies to solve yeah that's that's a really interesting sort of uh goal or Target but it helps us move faster but I think there's there's a little bit more of sort of risk when it comes to bias when you have these sort of higher Stakes um things when you get personalized ads and you're putting a little bit of trust in sort of the machine so how do you weigh risk versus um sort of customer engagement or speed to Market or some of those uh I'm glad you used the word trust because I I think that's kind of the big risk for me as a product person obviously in generative Ai and AI in general there's different kinds of risk you know legal societal uh technical risk and there's frankly there's better people that talk about those than than me so you know from a product perspective when you're in AI it's the biggest trade-off is between your own um the speed of your ambition and the level of trust that you're that your consumer that your customer sorry has in you and again this is one of these situations where we're all above average nerdy so we kind of like Ai and we're you know we're eager to try it most people are not like us most people uh you know don't particularly trust AI and if you work depending on who your employer is sometimes you'll discover your your customer doesn't particularly trust you um and so when you're faced with that you know the trade-off is well I can't show up with the solution that sort of says look I'm going to automate and make your life perfect because nobody will buy that and that you know is a recipe for failure um it's very much how do you adopt a a an assist augment automate kind of framework to your product where you begin by earning the right to speak to your customer by saying let me find the the sort of boring tactical but but painful parts of your job and let me use AI to kind of assist you in that process over time you were in the right to talk about how do I augment the productivity uh of of your job and and in the very long run you know maybe you'll get to the point where you can say well look let me automate this for you but the biggest risk is the loss of trust and if you lose the trust of your consumer then it's it's really really hard to get back yeah yeah that's uh it's really interesting and Delphine I'm sure you are seeing companies that are on a much sort of broader scale outside the tech industry perhaps trying to understand what these Technologies are what generally I even is and how they can implement it how do you and if you if you take say medical as an example or biotechnology risk is really important there and assessing trust and risk so how do you sort of convince the people that you counsel that the application is safe while still bringing value yeah and Healthcare is an area that's highly regulated so risk is obviously one that sucks it's an important conversation you know one thing I have to say about Healthcare and life sciences and Armin will will refer to that as well I'm sure is that generative AI if you if you define it more as large pre-trained models the Transformer models that are the underlying technology that's been around for a while actually and um and it's been around where there's already workflows that exist where whatever it's generating can actually be caught by a scientist and evaluated and then they decide you know what to do with it so an example is drug Discovery you know these models are known for large language models but actually you can do very exciting things treating a molecule like a sentence and or molecular compound and then you know words or different molecules what it allows humans to do is to essentially do something that organic chemist was not able to do which is understand the structure the relationship between the molecules um and that's been used now for you know at least three four three to four years by large pharmaceutical companies to significantly accelerate drug development only about 12 percent of what starts in the lab actually makes it to be FDA approved so it's a lot of cost and unfortunately a lot of unmet need um that area there's less risk conversations now we're getting to more the chat bot craze uh which is you know initially people think oh great we're going to start communicating with our hcps with you know AI generated emails and of course the risk Department says well hang on a second are you going to read that email before it gets sent so these are the kinds of you know I'd say fun conversations that are happening but my my prediction is where there is a workflow and there is good design thinking you're going to see it take off like we have where it's you have to invent new workflows especially risk needing to look at things I think that's where you're going to see a little bit of a Slowdown yeah that's that's super interesting um Armin you know feel free to continue pulling on that thread but I'm really curious if you can just kind of explain the the flagship process to start with just really briefly so that everybody understands kind of how startups are born from Technologies at Flagship and how you're implementing that or how it might have changed for generative AI thank you thanks for giving me an opportunity to talk about Flagship so Flagship was founded in about 2000 it's a company that creates companies and by creating it starts from scratch it starts from literally from ideation from exploration uh we don't necessarily go find IP and say hey come actually we're going to incubate we're going to accelerate your development we don't do that we have mostly phds actually and then this we work on ideation and coming up with about 150 Explorations every year six to eight of those converting to companies we call them protocols these are not real companies protocols stands for product companies so we start before they become real companies we are product companies and we have a very well defined process of managing those protocols after about a year some of those protocols will become new cost many new companies and then become growth costs meaning growth companies the idea in our case is that if we trust the process if we have enough people who know what they're doing greatest ideas will emerge we don't necessarily have to start with the greatest idea in the very beginning it's not that it's the one day I'm going to wake up and I'm going to have the best idea ever but the best idea over a period of time will emerge and once it emerges we'll know what that idea is and I can take it forward so then afterwards we do invest Capital we actually will hire CEO Partners to actually run our companies uh who are going to be Flagship by employees in addition to obviously running the company as well now relate to generative AI uh about a year ago we started this initiative and I'm heading now called partnering intelligence the idea is actually to build Central capability in AI at Flagship to help our companies but also start new companies that are more generative uh in terms of the kind of using the AI computational tools and techniques we are also at the same time I think Eric was mentioning about kind of maybe associations and causations when talking about AI so a lot of time I think what we do is uh as probably everyone else uses uh AI in terms of defining what our causal what are associative and try to project them forward in our case as delfino's eluding there is a lot of workers that has been done in Pharma in figuring out how we can actually use generative models in drug Discovery for example now we can use this diffusion models that Dal imagery stable diffusional use to actually create completely new proteins that the world has never seen for example and we have a company called generate biomedicines that has shown it actually published the paper it's available for everyone to go and have a look at at the same time something that we leverage a ton I think many people probably don't yet but we actually leveraged the hallucination aspect of these large language models uh partly because it's very useful for imagination so these are things that probably are not true but these are things that actually can force us to imagine much further when we ourselves could sit at our desk and imagine so we are we are very intentionally prompting these models to hallucinate and hallucinate a lot more than in many cases they are actually allowed to and then we are leveraging it in our company creation company ideation process as well yeah and you know you have these uh sort of experiments or or prototypes at the at the initial stages but do you have a process in place that is thinking about how these I mean I'm sure you do is how these actually become viable profitable businesses and uh can you tell us a little bit more about how it's kind of like a staged process and how you eliminate the ones that you think might not work and move on the ones that do um and they result in successes like moderna yeah so it's initially we in the very beginning we we don't focus on the business models we don't focus on necessarily on the market opportunity a lot of it is driven by science and technology meaning can we actually enable vet Science and Technology assuming that if we are able to do it then we'll probably hopefully find ways to also commercialize it it's not 100 always true but I think if you start from the very beginning trying to actually run due diligence on a market that doesn't exist it's not necessarily the most uh productive place to to spend your time on so we just start mostly with the with the idea and with with the technology what happens over time and see though is is that as the idea grows when we started actually narrowing down and trying to figure out what is the market we are going to go after how we're going to go after the market what are the new business models that potentially we actually need to think about to go after those markets but it mostly happens it mostly emerges rather than us having a stage gate process saying hey now it's time January 1st to talk about the business model or talk about the commercial uh viability of this it just happens over time yeah I like that fluidity um great thank you so so Paul um Armin has talked quite a bit about how startups are born from from these Technologies and from generative AI specifically how do things work in the big text like Microsoft [Music] um can you talk talk us through a little bit about uh building your own large language models versus uh integrating existing ones into say the Microsoft Office Suite how you weigh those trade-offs um and if you if you're able to um I'm interested in a discussion around the different modalities of generative AI so text image video audio whatever you can share there that would be interesting as well yeah sure um so there's uh when you work at a large company uh there is uh good and bad in terms of support for Investments uh such as you know large language models getting them into the products some advantages are there's lots of people so lots of ideas and so you end up benefiting from effective exploration of the possible solution space of the different applications of these models of the stack and you don't have to figure out all of it yourself you can talk to others you learn what they have done you apply it you experiment your own you contribute back to this community and that's a fantastic advantage another Advantage is that uh just there's just a lot of supporting structure right there you need to do user research there user researchers you need to have a to build a user base to do the experimentation well chances are there are already products that have pretty large user bases and you can do that experimentation uh so so that is awesome um at the same time the challenge is that well you have large user bases and large products and there's a lot of stakeholders and you know unlike in a startup it's a little bit more difficult to make sure everybody's got the buy-off everybody's comfortable with the risk you know we talk a little bit about hallucination ethical aspects of these things and uh you know there's a lot of users and not everybody's comfortable with those you can't just throw things over the fence and see what works what does it you have to apply a little bit more control a little bit more process um in terms of what we do um you know you asked about training your own large language models it is not a small investment right so training of this model takes a lot of money uh if you want power and so that's one of those things where it's there but you have to get enough of a buy off for that to actually have but you can um I think that it is a little bit of a question of where you are in that Journey uh because anybody can create a I'm going to call it a vanity startup on top of open AI apis from the comfort of their home office in probably 30 minutes throw it out there see what happens it is probably not going to be super interesting so you have to figure out what is the unique value that you're going to bring and then in case of uh Microsoft Office in case of large products how does it apply to everybody who is using your product because you have different user bases from edu to from education to Enterprises to governments consumers Etc once you figure out that maybe you find an application for okay now I gotta train my own and training is a load right it could be your distilling could be your fine tuning for a specific application things like that um and I think maybe after that you're going to okay now I need my own like really large language models then you're in the business of competing with open AI or different things and that's I I don't know that we're there just at the moment different modalities become really interesting we talk a lot about text there's obviously a lot of investments into images into videos into voice I think real power comes when there's a combined when you create a document in Word text is nice it pops when you add images right you work in teams um you might again text is nice images are great but the large component of these videos and augmenting that with AI becomes really cool so I feel like there's a lot of power in the combination of the different modalities uh and I'm very excited to see what's going to happen with that Christian do you do you sort of agree with that because I think you've mentioned before that you think that text will sort of reign supreme and but you've also mentioned that your team has worked on these multimodal models how do you see that changing or evolving towards the future uh yes I did mention that that I I thought text was ultimately where things where things will go but maybe not for not for that not for a particularly obvious reasons so I'll come back to that um our team sort of is is forced to work on multimodal things because that's the nature of of digital advertising you know when we use things like Tick Tock Instagram reels it's a reflection of the fact that consumers are using these surfaces therefore you need ads on them guess what you know Tick Tock is a video first platform Instagram reels is a video first platform therefore if you want to solve The Advertiser problems that I mentioned earlier then you know what now we have to do video and guess what video is the multimodal format by definition it's got visual aspects it got audio it has content which is usually text and so you're kind of dragged into this whether or not you want to do that um uh the reason why I think that text is ultimately a more interesting modality doesn't necessarily have to do with that it has to do with the fact that and that if you mentioned this um you can text doesn't just mean this is human understandable words what it actually means is that there's many things that you can abstractly call a language that large language models can learn the grammar of implicitly and a language can be anything it can be English it can be it can be French it can also be JavaScript it can be python it can be protein sequences it could be a building anything that you can sort of constrain in in a way where you can express it as code is learnable by a large language model which really means that text shouldn't be thought of as text it should be thought of as a convenient way to represent reality and therefore to generate reality um and and that's why I think that ultimately the the large language model research is more compelling but not necessarily in the context of ads got it and Paul you touched a little bit on um high costs um so for the the first question is do you think that things will remain that way for a while and secondly because of these high costs now do you think that the bigger companies like Microsoft have sort of an advantage over startups when trying to get to the Forefront and do you think there will be like this Rift that forms where the bigger companies rise up uh I I don't think cost will stay the same I think everything we've we've had in the computer industry for the last 50 more years has pointed to cost go down all the time things get democratized things get cheaper new technologies new approaches all of that is going to happen uh at the same time we also know that every new technology initially comes with high cost and so just by nature of that companies with deeper Pockets have a little bit more of an advantage over just a startup but startups have a history of disruption they're going to happen they're going to be the ones bringing some of these new technologies monetizing it and so I don't think it'll be that the Big Deck always has this big Advantage nobody else can touch it I don't think it's going to calcify I think we will continue to see the cycle of innovation bringing us down disruption new Tech High cost Innovation bringing costs down great um I wanna I wanna also ask you a little bit about metrics and measuring success so you're working on several products presumably across the office suite and some of these products will not maybe see the light of day what are some of these metrics that you're measuring on to to ensure that products become successful Revenue lines or revenue streams for let's say Microsoft yeah great question so in in our world a lot of our products have been around for quite some time if you think about Microsoft Word Microsoft Outlook Etc gosh tens of years I think ward has just celebrated its 40th anniversary that seems a long time um so it's like it is a successful product it is you can argue it could be better it could be worse it has more user few years it is a successful product I think a lot of what we talk about is applications of this deck like the way we're thinking about them today the features we're building are they going to be successful are they going to be less successful and in terms of metrics to judging it uh there are obviously some sort of technical metrics how fast are things right um how much hallucination exists in this given model output do we generate the um sort of the right amount of tax do we cover all the languages all that all that is good um but in terms of business success you have to look at things that are disassociated from Tech you have to look at things like user engagement uh you want to if you think about money and monetizing you want to think about your top of the funnel how many users are attracting how much retention you're driving and ultimately Tech any Tech including large language models is in service of that so whenever you're building an application you have to disassociate from Tech and you have to look at your business goals your objectives and make sure that the tech Choice you're making the Investments you're doing are aligned to those business goals yeah that's that's really a good point uh and I think you said um sort of user engagement is an important one Delphine I think you had an AIML deployment that like a med tech company recently and one of the key metrics was I think customer engagement or user engagement um can you can tell us a little bit about that but also I'm curious to hear about how you put these metrics in place when things are much broader and you're touching a lot of different companies so I think one thing that's important to note is the deployment of AI is is not new and you know gen AI is just a class of AI and so is that there's some I'd say truth that we've been finding over the last 10 years of doing these these deployments this is the Quantum black arm of McKinsey uh you know first of all it's a bit of a sobering fact but only release um 28 of what you would call a digital transformation which is just being able to put an analytical product out there succeed so this other 72 percent they get caught in Pilot mode they get caught where the users actually don't really want the technology they get caught because it's a great technology but it can't scale you don't have the pipes to give you the data for example so a lot of time we spend up front is educating Executives around that because if you are excited and if you invest up front you need to know that it is going to be a journey um now that said we absolutely encourage folks to think of the art of the possible and experiment because it's very cheap up front and the metrics are really no different than what you would use for business I mean this is what Paul was saying it's it's really I mean at the end of the day it's your margin profile many companies are willing to trade that off and and get Revenue first before and as they figure out the cost because users are are quite important um and then yes you you measure along the way adoption metrics that are going to give you some signal that the users like like to engage with uh with your product what is encouraging to me with generative AI I would say is you know if you look at sort of out of these transformation it's only about 15 percent of the effort is on data science and then if you add data engineering it's probably another 20 well that's still 30 we can cut um if we have some of these models already available and if we can democratize who's able to Tinker with them now unfortunately the other 70 percent that is all the various jobs that Eric was talking about this is risk this is Design This is the business owners who also need to be around and and take it through so no different but I think now finally uh everybody is starting to ask the question how do I deploy AI whereas before it was more like what is AI and not even understanding the journey yeah that's that's really interesting and Armin I'm curious um about your thoughts on this 28 success metric which um which Delphine just said because that's not what I would have expected uh and also when when you're when your Technologies reach the stage where you think that they will become successful businesses um how are you measuring that so it's kind of a similar question you know interesting I think 28 I probably knew partly because I also worked at uh delphin's firm in the past uh but it's a I feel like a lot of it is at least from what I've seen has to do with cultural issues as well it's how much kind of people buy into it how much people kind of daily daily Drive the adoption of it um I was just going to maybe comment very quickly on the metrics as well and then I'll get to your question um it's interesting because I was thinking uh at least in biology how do we measure actually success of generative Ai and in a way it's probably no different than anything else at the end of the day we are making we are creating proteins molecules at the end they are gonna take them and test them you've got to have a wet lab to go and see duvet bind the very dark or not what's Affinity or not so it's it's not necessarily very different except now you can actually accelerate that process and at least the generation process can be so much faster than it could have been before so at least I think for us it probably metric wise it doesn't change I think exactly how do things now in terms of question about kind of our companies and as they grow kind of how we think about them it's a it's probably a multi-layered uh at least solution that we try to I think put in place one from from uh from technology perspective we always try to think um all these Technologies actually or are the companies that we are creating do they actually need AI generative AI or not a lot of the time we don't want our companies to just kind of go surf in the kind of generative AI hype if they don't have a surfboard they don't know actually how to stand on the surfboard yet so it's uh sometimes it's it's very easy to say oh so I'm going to use this I've seen this it's like a hammer I'm going to use but you don't need a hammer you actually need the right tool that's not a hammer and that's very that's what I think a lot of times centrif will actually provides to our companies partly because we have just seen so much more than every specific company has seen and the second is actually uh kind of maybe if we increase the 20 28 that Delfino is saying hiring the right people into the companies very early on uh one thing maybe culturally I think we have noticed is and this is probably this is not Flagship specific I think this is generally uh probably true is some of the first people you hire in ml in digital in AI is probably going to define the trajectory that you are going to have in that space it's interesting from what I've seen personally before even Flagship it's people don't like to hire people that are that seem to be smarter than they are in many cases it's interesting observation my own but if you if you try to hire people actually create a collaborative environment where they know where to take the company that helps a lot with the transformation in the future and hopefully 28 won't be 28 but will be more yeah I think a lot of people have been saying that um you know the technology is of course a challenge and a problem but there are a lot of other peripheral things um that really make it a success and one of them is uh people and and hiring like you said um I I also thought um it was a yeah a couple of interesting points there around uh how you how you launch the businesses out of there it's it's pretty uh intriguing uh so I just want to ask each of you a sort of interesting question which I think the audience will appreciate um is there something surprising in the world of gender generative AI that you expect to take place over the next let's say couple of years Paul you want to go ahead yeah um I I think the thing that we'll see is that it is going to move a lot faster than we predict um I think um you know when Eric was talking he was talking on the horizons of five to ten years now that's not just generative AI they're just the AIO lab but I think it's going to go extremely quick look at the journey we've gone even in the pure technical level in terms of model sizes the number of hyper parameters um just even in the last six to 12 months uh and every doubling of that represents an exponential Improvement on the quality uh and I think so I think it is going to surprise us how quickly this will move and it sounds like you're excited about the speed I yes yes excited and a little bit worried because uh keeping up with that is not easy it's almost by the time you are done releasing something it's outdated so you got to go really quick yeah that was going to be my question is if you have a concern around it um great Christian uh similar question you know something that unexpected and maybe a concern around it you ask me if I expect the unexpected which which is a neat question um I the the one that excites me intellectually is this notion of emergent abilities of large language models in particular so as Eric pointed out the we don't entirely understand why these models do what they do and neat stuff seems to happen when the models get complex enough and I sort of hand wave when I say complex because it can be a number of parameters it can be training data size it could be compute that you throw at it but they develop these surprising abilities like the ability to do arithmetic the ability to do inductive logic there's there's certain tests like that at a certain point these things emerge for reasons that we don't really understand and we can't enumerate what these abilities are um certainly some of them haven't been codified yet you see similar things although to a lesser extent in image generation models that develop the ability to do background segmentation automatically and so I think building on pulse Point things are going to move so much faster but what's really exciting to me is what are these models going to develop the ability to do without having necessarily been trained to do it and I'll join you with the idea that this is very exciting it's also moderately terrifying because already I feel the first five years of my career could be replaced by an llm so now I'm thinking man this thing's going to catch up to me pretty soon uh so that's that's a little bit scary yeah I think this idea of some people call it AI explainability I think is one of the things it's going to get touched on on a few different panels today it's it's a really important topic a topic people in the past before even generative AI have talked about the black box of deep learning and um it's like a convenient term to use but it's not convenient for long I think so I agree with that um Delphine any anything surprising that you expect so I'm actually quite excited to see what our kids and unborn kids are going to do with AI you know we talk about digital natives I think we're going to have ai natives uh and I actually you know I mean you guys can resonate here at MIT I was highly frustrated that MIT wasn't teaching me things from the textbook initially and I realized after a while no they're teaching me how to think and I never had to memorize anything at MIT and now the educational world is sort of trying to come to terms with that but I think it's very freeing if you don't have to remember things and instead you can get a computer to tell you what you need and then you add and you think on top of it and so I can already see with my kids you know the kinds of questions they asked chai GPT I would have never thought of it I mean my 11 year old ass chat GPT should try it are there more wheels or doors in this world which by the way he's asked my husband and I a lot and you know we both being MIT Engineers we're trying to figure out the answer but try it out see what chat GPT is going to say which you know allow me to have a really interesting conversation with them around logic by the way because it was a logical answer but yet it didn't really use logic I see you're getting your your children all ready for the McKinsey interviews you know we're trying to keep them away from both of our jobs but it's not working I I just want to ask on that because it you you gave a kind of personal example which is really nice do you see any concerns or risks around that if you draw the timeline out to several years or a couple of decades and generations I mean I'm an optimist so yes I have lots of concerns but I think there's plenty of smart people in this world to put enough safeguards around technology so um I I would say that the concern is if it's makes education less accessible which I'm hoping that's not the case but that's always my concern if there's sort of if it's not inclusive enough great great and Armin um what what surprises you I think what surprised me in the last few months is what Christian mentioned actually is emergence aspect of terrorist models in the field that I work in now it's for example we we took just a regular llm large language model and with very slight tweaking it was actually able to fold proteins like what two years ago before Alpha fold or ES unfold was completely impossible to do at the level of accuracy these models are doing right it's not that's not something you just expect out of the boxers models to do very well but they do and there's again this is just we are probably scratching the surface of what's possible with some of the emergency aspects of it especially assuming these models are going to grow in size 10x 100x thousand decks probably in the next next few months or years uh what I'm looking forward to is actually how people are going to be using or developing middle layers on top of his llms and probably providing types of services that we actually don't imagine today and I think it's going to happen very quickly I think these things are also going to emerge badly on top of each other and we are actively thinking what those are going to be but I don't I don't think we are the most creative people so maybe we can do some of these models to help us they're saying but that's what I'm looking forward to trying to see what middle layers are going to come up to actually help solve problems but the the regular base models themselves probably don't yeah I think this the application layer the middleware so to speak plays a really important role and we didn't get to touch on that much today but um uh I just want to say uh I just want to ask everybody to thank our amazing panelists because we're out of time thank you so much

Channel: MIT AI ML Club

Views: 20,658

Keywords: AI, ML, Artificial Intelligence, Machine Learning, Generative AI, GenAI, MIT

Id: CGAU9iQzFfg

Length: 44min 15sec (2655 seconds)

Published: Thu Mar 23 2023