SBTB 2023: Anthony Annunziuata, Keynote: AI, Quantum, and AI + Quantum.
Recording: SBTB 2023: Anthony Annunziuata, Keynote: AI, Quantum, and AI + Quantum.
[Music] hey good morning everyone I'm Anthony and uh thank you Alexi for the introduction very generous I wanted to start uh just very brief background about me personally I am a physicist not a software engineer for anyone who knows me that becomes abundantly clear pretty early on uh but I have been working and leading teams uh in software development in product and so on for a little while so I'm trying to impedance match to that world and maybe some people here can attest to progress there I will say that recently I've gotten a lot more comfortable with things because if you look at llm outputs the non-determinism there and uh you know their General kind of predictability within an envelope of unpredictability smells a lot like quantum mechanics so to me that's really good stuff and makes a lot of sense to you all of you maybe has some new challenges so today I uh thought what I would do is go through uh a bit of the history of what I've been doing uh including with a bunch of key people that are here and under the theme trouble trouble that scientists like me can cause Engineers like you now we're at a very interesting time right because speaking of trouble we can cause I can go find eight or 10 lines of python or maybe a little bit more and creates uh what looks like kind of a pretty capable application right sort of kind of I mean that's where the trouble begins isn't it right it's really easy to put something together that looks pretty compelling but behind the scenes there's all sorts of the usual challenges and all sorts of new ones as well so with that is kind of a theme and you know you could look at that last line I'll get to that at the the final slide or two of the talk I wanted to talk about three major areas first accelerating scientific discovery second uh experience launching a Quantum Computing business and I'm not going to be very business oriented I'm going to talk more about the tech and then I'll leave off with talking about open source AI a little bit of the general landscape as I see it and then uh what a number of organizations are Brewing up to try to really support the community in a much more forthright fashion here first science now if you know science science uh one perspective and lens on it is that it is an excellent big data in AI problem because the problems of science the scale of data collected in experiments the scale of the space you need to explore to find for example a compound that's active uh as a potential drug is enormous right far more than you can uh probe directly uh or even sample efficiently since there's so many on knowns science hit us uh in its importance pretty hard during the pandemic that's when this particular program really ramped up at IBM but the perspective here is science is a really great application area for AI Technologies in particular so let's start with some interesting progress here now it's uh it's very possible and many people are now working on applying generative techniques to molecular Discovery in particular using generative modeling uh to predict molecules that are active uh against certain targets or prime targets using generative AI to predict properties of perspective molecules that could be useful replacing the direct kind of HPC style simulations uh with AI as a proxy for that or mixed versions of that uh protein folding property prediction and many other things there's a lot of great progress here in IBM has uh has taken a a significant role in some parts of this you can see some snapshots here of some of the generative modeling work that we've done uh in uh finding molecular candidates especially for um timely challenges including uh a few years back in the uh the the the race to find treatments for covid now this is all open source by the way and and much of what you see here almost all of the software that I'm going to talk about here is open source that's a theme it's very important to IBM uh and you'll see that's going to extend very strongly into the future um with open s AI here's another interesting example this is using AI approaches to uh now say you have a molecule that you think is useful now you want to synthesize it right so you can come up with recipes you can come up with uh reaction predictions you can actually deploy that directly in an automated fashion to a tool uh to build molecules so that's really interesting stuff and it's also some of the transcription from the pure Digital World to the physical world now earmark that because there's a lot of learn in there which I'm going to come back to when people start to talk about these existential threats of AI right especially in creating bioweapons and the like that's not so easy we'll uh we'll get to that though now this is great we could go through science papers and results all day uh but this is of course an engineering audience so the question is how do you make this um into a product how do you make it robust how do you make it useful across many use cases how do you handle all the details in practice so if you think about an exemplary use case in a nutshell of the space of things what um so drug Discovery is actually a good canonical use case to represent the challenges and opportunities today that is a very manual process there's multiple different scientist uh Persona types involved engineering uh huge amounts of lab work the top flow represents you know very roughly how that works huge amounts of Hands-On huge amounts of manual iteration takes multiple years and lots of money if if you can Infuse Ai and automation different parts of that including right in the uh replacing some laboratory methods with in Silicon methods and augmenting and accelerating some lab methods that are going to be needed no matter what the promise is that you can accelerate that significantly lots of challenges there data challenges compute challenges uh tracking and Ops challenges um user experience challenges many many interesting problems we put a lot of work into building out science platform to try to address these sorts of use cases um and there's there was a lot of interesting work at the infrastructure level at kind of science oriented platform Services level uh and targeting uh a few different use case areas that were important and are important in science now I'm going to uh show you a little snapshot of one kind of app here uh that version of synthesis Automation in production this is a service that we have out and a lot of scientists use uh it has been you know fairly successful this is not a billion user application right this is more of a very specialist application but nonetheless there's been a lot of resonance right in in using digital methods to replace what to date has been very Hands-On very manual very in the lab uh experimental pure experimental uh trial and error type of approaches two people in the audience are going to tell you more about this so both Dean and Carl are right up here up front they have talks later today so you're going to hear more about uh the evolution of this platform and you hear more about engineering challenges in bringing the worlds of software engineering and science together all right so that was an intro hang on tight because we're going to switch gears to Quantum Computing which is really different and a little bit crazy so we're going to start by relaxing a basic assumption ones and zeros it's so abstracted we we rarely even maybe think at this level anymore but uh forget about keeping track of State Quantum Computing eliminates State as a deterministic concept so that's a lot of fun now what do we mean uh okay obviously classical Computing and yes we're calling it classical Computing I know that may sound quaint to you all but yes physicists like me call it classical Computing let that sink in for a second classical Computing is based on ones and zeros it's based on logic at the low level you all know that Quantum Computing is is based on something called the Cubit which is a richer version of storing information that can occupy a Quantum State that's a super position between one and zero and where one cubit can be entangled in a very strange and interesting way with another where its state depends on the other and before you go and measure it right think about um calling an API to look at the State uh it doesn't exist so don't worry about keeping track of State not so simple obviously there's remember back to that analogy I started with there's an envelope of predictability that propagates through uh is the basis of computation whereas specific uh measurements and processes don't have determinism how does this work in practice there are quantum Gates which are classical uh analoges to classical Gates and there's a concept called a Quantum circuit which is kind of a primitive uh program construct that you should keep in mind because that's really the building block that you're going to program a quantum computer with what does it look like so you uh have these strange Gates that can take what look like normal bits ones and zeros and create them uh as these mixed superposition states you have other bits that can entangle them so that's this uh little connector with the plus sign a cot gate and then you've actually got uh measurement Gates so remember until you go and probit this physical Cubit it doesn't exist in a specific State because you've just introduced it in this indeterminate um state in between that you can actually construct uh Quantum logic Quantum algorithms to do potentially interesting things now you might think well how does this manifest in practice with a developer um the TAC that the industry has started to adopt and IBM is all in on this is to try to infuse these low-level Quantum constructs into known abstractions so that you can use similar tools you can program the way you know you do have to learn some new Concepts within that so that's the general take um so you can use your good old friendly python code with some new constructs to to build these programs uh you can uh pretty you know so we've built up a lot of the the software infrastructure and uh Computing infrastructure to do this and you actually can get started relatively easily you do have to learn some of these Concepts like what a Quantum circuit can do in the context of running a Quantum algorithm um and then underlying that is a lot of really good engineering uh that uh that Carl here in the audience in particular was a significant part of and maybe I'll tell you about during a break uh lots of interesting Quantum platform engineering challenges there uh the idea is to present this uh circuit based abstraction so you can think about a Primitives library that has building blocks that you can build circuits lift and you know you have uh subcircuits such as the ones listed here we won't go through through all of this in detail um but you know this is a a key way that we have tried to make Quantum understandable digestible and usable again U just to mention again this is all open source software this is part of uh what we call kkit which is an open source framework in Python primarily for programming quantum computers it's a whole Quantum platform and infrastructure story uh behind the scenes here again very interesting how we take uh the world of cubits right which are these physical devices that you know that that behave Quantum mechanically and then match and and Infuse that into classical infrastructure in a in an accessible fashion um we have a Quantum Cloud you're seeing some resources in that Quantum Cloud think of these as uh Quantum um so these are accessible Quantum Resources you can use to run your Quantum programs and then in the theme of bringing it to the application here's a fun demo bringing Quantum computation to uh your favorite spreadsheet program and you know you could keep in the back of your mind in a few years are we going to see AI generated Quantum accelerated pivot tables as our killer app here maybe I don't know about that all right hey it's real people use this stuff um you know it's got um close to half a million registered users active users are obviously a subset but there's about 4 billion API executions a day pretty steadily and you know steadily growing um Alexi mentioned our partner Network and our community efforts huge amount of activity here wish I could go into it we can talk about it you know during the break maybe um but you know this is a very active program but it is a longterm program uh it is something that is definitely still on a 5 to 10 year Horizon toward productive use all right so final subject open- Source AI which we'll call the latest challenge or opportunity depending on how you see it what is clear something that you all know the community really wants open- Source tools and open source models for generative AI for foundation modelbased AI uh you've seen incredible popularity of some of the open source projects uh you know a couple of them represented here llama for uh pre-trained models and Lang chain for application orchestration so there's a lot of demand and you know you might convince yourself well things will work out fine we know the story in open source you you know get the you know get Community Focus you get resources you get some big companies on board to help but things are a little bit different here the Primacy of pre-trained models as a compute construct right as an abstraction concept is changing a lot of things it obviously changes application architectures and many many things around that many things you probably know there's a tutorial yesterday around this but if you think about take a step back and ask well is open source just going to work here well there are some new challenges right the state of is that open- Source AI especially at the model level is vibrant but uh the State of Affairs is the proprietary models are still better they're still more capable now will that naturally change maybe but you can make an argument that things are a little bit different here with pre-trained models as the as the primary construct of an open source based technology you have new bottlenecks especially resource bottlenecks it takes huge amounts of compute in particular large amounts of curated data uh to make these things and to maintain them so there's some new challenges here as well another lens is policy which we'll get into just a second there are some new challenges and headwinds along with the open source opportunities and this is a bit of a foreshadowing you're going to see more and a more focused effort among uh organizations and the community in this area soon IBM and others are very active here in trying to figure out how we solve some of these relatively newer bottlenecks especially the re ource challenges uh I wanted to just mention a few pieces of the story uh this is a you know again um part of a of a broader hole that we're going to be rolling out along with others in the coming months but there are some interesting projects on the on the on the front TI here uh here is one in principle based alignment so this is uh essentially how you can use natural language based principles uh you know to align a model not just for safety and ethics but also uh to Brand values to uh you know product uh use case space and here is something that is also very interesting um you may appreciate that um mixture of experts approaches versus DSE llms are kind of the frontier it seems to be what gp4 is based on um IBM has been particularly active here we've open sourced some early work and we intend to engage Community uh to build this up in a in a much stronger fashion and if you think about this the modularity here allows for um extensibility to different uh subdomains of knowledge that are very technical think back to that science uh you know the science problem space there's lots of uh science specific technical knowledge that uh is natural language like but not quite and then there's obviously new modalities uh that need to be extended here so there's uh this is uh some interesting work that we're planning to to go much bigger on in in the near future um this is just kind of an example of how this can work the ability to add subtract pair down um these sorts of modular models um and you can you know therefore adapt to specific use case context specific deployment context and so on okay in this space though scientists not the biggest troublemakers we have new troublemakers um the policy realm and the public discourse and I don't know if you've noticed is a little charged and there are uh some significant headwinds to open source Technologies in AI especially open source models um now uh there are various conversations discussions regulatory efforts happening globally but you should pay attention to this you should weigh in when you can because uh there are some players in the tech industry that would like and benefit from keeping AI closed so I'll leave it at that and in closing I want to come back to one more point in that big list of where the trouble begins when people like me can use a few lines of python to maybe sort of create a toy app and give you guys a lot of trouble there's one big one that we're all talking about am I contributing to an existential risk so let's just dive in a slight bit deeper some people are worried that uh gen is going to lead to Humanity's demise how do we think that might happen people are talking about AI helping to build nuclear weapons bioweapons some sort of malicious nanom machine army or a big army of Killer Robots all right uh should you be worried about this well I'm going to put my scientist hat back on here and if you've all ever worked in a lab or a complex manufacturing environment uh it's pretty tough the recipe to build the virus the recipe to um build a nano-based piece of techn technology never mind a nuclear weapon that's not the hard part the hard part is actually building it uh building it getting the materials engineering it testing it right that's not so easy if you've ever been in a lab it usually doesn't work for years that's why PhD takes years for even a small incremental progress to build anything at a physical level so AI will not cause the extinction of the human race the neck are not the existence of better recipes to follow and if you don't believe me go try to build something good luck so keep coding fix real problems you're welcome from a scientist oh uh questions yeah okay happy to take questions thank you Anthony for the great keyote so uh I'm I'm curious like what's on top of your mind uh to make you know open source AI a reality what do you think is the direction of the next year yeah so if you think about what's happening in open source AI there's like a huge amount right there's something like 30 I don't know 50,000 open source models on hugging face right now and growing all the time um there's lots of organizational level interest but um I think what if you think about the history of Open Source and how like open source software was catalyzed there is a big role to play for industry and for organizational level cooperation think about what Linux took to to really kind of win out took a huge amount of investment um collaboration focused effort along with obviously a really robust individual level Community effort so I think if you if you look at the present there are some analogies and uh I do think that organizations like IBM and many others need to step up and directly support in a much bigger fashion the open Community that's building the Leading Edge of gen uh in the open So speaking of existential threats one of the worries about Quantum is it'll be used to break all of our encryption protocols although they've got Quantum safe ones now do you believe that's a real problem and if not are there other problems we should worry about with Quantum so do I believe it's a real problem to solve yes but it's a problem whose Horizon is a ways out and we have plenty of time to solve it and we know how to solve it there are better encryption techniques that we know how to implement at this point it's a significant engineering challenge but there's not a science challenge in how to protect against you know far future Quantum attacks you know so these days AI is putting so much carbon in the atmosphere you know like do you think that Quantum somehow going to help to kind of think more about like the impact is that a lot of the kind of people looking for gpus we there so that's a great question which I'll take a kind of personal stab at answering I personally don't believe Quantum even though as much I'd love if this was the case I don't believe it's going to significantly move the needle on reducing carbon impact of compute and that's because car uh Quantum will really be for a long time um a use case specific acceleration methodology right um the ramp in uh AI workloads um is going to far surpass any negative contribution from that uh in the total carbon profile for a long [Music] time