Transcript for:
Andrej Karpathy -

[Applause] welcome to the closing ceremony  of UC Berkeley's AI hackathon uh I want to   call on stage the awesome incredible executive  director of SkyDeck Caroline Winnett thank you   very now you keep that hi everybody how you doing  awesome you ready to hear who won the hackathon   yes you are how many hackers here how many in the  audience oh nice very good all right we're going   to get started because I think you want to hear  Andrej yes you want to hear Andrej yes you want to   hear Andrej all right let's quick run through you  want to hear some cool facts about what has been   happening this is what we're going to do today  we're going to get to our pitches soon this is   some pictures all you hackers did we have fun did  we have a good time I had a absolute blast and yes   there were llamas for sure I was there most of  the time I was not there at 3: a.m. but I was so   impressed with all of you you hacked your hearts  out and I'm so proud of all of you whether on   stage or an audience you're all completely awesome  all right how many people it took to make this   happen this giant number 371 UC Berkeley SkyDeck  which I represent and Cal Hacks educational   program and student organization so I think we did  a pretty decent job of getting this all together   this is how it breaks down Hackathon at Berkeley  directors, Skydeck staff sponsors we're going to   give some love to sponsors as I mentioned we're  an educational program Cal Hacks is a student   organization this is all because of the sponsors  so we're going to give them a ton of love when   they come up on stage you with me awesome okay 140  judges 100 plus volunteers and 80 mentors hanging   out helping everybody let me tell you a bit about  sky SkyDeck who hasn't heard of SkyDeck anybody a   couple of you SkyDeck UC Berkeley's Flagship  accelerator we host 250 startups a year our   accelerator track gets $200,000 in investment from  our dedicated Venture fund pretty cool let me tell   you about Berkeley SkyDeck fund our dedicated  Venture fund investing about 40 startups a year   that's a lot of startups for a venture fund by  the way the 200k investment and who wants to   apply to SkyDeck July 16 I want to see all of your  startup applications coming in that's in a month   and hackathons at Berkeley are amazing student  organization truly extraordinary people who helped   put us this on this event this is of course what  they do they do hackathons they've been doing it   for 10 years they do about 2500 students a year  and of course they reach a ton of universities   how many people here not from cal hacking not from  cal fantastic welcome Berkeley is a place where we   bring great talent in you all are great talent  we brought you here that's what we do that's   what Berkeley hackathons does come to their 11th  big hackathon in San Francisco in October check   them out on social media get on that linked and  all of that okay who's coming to San Francisco   y'all coming yes okay fantastic all right thank  you to our partners all of you who brought your   hackers here including our friends down in the  South Bay thank you for joining us and all the   other great universities fantastic really happy  to have you you want to hear Andrej do you want   to hear Andrej yes please give a huge round  of applause for our keynote speaker founding   member of open AI I need the Applause come on keep  going Andrej come on out, Karpathy, yes big Applause thanks you hi everyone yeah uh so thank you for inviting  me it's really a great pleasure to be here um I   love love love hackathons I think there's  you know huge amount of energy huge amount   of creativity young people trying to do cool  things learning together creating I don't it's   just like my favorite place to be and I've had  my fair share of hackathons so really a great   pleasure to be be here and talk to you today um  so one thing is this is bigger than I expected   when they invited me so this is really large here  um I kind of feel like actually the scale of uh   the hackathon is quite large and I guess like one  thing I wanted to start with is that just in case   you're wondering uh this is not normal for AI  I've been in AI for about 15 years so I can say   that with confidence and uh you know it's kind of  just like grown a lot so for me AI is is you know   a couple hundred uh academics getting together  in like a workshop of a conference and you know   talking together about some esoteric details  of some math and uh so this is what I'm used to   uh this is when I entered AI about 15 years ago  you're working with say when you're training um   neural networks you're working with these tiny  digits from MNIST you're training a restricted   boltim machine you're using contrastive Divergence  to train your network and then you're scrutinizing   these G on your first layer to make sure that the  network trained correctly and I know none of that   makes any sense because it's been so long ago uh  but it was a different vibe back then and it was   not as crazy I think things have really gotten  out of proportion to some extent but it is really   beautiful to see the energy and today 15 years  later it looks a lot more like this uh so this   is I guess where AI is today uh and that's also  why this event is large I expect um so yeah Nvidia   the manufacturer of gpus which is used for all  the heavy lifting for our neural networks is is   now the most valuable company in the United States  and has taken over and uh this is the day that we   live in today and why we have so many hackathons  like this and so on which I think is quite amazing   but definitely unprecedented and this is a very  unique point in time that you're many many of   you maybe are entering the AI field right now  and this is not normal it's super interesting   super unique there's a ton happening now I think  fundamentally the reason behind that is that I   think the nature of computation basically is  changing and uh we're kind of have like a new   Computing paradigm that we're entering into and  this is very rare I kind of almost feel like it's   1980s of computing all over again and instead of  having a central processing unit that uh you know   works on instructions over bytes we have these  large language models which are kind of like the   central processing unit uh working on tokens  which are little string pieces instead and uh   then in addition to that we have a contact window  of tokens instead of a ram of bytes and we have   equivalence of dis and everything else so it's a  bit like a computer and this is the orchestrator   and that's why I call this like the large language  model lmos and uh I've sort of like tweeted about   this in some more detail before and so I see this  as a new computer that we're all learning how to   program and uh what it's good at what it's not  as good at how to incorporate into product and   really how to squeeze the most out of it so that I  think is quite exciting and I think maybe many of   you have seen the GPT 40 demo that came out from  open AI two three weeks ago or something like that   and you're really starting to get a sense that  this is uh this is a thing that you can actually   talk to and uh it responds back in your natur um  natural interface of like audio and it sees and   hears and can paint and can do all these things  uh I think potentially many of you have seen this   movie if you haven't I would definitely watch it  it's extremely inspirational for us today uh movie   her and actually kind kind of presently in this  movie um when uh this main character here talks to   the AI that AI is called an OS an operating system  so I think that's very precedent from that movie   uh and it's a beautiful movie and I encourage  you to watch it now the thing is that in this   movie I think the focus is very much on like the  emotional intelligence kind of aspects of these   models but these models in practice in our society  will probably be doing a ton of problem solving in   the digital space and so it's not just going to  be a single digital entity that kind of in some   weird way resembles a human almost in that you can  talk to it but it's not quite a human of course   but it's not just a single digital entity maybe  there's many of these digital entities and maybe   we can give them tasks and they can talk to each  other and collaborate and they have fake slack   threads and they're just doing a ton of work  in the digital space and uh they're automating   a ton of digital infrastructure not just uh  digital infrastructure uh but maybe physical   infrastructure as well and this is kind of an  earlier stages I would say and will probably   happen uh slightly lagging behind a lot of the  digital Innovations because it's so much easier   to work with bits than atoms uh but this is  another movie that I would definitely point   you to as one of my favorites it is not well  very well known at all it's called I Robot and   it's from from 2004 Will Smith amazing movie and  it kind of explores this future with like human   robots doing a lot of tasks in society and kind  of spoiler alert it doesn't go so well for these   people in this movie and the robots kind of like  take over a little bit uh but um I think it's kind   of interesting to think through and I definitely  would encourage you to also watch this movie and   this movie takes place in 2035 allegedly which  is 10 years away and so maybe in 10 years you   can definitely squint and think about that  maybe we are going to be in a place where uh   these things are walking around and talking to  us and Performing tasks in physical world and   digital world and what does that look like what  does that mean and how do we program them how do   we make sure you know they um that they sort of  do what we want them to Etc so when you put all   this together I think the feeling that people  talk about often is this feeling of AGI like   do you feel the AGI quote unquote and what this  means is that you really intuitively understand   the magnitude of what could be coming around the  corner if the stuff actually continues to work   um the amount of automation that we can  potentially have in both the digital space   and the physical space now I don't know about you  but I actually find this picture kind of Bleak uh   this is what came out when I put a bunch of the  last few minutes of talk into a image generator   and I don't actually like this picture I think  we can do better and you know you have we have   a few thousand people here you're about to enter  the industry and you're going to be working on a   lot of this technology and you're going to be  shaping it and you'll have some active sort   of power over it so I don't know maybe we want  this to look something like this I this is what   I would like um so this is humans animals and  nature coexisting in Harmony and but secretly   this is actually a high-tech society and there  are robots and quadcopters and there's a ton of   automation but it's hidden away and it's uh it's  not sort of like in your face and uh so maybe this   is something that we want instead and you should  feel a lot of agency over what you want the future   to be like because you're going to build it uh so  maybe we can agree right now that this is better   than the previous picture but I don't know about  you but I would hope so because I'm going to be   living in that future I think so the question for  this hackaton I mean a lot of you have worked on   a really a bunch of really cool project over the  last day or two and the question is how do we go   from hacking to actually changing the world and  building this future um whatever that may be for   you and so what I thought I would do in this  talk is go over maybe like my last 15 years   or so in the industry and I think I had a bit  of a window into how projects become real world   change and I have some takeaways and things like  that and that I maybe wanted to talk about so the   first thing that I find really incredible is how  projects that are sometimes very small projects   like all of snowballs can actually like snowball  into really big projects and just how incredible   that is to watch so as an example I have my fair  share of hackathons like I mentioned these are   some projects from a long time ago that I worked  on over the last 15 years or so so I had a little   Rubik's Cube color extractor I put up some game  programming tutorials on YouTube like 13 years ago   and tried to teach people programming for games uh  I had a video games and a lot of them I had this   like kind of jankie neuroevolution simulator which  uh was kind of interesting and unsurprisingly not   all of these projects actually go on to snowball  a lot of this is just exploration you're tinkering   and so actually these three projects didn't  really go anywhere for me I wouldn't say that   it was really wasted work it was just like it  didn't add up to and didn't snowball but it was   still like helping me along the way I'll come  back to that later uh but the game programming   tutorials actually ended up snowballing for me  in a certain way because that led me from game   programming tutorials to a bunch of Rubik's Cube  videos actually that became kind of popular at the   time and this is kind of sparked an interest  in teaching for me and then when I was a PhD   student at Stanford I uh got to teach this class  cs231n um and got to develop it and teach it and   this was the first like big deep learning class  at Stanford and uh a lot of people have gone on   to like this and then after that I ended up  making another YouTube channel which is um my   Zero to Hero series for deep learning and all so  a lot of people like that as well and then on top   of that continuing the snowball the project I'm  currently very interested in is this next class   and what it could look like and how I can make  it better and I'm calling that llm 101n and it's   about building a Storyteller something like kind  of a chat GPT that you can work with to generate   stories and the idea is you build everything from  scratch uh from basic prerequisites all the way   to like kind of a chat GPT clone in the domain  of Storytelling and building that from scratch I   think will be really instructive could be really  fun I only published this on GitHub like two or   three days ago so it's pretty raw and still  very much in the early stages but I'm really   excited for it and this for me is an example of a  snowball it started with like 13 years ago little   game programming and I'm working on a course  that I think will be really interesting um thank you another example from my life I think is the  snowball that I've witnessed with open AI so   as was briefly mentioned I was a founding member  researcher of open AI and so I was there 7 years   ago these are some images that are public of  what it was like um uh working out of Greg's   apartment like eight of us and uh open AI was  founded to be kind of like a counterbalance to   Google and Google had was like this gorilla with  70 billion free cash flow and back then Google   employed like half of the AI research industry  almost so it was kind of like a uh you know um   an interesting setup I would say and we were just  like eight people with a laptop so that was really   interesting and very similar to my background Open  AI I ended up exploring a large number of projects   internally we hired some really good people and  many of them like didn't go uh too far but some   of them really did work and so as an example  here's a project that uh was in an early stage   a very small snowball at in the early history  of open AI someone worked on a Reddit chatbot   and if you come by their desk and you're like I  what does this look like when someone's working   on a Reddit chatbot we're trying to like compete  with Google and you working on a Reddit chatbot   like we should be doing something bigger uh and  so it's very easy to dismiss these small snowballs   because they're so fragile right these projects  are so fragile in the beginning but actually this   reddit chatbot and by the way don't read too much  into the specific details these are kind of like   random screenshots just for illustration uh but  this was a Reddit chatbot and it looked naive   but actually Reddit chatbot what is that it's a  language model and it happens to be trained on   Reddit but actually you could train a a language  model on any arbitrary data not just Reddit and   when the Transformer came out this was spun into  something that worked much better and then the   domain was expanded from just Reddit to many other  web pages and suddenly you get gpt1 gpt2 3 4 and   then you get GPT 40 so actually this Reddit chat  bot that was so easy to dismiss uh actually like   ended up leading uh and snowballing into GPT 40  which we currently think of is this like change   in the Computing Paradigm and you can talk to it  and it's amazing so it's really incredible for me   to have witnessed some of those um I guess  snowballs and today opening a of course is   worth uh maybe somewhere just below 1 billion  or something like that so A really incredible   uh incredible to see some of these snowballs in  practice so I would say a lot of you over the last   two days have also worked on small project small  snowballs maybe and it's really incredible to me   that some some of them probably won't go anywhere  but probably some of them actually will and uh you   should continue the momentum of your projects and  maybe they can add up to a really big uh snowball   and that's really incredible to watch the next  thing I wanted to briefly talk about is this   concept of 10,000 hours that was popularized by  Malcolm Gladwell I think I actually am quite a big   believer in in it and I think that to a very large  extent success comes from just repeated practice   and just a huge amount of it and you should be  very willing to put in those 10,000 hours and   just literally just count don't be too nervous  about what am I working about am I succeeding or   failing etc just do simple B counting of how many  hours you're going to you're doing and everything   adds up even the projects that I failed at and  didn't snowball into anything those add to my   counter of number of hours I've spent developing  my expertise and getting into an empowered state   of being able to take on these projects with  confidence and getting them to work so a few   examples of that um I made this like really janky  website a few uh weeks ago this was a weekend   project and it's called awesomemovies. life and  you you can visit it I think it still works I'm   not 100% sure I wouldn't recommend you go there  it's trying to be a movie recommendation engine   because I was trying to figure out what to watch  on that Saturday and then I was like okay I need   to build myself a movie recommendation engine so I  put this up and one of the tweets that was a reply   to mine was wow that's so cool that you got this  to work in weekend and I was kind of reflecting   on that at the time because it wasn't as amazing  to me and the reason for that was that what this   person is not seeing is that this is my 20th time  like making a website like this uh like and so I   see all the steps that what's going to follow okay  I need a linode I need a flask server I'm going to   write some some of this JavaScript stylesheets  HTML I'm going to spin this up together I need   all I need to scrape all these web pages I need  to extract tfidf vectors I need to train svm   and and all these things are things I've already  done before 20 times I already have code Snippets   lying around from previous projects and I'm just  remixing what I have and I've already done all   of this and so remixing everything into a new  form isn't actually that much work and allowed   me to put this up over the weekend and it's not  that crazy and this only comes from expertise   this only comes from having done it 20 times that  you can do this so confidently the next example I   would say in my life was a Tesla autopilot so um  I was hired to lead the computer vision team at   Tesla autopilot about seven or eight years ago and  uh one of the first things I did actually when I   joined the team was I basically ended up rewriting  the computer vision uh deep Learning Network uh   training codebase uh from scratch in pytorch in  some of the first few months that I entered the   team and I sort of agree with the whole thing  from scratch and that ended up being a kernel   of what it is now and I think to some extent to  some people that looked impressive at the time but   for me it wasn't because I was coming from my PhD  and I spent five years doing stuff like that and I   knew exactly what needs to go into there I need my  training set my evaluation sets I need my training   Loop in pytorch I need my um uh sort of configs  I need my log directories I need to bring in a   resonet I need to put in detection we're doing a  regression classification and so the whole thing   like I'm anticipating all the steps and that only  comes from experience that only comes from having   done it 20 times before and so I think this makes  a huge difference and things that look impressive   are may be much less impressive to you if you've  done it 20 times before so so really try to get   to this point where you have your 10,000 hours  it makes a huge difference and uh just uh yeah   that's it by the way 10,000 hours if you're doing  six hours per day I think this works out to about   5 years uh so it's about a length of a PhD that  you need to develop expertise in an area uh so I   think it's roughly correct that that works out to  about a PhD length the other thing that I found is   actually quite useful is uh to keep the dopamine  flowing be aware of your psychology your brain how   it works and what it needs to keep going and how  to keep inspired and so in particular your brain   is a reward machine and it wants rewards and  you need to give it rewards so what is a good   way to give it rewards and in my practice It Is by  doing projects and work on projects and continue   publishing them and so here I have a web page  snippet of some of the projects I have worked   on in the past and these are hackaton projects  and random projects and not all of them are good   some of them are not quite good Etc but what I  love about project is a number of things number   one I love that projects get you to work on  something end to end and depthwise like normally   when you go to classes you're learning in a  breadth wise fashion you're learning a lot of   stuff just in case you might need it in the future  well when you're working on a project you know   what you need and you're learning it on demand  and you're just trying to get it to work so I   think it's a very different mode of learning that  really complements the breath wise learning and   is very important so 100% encourage people to work  on projects the other thing is putting them up is   actually also like a really good Jedi mind trick  in my experience the reason for that is that if   you're going to put something up you're thinking  about all the people who are going to be looking   at it your friends and teammates and family and  future employers Etc and so that really increases   the bar for your own work and it it makes you  work harder because they're going to be looking   at it and you feel shame if it was crappy and so  you work much harder and you're going to go that   extra mile to make it really good and that really  uh really helps um and lastly when other people   are looking at your projects uh you're going  to get that reward because they like it they   appreciate it they Fork it they work on top of it  and so that feels good to your brain and so the   way that this comes together is you are getting  your dopamine you feel good that way you can   build up to 10,000 hours of experience and that's  what helps you a lot snowball your project from a   small snowball all the way to a really big one and  actually make change in the world so in summary   that's I think how it works like on a high level  and the message is just keep hacking that's it uh and then hopefully uh this is  the feature that we're going to   build together when we snowball all  of our stuff or something like that   uh but not the not the first picture  I showed hopefully and that's it thank you Andrej Kaparthy everybody thank you Andre  that was awesome thank you thank you all right   let's get to those pitches the grand prize coming  up you're going to hear eight pitches by eight   projects filtered through 290 submissions narrowed  down to eight so you all going to see some cool   stuff the grand prize is $25,000 investment an  actual term sheet from the Berkeley skydeck fund   they must commit to hacking all summer on their  project and they must appropriately form of course   a legal entity how do you get money otherwise  all right I would like to now uh tell you briefly   about how this is going to go um eight projects  as I said three minute pitch you guys ready three   minutes yes they're ready um the judges will then  provide three minutes of feedback and then after   all the pitches the the grand judges will go and  deliberate and pick a winner well we show you   some other cool stuff all right I would like to  introduce now to Great Applause everybody please   because we have an incredible panel of judges is  we are so pleased to have them please welcome our   first judge Brian Bordley with the Berkeley skyc  fund welcome Brian Marcy Vu with Greycroft welcome Ninamdi Iregbulem with light speed welcome Nami  Irving Sue with Mayfield fund welcome Irving Kurt quiter UC Berkeley faculty and serial  entrepreneur welcome Kurt and Mark nitzburg   Berkeley faculty and director of the UC Berkeley  Center for human compatible AI thank you judges all right we got eight startups warming  up backstage let's give them a little drum roll   let's give him a little drum roll we're going to  get him going I first have to hear if the slides   up the slide is up first are you ready you ready  are you ready yes please give everybody a warm   Round of Applause they've been up all night  hacking and they're ready to share with you   please welcome the first project revision come  on out come on out [Applause] revision okay oh   yes the mic that would be helpful yeah thank  thank you so um good evening everyone it's my   pleasure here on the behalf of my team also um  for the revision project and my name is Danica   I'm a rising senior studying computer science at  UC Berkeley we have masters of design students   as well as data science students on our team  and we're really excited to tell you about   our project so our project we're focusing on  building an AI co-pilot tool for stem textbook   authors capable of detecting and mitigating bias  in textbooks to create inclusive education content   and there is a reason why we're doing this when  considering overall representation of scientists   across textbooks o only 13.1% were women and  compared to 86.9% men in a 2020 study that   featured seven of the most frequently used biology  textbooks within the US and on average people of   color only appear every 320 pages of text while  white figures are observed every 24 pages across   10 college stem textbooks published between 2016  and 2020 so we thought about this problem deep   and hard and it has been something that I've  seen from my personal studies and starting from   elementary school to middle school we constantly  see different examples of word problems and other   situations where text is always there and it's  not always reflective of the actual true history   and this research has been done by numerous  scientists who have gone through this process   of identifying people creating databases but there  is just no current fix that and no one is really   hoping to create this problem but there is no  current fix that helps address this problem so   the text companies actually is who we our team  identified as our buying customer the current   revision process actually takes six to 12 months  of a committee of five or more full-time employees   working on bias checks and the issue here is that  employees are actually not experts on their topic   they also bring in their personal biases as well  so our tool would come in right in between the   writing and revising part of this entire um this  this entire cycle that developers go through   through when writing textbooks so again here  is our competitive analysis I'm sure many of   you have used turnitin or grammarly when you're  submitting even essays and we really think that   there needs to be an additional check here for  bias and checking gender racial political and   other biases and making this process affordable  and automatic so it's not uh so it's not a cost   it's not a costly process for anyone and through  throughout this process we're addressing supply   chain diversity so starting from a younger age  the elementary school students could be able   to use textbooks that truly reflect the true  history as well as themselves and here is our   prototype so we have our text box here on the  left side of the screen where you get to show   in real time um the examples of like some sort  of text that a writer is creating at the moment   and on the right we have an overall score and  the bias checks for um different categories and   we're using machine learning models on the back  end to actually identify these as well as llms   and I'm not sure if I can play the Prototype  but okay yeah it does play so um essentially you   can click through the different links to see the  breakdown and once you actually um highlight one   of these we are also adding in an API through  Hume API uh through through a couple of the   sponsors here um like such as hume API and more  to actually identify emotional analysis as well   um in the textbook writing and in addition to  this we're hoping to build a chat bot that can   actually help you also get bringing data databases  from the unrecognized scientists and being able to   sort of represent it um because bias actually  exists in three different ways one of them is   through actual like text such as um representing  firefighters um would be nicer than saying Fire   Man and the other way is that um the entire tone  and emotional analysis which is why our team used   hume API to actually um detect that emotional  component and the third one is mitigating bias   so um we also considered adding in the chat  bot so say for example if you want to highlight   scientists that are like for example contributing  to physics you wouldn't just say list a few male   scientists in call it a day we would also suggest  equivalent contributions of female scientists   as well so please join me and our team in  revisioning our future of education and work thank you so I I think everybody in communication  today nonprofit or profit is concerned about   diversity so it seems like you have a much  larger market than just um textbook Educators   um also a comment on kind of like Market sizing  and whatnot I would I would think about you know   potential ways you could expand the market here  because the number of people who are involved in   writing textbooks is a relatively small group but  one way to think about it is like maybe in this   new era of AI generated content a much wider  array of people can be part of this textbook   generation process so that's one thing and then  I would also maybe consider selling directly to   the consumers of textbooks in some sense the bias  you're talking about is internalized on that side   of the equation not on the manufacturer side and  so there could be an incentive there for people   to want to pay for something like those yeah  definitely that's something we're considering   so like the textbook would be our official  buyers that we're marketing to but eventually   it would be more of like a grammarly checker  type of tool that anyone can use yeah I had   a similar comment on Tam and Market opportunity  and as think about just how a textbook gets put   into production that if you actually had it as a  tool for whether it's news or other areas You have   more velocity um both in terms of getting the  data to improve your models but also um greater impact yeah I'll just I'll just like as well  I mean similar I think everyone here is kind   of hitting the theme of how do we think  bigger um so even Enterprises right like   companies setting out communication  internally or externally I know this   this this problem exists everywhere so  that's kind of where my brain would go too okay thank you yeah thank you whoops agent OS please welcome agent os hey there everyone my name is Shashank I'm here  today with my friends Agam and Dhruvsomewhere in   the crowd over here we built today agent OS  picture this you work at a hair salon and   you guys are bombarded every single day and every  single year by your accounting and tax preparation   qualms these are things that are very hard to deal  with and you've heard of tools like open AI chat   GPT llm this chat GPT that everything but you have  no clue where to start using these Technologies   and that's no fault of your own the current state  of the technology right now is very bad at multi-   functionary tasks more so it's very hard as an  individual developer sometimes even non technical   to even get started with even the simplest  automations or workflows or tools with such llms   even Engineers with years on years of experience  in this space take tens of hundreds of hours and   even thousands and thousands of dollars to even  get started to build something this is where agent   OS completely transforms the landscape with agent  OS you're able to create multi-agent workflows in   a matter of seconds from natural language what  does that even mean take your average corporate   org structure you have your managers you have  your workers and sometimes you even have your   interns everyone is really good at what they do  they have their tools their skills let's say John   is really good at charting and making PowerPoints  let's say Steve is really good at python coding   everyone's really good at what they do in this  you have a very collaborative working together to   create this common solution for someone coming  from higher up that's that's how agent OS was   designed our Engineers Dhruv and Ugam were able  to replicate this human collaborative process   programmatically using llms what does this do  this allows everyone from the common Joe all   the way up to Enterprise clients to be able  to interact and use these multi-agent agentic   workflows in their day-to-day life to improve  their quality of life or productivity in all in   a matter of seconds and a few sentences let's  go back to the study of the hair salon in the   process of doing your taxes and accounting you  have multiple steps you have your collection   from your receipts and your invoices you have  calculating your cash flow all the calculations   you have to do you have to manage your workers and  then you also have to do your general summary what   about your insights for the year how you were  spending what you were spending on and you have   to also do a lot of clustering and analytics on  this this is a very complex workflow that's nearly   impossible for modern-day llms at the current  state to do right now you can take chat GPT you   ask it a question for even more than three things  it'll forget what the first thing was by the time   you're at the second it doesn't work that way with  agent OS this completely changes where you're able   to have these complex workflows let's dive into  another demo so let's say I'm an analyst at JP   Morgan and my boss tells me every morning I want  a report of uh XYZ stock in the morning a detailed   report on paper how do I do that I use agent OS  on the screen you can see a bunch of other complex   use cases of multiple agents working together  collaboratively but in the toolbox in the search   bar you can see the use case of the analyst here I  have to do market research live stock data I have   to search the internet go on Yahoo finance then  I have to create my analysis technical analysis   qualitative analysis then I have to do what my  boss is telling me to do and after all of that I   have to create charts graphs and visualizations  here you can build tools using natural language   like the one right there that says write me  a tool that fetches the meta stock price from   Yahoo Finance in a matter of seconds the common  Joe or anyone is able to create that tool connect   them to workers you can think of workers as your  everyday employees agents people that perform   these actions using the tools and then connect  them to Super teams and these teams are able   to on this screen you see four but you can scale  this up to 40 400 basically complex vertical and   horizontal organizations that are able to perform  complex decision- making and complex analyses for   anyone from Enterprise to Consumer what does this  do with the multi-agent multi-team framework this   completely opens the landscape up for anyone and  everyone to take on the power of llms into their   own hands from natural language take your average  Farmer at a farmer's market he's trying to create   his marketing campaign for the upcoming Farmers  Market this Sunday he has no clue where to start   looking at his metrics looking at the customers  looking at the weather and creating these   brochures papers pamphlets and whatnot with one  line and one minute using agent OS he can create   all the documentation he needs in order to enact  this stuff and be able to perform successfully and   continually and grow his business at his Farmers  Market things like this are completely opened up   with agent OS and we hope to completely  democratize the process of using llms at   all scales at all geographies and all use cases  within sentences and seconds thank [Applause] you that's a a compelling proposition  um the one thing that I worry about is   right now um the agents are the the  uh the llms you know uh uh performing   these tasks and there's a certain uh uh  certain question about the veracity and   reliability of what they're doing and so  I I I think this that in a future where   we have that reliability uh this this would  make perfect sense but I I would want to add   a kind of Tandem subject matter expert maybe  looking over the shoulder of each of the agents   I think next time I hear this pitch I'd love to  hear about the one market you're going to crush   uh it's hard for me to imagine uh serving a hair  stylist one day and a Morgan stanley analyst the next this is a huge opportunity and a big  bold mission that you have I would want to   dig a bit deeper into your Tech staff and the  people you have on your team because these are   really complex problems and issues and also agree  that um what would be your first area of focus   because it's it's pretty Broad and wide I'll say  I kind of like the broad focus and there's a lot   of individual startups tackling each of these  individual you know problems if it's invoicing   or research so it might be interesting  to figure out how to like Loop in all   these other tools that are out there and really  kind of just be like an interface layer and let   these other companies solve the the technical  challenges yeah I think the value proposition   of creating multi-agent workflows in a matter  of seconds is really compelling I think the next   step would be trying to figure out how can you go  from Simply performing these tasks to becoming the   best at these tasks so for example going after  the outliers sort of the thesis around coaching   networks um some startups do this and they do  it better for like certain verticals than other   so I think um doing more research around that  could be really compelling only only thing I   would add is just think about you know Enterprise  security and how you solve for that there's a lot   of authentication and authorization you're  going to have to do for all these agents so   just have a answer for that well yeah thank  you so much thank you everyone thank you AGL all right next up Skyline come on out Skyline hey everyone hey so my name is Rajan and I'm  a first year student at the University of   waterloo and I study software engineering and  I fundamentally believe that cities shouldn't   be static they should optimize and scale for the  people who inhabit them and so we built Skyline   Skyline is a uh an an Optimizer and it allows  you to better understand how to model cities   using agents and optimize things like traffic  and and transit to uh to to inherently increase   mobility and reduce things like carbon emissions  so this is a very weird problem that we solved   but I want to walk through through the case  study of Los Angeles so Los Angeles is one of   the largest carbon emitters in North America uh  this mostly because of their Transit because of   the amount of cars and so what are ways in which  we can optimize this well let's look directly at   the people who inhabit Los Angeles we can extract  Census Data things like age we can look at things   like um gender we can look at things like income  uh we can find population density graphs and using   this information we can start to find patterns  specifically what we did is we created 500   distinct agents each agent is a different citizen  with different with different interests and what   we can do is we can give them each their own Chain  of Thought each person here has their own day in   their life for example this person is a very young  uh I believe this was a a 22-year-old with a very   large income he's a long day at work and after  work he goes to the gym we can now reason about   what this person may do and now model this on a  map now once we have how these different agents   are moving around what we can do is we can try and  optimize things like Transit so what what we do   here is we have our own proximal policy analyzer  and this allows us to create simulations on what   we believe to be the best way to understand  um how how we can move around from any point a   be in in in the fastest way at the lowest  carbon cost we use our own carbon cost analysis   mechanisms uh our own machine learning models to  better understand how we may be emitting carbon uh   and and how to reduce this through our Transit  so this is a lot that it is through you and I   think the best way for me to represent this  to you is through a video I hope this video loads it's possible to play the video so what we first do is we have an agent  based simulation these are 500 distinct things   in parallel that are running now they each  go around throughout their day and what we   can do is we can find patterns in how they  move around now the best part is what we   can do is now that they're all back in their  original position we can start a generation of   Transit and we're using these patterns to now  generate live different transit systems that   we believe to be the most optimal so what  Skyline is we're not we're not a company   that does uh analysis of Transit we are a  human modeling company and that allows us   to better understand and better predict how  things around us will change and how we can   optimize them using these patterns um yeah so so  that's Skyline um happy to take any [Applause] feedback wow uh I I just want to uh observe that um what  what you're doing in creating a sort of digital   twin of a city uh is for the the the essentially  the the you know the the um uh each citizen is   being simulated using you know one of these  really powerful expensive things a language   model and um uh it it it will be uh probably  an important step to uh to to draw from the   language model some of the uh statistics that  that are actually fed in in the first place make   sure you're getting what you're what getting  out what you know something representative   but I I that's very impressive yeah similar  comment I think you know there's all sorts of   like economic theory about you know agents and  modeling their behavior and their values and   whatnot and the thing that usually gets you is  a sort of heterogeneity across the population   so making sure that that actually represents  the populations being modeled is important um   and then the other thing also related to Value  I would think about is just value capture for   your own sake because I feel like this is like a  category of software where like yeah the economic   impact of this could be massive but how much  of that do you get to capture as a software   vendor is like less clear to me but it's very  interesting I guess I would be curious about   maybe some more like nuanced Enterprise use  cases as well if it's concerts um or security   or stadiums so kind of just thinking about like  are there more micro use cases that there's a more   direct Roi with um for this sort of modeling  yeah we we we tried to consider ourselves to   be a human modeling software and this is just  one of the most visual applications which is Transit okay awesome thank you so much thank  Youk thank you Skyline all right next up we have we have spark please welcome [Applause] spark  hi how's everyone how's Cali we are spark and   we're giving a voice to new entrepreneurs young  entrepreneurs so let's admit it cold calling is   really hard I mean resources are hard to get it's  a steep learning curve and getting attention is   hard if you've cold called someone you know they  don't have time they'll say oh sorry call me back   later I mean they're busy everyone's busy we have  things to do so we have to figure out how can we   earn the time of working people there's existing  Solutions it's long and arduous for trial and   error it's expensive for a sales coach and finally  it's if you have a sales partner chemistry isn't   easy if you're just meeting them right well  we have a process you upload a transcript   to our software we go through and analyze the  emotion we aggregate this data and we give you   productive feedback who's our target market well  look around you guys are our target market people   who are Engineers people who love to build and  say this weekend you made some sort of product   you want to sell you don't have much experience  with sales or Outreach with our software you can   record your cold Outreach and we can tell you  what you've done right and how you can improve   to hopefully land your product where it needs to  be and later on we want to expand to call centers   and sales staffs because we think we can spread  this across an organization and it can be highly   profitable we have uses usage based teering so  75 cents a minute for 1,000 hours going up to   40 cents for 10,000 so this is our software and I  want to tell you guys a story I started being an   entrepreneur around 6 months ago and we made an  AI body cam analysis startup so I did 100 phone   calls 100 cold calls I got no clients 200 300 500  and 700 no one was responding to me so by 800 I   got actually three and I realized something the  human brain is pretty amazing we're able to pick   up on patterns but at the same time it's kind  of inefficient because it took 800 here we look   at the emotion between every single sentence  and we figure out out spikes of emotion and   decreasing emotions we see that when we talk  about security and data privacy with police   officers it shows an increase in their interest  and this was a trend among many conversations   we had so in our analysis page we see in the top  left that mentioning AI accuracy and efficiency   increased officer safety and discussing cost  savings really helped us when we were talking   to officers because we're some college students  right we're dealing with some pretty confidential   data bringing this up early really helped improve  our rates and the four things you see here in the   corners are the different triggers we generated  automatically based on the cold calls we had so   one is positive reactions negative reactions  escalate or de-escalating tense situations and   normalizing exciting situations we also generate  insights too based on whatever cold calling Trends   you make we we also have a rag so you can upload  your company knowledge your target audience and   your pricing information so if you make a mistake  don't worry we got your back we'll tell you hey   maybe instead of saying this you could have  said this because might have helped you out   a little bit sorry my team picture isn't on here  but thank you to tusar and Nick and Krishna you   guys were a great team and I'm honored to be here  representing you guys today I'm open to feedback I guess I need to be the first person to say that  you're uh entering a pretty competitive uh Market   with other offerings here so yeah yeah I'll say  something that stuck out to me was this idea of   of insights but I I I think you know there's at  a organization there's going to might be a sales   team and a marketing team and an online web team  and those teams don't really talk to each other so   it's interesting to think about how do you  pull insights from like one channel of sales   or marketing and actually bring that into like  another channel so maybe the insights from cold   calling are actually influencing what's going  on the website maybe there's some interesting um   spots of opportunity there yeah I could actually  talk about one facet of this we want to explore   deeply uh I want to give you an example say we  have three founders in the company right I have   a first cold call with one person and later on my  second co-founder wants to set up a warmer call   right in the future and then my third founder  wants to set up a third call we want to build a   profile for this client as they go along so we can  truly understand them and also we want to develop   a profile on ourselves too so we can learn more  about ourselves as we go and how we're behaving   uh make sure that we're learning as we go so  we're thinking if we develop a CRM on top of   this data that we leverage then we can connect  multiple teams and enable cross functional uh   benefit yeah I had a similar comment I think it  would be really game-changing if in addition to   some of the real-time analyses you guys are  doing around sentiment where you can see the   system with information on prior calls or this  person's particular strengths and weaknesses and   how they complement those of the other people  on the team and to really build the CRM this um   Knowledge Graph around each person's strengths  and weaknesses on the team to be able to better   find tune the system yeah thank you you guys saw  the analysis but also there's a long list of past   conversations you can actually go into every  single conversation you've ever had and look at   it deeply the same way you did in the like latest  conversation I would think about the full sales   funnel um this is pretty deep down in it and as  you think about where are you really going to be   able to convert or where is where is the wedge  that really matters because there's a lot that   goes into converting a sale um and it's not just  the cold call so is the issue are you actually   calling the right people or is the issue like are  you actually speaking to the right decision makers   so just thinking more broadly about that funnel  and where you might actually be able to have the   most impact and and have the right wedge into  the broader product weet yeah s thank you thank   you all right thank you guys we appreciate  it bye enjoy your question clicker Clicker   thank you okay next up we have hear me out  please welcome hear me [Applause] out the big   one okay thank you all right guys hey hi my name  is Marcus um I'm from here me out and what we've   built is an AI driven customer service pipeline  optimizing call matching and visibility so that   might leave you scratching your head so let's just  talk about the problem so let me give you a bit of   context I'm an exchange student and when I  first came here I had to deal with so many   things I had to deal with banking I had to deal  with insurance I had to deal with deliveries and   even my I even had my car break down to me  and that was a real pain um in short I was   overwhelmed by the sheer number of customer  service calls because for each of these things   I had to make so many calls just to get things  done and I think a lot of you guys can relate   to that we've all had our fair share of bad call  experiences where we're upset the customer service   representative is also upset and nothing gets  done we've also had good experiences as well   and I think that's the core of what we're trying  to tackle here we want to create a pipeline that   tries to provide optimal matching and provide  visibility on emotional data to businesses so   we also did the research and I think the numbers  speak for itself this is a key problem with a   sizable market and a sheer number of people are  affected by this as well and this is our problem   statement which is how might businesses which  offer customer service calls provide a better   experience for their customers so we think we can  tackle this in four key components first of all   an improved callbot we all are common with that  initial robocall that we have to deal with and   sometimes it's really really frustrating how  many times you had a call talk to you and it   just doesn't direct you to the right person I  think we've all experienced that before second   and third and I think this comes hand in is  just business visibility we want to provide   businesses with better visibility of both call  and uh both call experiences data as well as   customer Representatives bandwidths over the  day as they continue to take calls and finally   the this is where we put those two together we  want to take that data and optimize a customer's   journey through a best through a better customer  to service representative matching so I won't bore   you with this um data but in our with with that in  mind we developed a set of decoupled microservices   and I just want to uh point three key Parts  out to you so first of all um we want to assess   customer agreeability with an initial robocall  but this won't just be your normal robocall we   want to use hume's API to manage the robocall  in an empathetic manner such that it measures   the customer's emotions as they go through the  call and eventually outputs um an agreeability   score for the customer second of all we have  a call analysis feedback loop and that's that   whole thing on the right that goes down below and  what this does is once you have a call connected   between a customer and a representative it takes  in multiple factors of data such as the call   duration the emotional change over the call and  the overall call outcome using hume's emotional   measurement API we can then also generate a call  score finally and this is the third and key part   to this it's the matching API using the two things  that I just mentioned we can best match a customer   to a customer service representative which matches  their Vibe their energy and their emotions based   on how our C our custom model is developed  so what's the outcome of all of this as a   representative goes through their day their state  changes depending on how their calls go and their   bandwidth adjusts accordingly this affects the  subsequent customers which they are matched to in   a positive manner and creates a better experience  for both parties so there's a lot more which we   hand build with what we have but with this  foundational pipeline We believe We effectively   tackle the problem that needs to be solved that's  all I have for today thank [Applause] [Music] you nice thank you yeah I mean a a little bit of feedback similar  to the last company as well just there's a lot of   companies working in this space too so I would  just continue to think through you know how to   find that core differentiation um you know if you  continue to work on this after the the hackathon   yeah I completely agree I think a key part um  that we thought was really exciting was just   what you can achieve with custom models what  we're doing by developing a feedback loop is   we're creating something where we can create  in a sense a model which trains itself we can   assess how calls might improve or get worse  after the matching and that feedback gets   fed straight to the matching API so that it knows  whether or not it's done a good job or not and we   find that really interesting and we think that  that's a key differential factor which we can achieve there might be some there might be  some opportunities for building some sort of   like synthetic data pipeline here where you  could like just sort of simulate calls with   like an AI bot of some sort and um use that  as feedback I don't know how good that data   will be or not but could be interesting yeah  know that's a really interesting thought thank   you yeah I know right now you guys are targeting  customer service agents as well as call centers   um something that could be interesting to think  about is as you think about the different stages   of the software adoption life cycle as you go  from your early adopters to your early majority   and then your late majority who's eventually  going to justify your evaluation in terms of   like what those ideal customer profiles are  going to look like down the line yeah thank   you for that um I think one key thing was like  we actually had a a judge come to us and talk   to us about how they're doing something similar  for sales representatives as well and we found   that really interesting so happy to figure  out how we can pivot if that need arises yeah thank you so much cool thank you thank you   hear me out all right next up we have  dispatch AI please welcome [Applause] Dispatch AI everyone my name is Spike and I'm  with Dispatch AI in the United States over 80%   of 911 call centers are critically understaffed  this this means that in a crisis people with real   emergencies aren't able to get the support  they need because all the human agents are   busy and they're often put on hold this is  particularly an issue in our neighboring city   of Oakland where last year the average weight  time was over 60 seconds to pick up a 911 call   now in an emergency every second counts and this  could be literally the difference between life   and death this is unacceptable and that's why  we built dispatch AI the world's first AI 911   call Operator designed to eliminate these wait  times our system is powered by two key components   first is the voice AI The Voice AI will step into  calls when all human agents are busy and it will   work with the caller to evaluate their situation  extract the location and optionally dispatch First   Responders Direct directly to the scene and  the second part is our powerful dashboard for   the operator themselves so the operator will have  access to a bird's eye view of all of the ongoing   calls which will be automatically triaged by the  AI into different forms of severity or priority   further they'll see that the AI will automatically  extract the location and we'll provide a live   transcript of the call so that they can quickly  see what's going on and even step into the call   once they're available further they have buttons  that allow them to directly with just one click   because the location's already fetched dispatch  police firefighters or paramedics all of this is   done from a human Centric angle and the way how  we achieve this is by taking into account the   callers emotions so for instance one of callers  shows signs of anxiety or fear the system could   work more to calm them down and make them feel  before taking the next safe step this system is   fully designed with ethical safe guides in mind  and part of that was fine-tuning a model on over   500 911 calls so that it could understand the  proper protocols and have a wide or have ACC or   be knowledgeable on a wide variety of possible  scenarios in which a 911 operator could assist   him including fake calls or in instances where  it may not need assistance this is all powered   by our Innovative Tech stack that utilizes a  variety of AI components including the voice   AI the um emotional analysis and of course a  key component of this the fine-tuning itself   our mission is to make requesting Emergency  Services more effective and efficient thank you I'll go first I thought you did a great job  um I thought you presented the problem set   the opportunity and the product very clearly  um and you only had three minutes but you hit   all the relevant points thank you the the one  thing I would encourage you to think about a   little bit is um sort of like the optimization  function for these municipalities right CU if   people Oakland are waiting 6 seconds to  get their 911 call answered like there's   a reason for that I don't know what that is  but somehow these municipalities have decided   that's how they want it to be and so I would  just think about like you know as you bring in   AI to this problem doing the like potentially  difficult AB test of making sure that whatever   it is that these municipalities are actually  optimizing for is actually improved by this   because it seems like a no-brainer when you  first say it but like clearly it's this way   for some reason that is probably nuanced and  and tricky so just something to think about any other feedback or I don't think so well  just just following up on that I think the   key is ease of adoption I mean I think you  you it's easy going to be easy to make a   productivity argument to to the city of  Oakland but then you got to think about   who's actually installing who's who's  paying for this and who's installing it okay think that's good thank you so much thank  you the dispatch all right and next up we have   ASL bridgeify please welcome ASL [Applause]  bridgefy hello my name is ISA and today I'll   be presenting ASL bridgefy the next generation  of sign language Interactive Learning oh oops uh oh this okay sorry so what was the inspiration  behind this well ASL is the third most studied   language after Spanish and English and over a  billion people are projected to have some sort   of hearing hearing loss deficiency which is why  it's even more important to have a seamless way   of for people with hearing loss deficiencies  to communicate with people without them and   vice versa and next there's over a 15,000 per  return return on investment over a 10-year   period demonstrating the value proposition and  existing platforms like duolingo surprisingly do   not take into account ASL learning which is why  it's important to build a plat an interactive   platform where where individuals can retrieve  the accuracy of their signed texts as well as characters okay now our solution includes three  proprietary AI models first we use random Forest   the random Forest algorithm in order to in in  order to map input po pose estimation frames   of frame length of 200 to to the predicted um to  the predicted Al alphabet from A to Z next we also   use an LS lstm model which captures sequential  dependencies to map from to to map from hand hand   pose coordinates to the actual to to the actual  word and then we also have our IND individualized   RG calling in calling in link chain as well as as  well as PDFs that PDFs that are specific to ASL   learning that get that get chunked and transformed  in a vector Dimension space now as you can see   here this is this is a hand handos estimation  um extraction using the media pipe Library so   you can see a b and c and here's our platform  where you can there different modules to learn   alphabets signs as well as sentences and we even  have we even have our real time ASL practice so in   real time to cap to capture the sign that you are  actually the letter that you're actually signing   and give you the accuracy for that so here's an  example of us using the media pipe library to   actually extract all of the hand hand key points  and here are some videos where they're over there   over hundreds of words that you can actually view  to learn to learn each of the hand signing frames   now this is our this is our proprietary RG and  the way in the way we've trained this is we we   we've collected we've collected a variety of  PDFs that that are essentially manuals for   for a ASL learning and potentially in the future  we we would want to incorporate things like you   YouTube transcriptions that can actually be that  that can actually be transformed and and embedded   within our within our Vector Dimension model  now in now in the future this doesn't or ASL   doesn't just hand hand pose estimation doesn't  just have to be H H have to be localized to ASL   can there are plenty of other opportunities for  for human pose estimation including including   Fields like dance martial arts where you can  not only identify certain techniques but you   can also you can also get feedback Generations  from from certain certain input frames and in   the future this could also be integr it into  existing Solutions such as such as Zoom Loom   and FaceTime video so if if there's so given given  a signing of of a certain um sentence transcript   you can you can get in real time the actual  predicted sentences and words okay [Applause] that's nice work for 36 hours I would be um I I I spent some time  creating assistive Technologies for the   Blind and and I would be just just very aware of  the market and how you'll you'll you'll approach   it and and who will be paying I think that will  be a good thing to pay attention to thank you yeah as you as you think about the market you  know I feel like these language learning apps   are tricky to kind of scale to meaningful  businesses you know there was sort of like   Rosetta Stone whatever 20 years ago and then  there's been like duolingo on this most recent   gen but there aren't like that many that get to  meaningful scale so might be worth just thinking   about that market and what are the kind of  success drivers I think even as I mentioned   previously apart from just hande estimation I  think that there's a big big market for body   post estimation I think especially in things like  combat training especially like if you look at the   military even Dance Performing dance performance  companies where they have to train dancers and   they're actually specific techniques in which they  want in which they want ground TR feedbacks for I   think those are also potential markets that  could be ready to penetrate into you chose   more traditional machine learning algorithms  and early neuron Nets like lstm and um [Music]   that may be the right answer that's not obvious  to me but I think you would for today's audience   we need to explain why you're not using more  contemporary gen algorithms yeah so initially   I we were actually thinking about using more en  encoder um encoder based Transformer models but   um we also we ran into some struggles so we just  ended up we ended up settling on the lstm on the   lstms but in the future we would obviously  trans we would obviously adapt more um more   of the state-of-the-art Transformers and even in  the case for um feedback generation for given hand   poses that could be an easy encoder to decoder  um multi like self attention model that you could train okay thank you so much thank you thank you  all right our last contestant for the grand prize   is green wise please welcome green [Applause] wise when I was 14 I stopped  eating all meat I lasted about two months now even though I still eat meat there  are a lot of small changes you can make that   have a huge impact on your carbon footprint for  example by switching from beef to chicken you cut   the carbon footprint of your Meals by six times  what we do is we help you make that switch from   beef to chicken for everything for your shoes  your shirt household supplies food everything   has a carbon cost and a carbon footprint that  we can mitigate so how does a consumer analyze   all their purchases and the carbon footprint of  anything and try to make all these very difficult   decisions and research about how they should  change their actions well this is where green   wise comes in we we seamlessly integrate with  existing customer purchase models to basically   input what the consumer is already doing for  example through receipts or through emails   and we have integrated with Apple pay with  Amazon and with square to automatically get   their purchases into our system from there  we compare we vectorize their purchase and   compare it to our Vector database this database  has all the carbon Footprints of over 10,000   products that we've analyzed and made sure  that these are accurate carbon estimates   additionally by using the vector embedding we  make sure that these similarity scores are very   accurate it's not an llm that can hallucinate  these are real accuracy scores and real carbon   predictions from there it directly can tell them  an alternative product that is very similar but   has a less carbon footprint additionally this  presents a lot of room for scaling when other   businesses want to analyze their carbon footprint  for their products or for events and other bigger venues so from Good Intentions to  reality let's make it happen [Applause] it's a very Innovative rag use case I would have  never thought of that um it's pretty interesting   we not using grat here oh it's not it's uh  similar in that it uses a vector embedding   for um finding similarity but the similarity  is directly the output yes that's right y APIs   is this a subscription product so you would  imagine it being a subscript sorption prodct   we would or we can talk uh probably not um we  ideally we'd integrate with existing businesses   like instacart or Safeway so that they can  show uh our results on how green or how uh   the carbon footprint of certain products is on  their uh app um but it also works for consumers   to use on their own as demonstrated here  uh people wouldn't pay for a subscription though okay I think that's all the comments  thank you so much green wise thank you all right   I would now like to invite our esteemed  judges to convene in this secret room   where judges make their decisions and we are going  to have the special prizes so as I mentioned a   bunch of sponsors came to make this all happen  we're an educational program and it is entirely   the support of these sponsors and they're not just  providing support they got cool prizes so let's   bring them on in just a minute you're going  to hear from each one these are the sponsors   for today and I also want to thank our community  sponsors these are startups very cool startups who   hung out and helped our young hackers with their  needs and their Cool Tools all right so our very   first special prize is going to be announced by  a very special campus partner I'd like to welcome   the academic Innovation Catalyst I'd like to  welcome out here Matt cini and Paul work to tell   you about AIC one of our newest campus Partners  doing very cool stuff please give them a welcome   thank you thank you so much Caroline it's just a  thrill to be here so my partner and I Paul and I   created academic Innovation Catalyst or AIC to  release more Innovation from academic research   for positive social impact and we're focused  initially on the areas of climate Tech and AI   for good which is why we're here today how do  we do this well we make proof of concept grants   so no strings attached non-dilutive grants to  academics with with an idea then we help them   take that idea carry it through commercialization  to scale and sustainability and so that's what we   do we're thrilled to be here today and we'll be  making two Awards to the most compelling business   plans or Innovations involving the use of AI to  make the world a better place and we couldn't   be more excited to announce them in 5 seconds  here I'll just say that we met with many amazing   teams it's been an extraordinary weekend thank you  so much for including us we had to narrow it down   to two it was tough but I think you'll see that  they're well-deserving so with that let me hand   it to my partner Paul work to announce the winners  of the AIC AI for good Awards in 19 or I'm sorry 2024 this is what happens when you get old  people up here on stage so anyway uh we are   really thrilled to be here as Matt said  and we're especially thrilled with the   fact that so many of you are putting your  talents to work for such great causes and   for the betterment of of humanity and AI has  so much potential in so many Realms but among   the most important is to make the world a  better place and to make a social impact   and so with that we're thrilled to announce  the first two two winners dispatch Ai and ASL rifi so the these are you know tremendous  companies we we were again the the competition   was so strong um may I ask actually uh both  sets of of winners to stand in in the audience   here and thank you again so much for the for  the terrific work um I think as as you heard   ASL bridgefy is doing for uh sign language what  dualingo has done for learning other languages   and it is so important it's incredible and and  shocking that it's an underserved and currently   not served market and their Technologies are  going to change that and dispatch AI uh what   what can you say I mean it's such an important  issue to be able to get emergency response to   be able to get a response when you need it and of  course the reality is when we have unfortunately   too many mass catastrophes the time when you  need the response most rapid is the time when   you're often most short staffed and so dispatch  AI using artificial intelligence and a variety   of Technologies to speed that process up and  to help both the dispatchers and the people   that the dispatchers are helping and so can I ask  the dispatch AI team to stand up as well and be recognized it's a great job congratulations  to all of you and and to everyone who is here   today thank you so much thank you madam Paul thank  you academic Innovation Catalyst our next special   prize is going to be introduced by our very own  general manager at skyc syil Chen give her a welcome hello everyone hope everyone has had a  great weekend at Sky Deck we have we about a year   and a half ago we launched the skyc climate Tech  track uh in part thanks to a nice Grant from the   University with $150,000 to build out the track  and right away we started putting that to work   we grew our advisor network from you know maybe  like five advisers in climate Tech to now over   30 advisers that are in the climate Tech space  and beyond that prior to the grant we had maybe   three to five startups um every batch that were  in climate Tech and now we average 15 startups   per batch in the climate Tech space and we really  hope to see that grow um so I'm very pleased to   announce that the winner of the ske climate  Tech track is Team Green wise I think they're   still in the green room cuz they just pitched  they were the last ones to go on stage um but   they really kind of represent the type of start  the you know team members that we like to see at   early stage startups it's three team members  that are best friends from middle school um   oh they're all here on stage come on out I wasn't  expecting that but um Anthony Ben and Ethan three   best friends from middle school representing  UC Davis UC Santa Cruz and UC Barbara oh UC   Santa Barbara um and they've built a platform  for carbon footprint track footprint tracking   uh with actionable recommended recommendations for  vendors um so that people and companies can reduce   their overall carbon footprint so please help me  and congratulating this team winners of $25,000 all right thank you simol thank you green uh  clicker clicker thank you all right next up   specializes from Intel Intel come on out intel was  here their table was swamped I'd like to introduce   Gabriel Amaranto hi everyone um thank you all  so much and thank you to the organizers for   having us we've had such a great weekend um and  we your projects are so amazing so thank you to   everyone who joined our track as you can see the  winners behind me congrats to all the winners uh   we have our raffle winner Isa RS third place is  aell second place is batteries by llm and first   place is dispatch AI so let's give them a hand  of a r Round of Applause um yes great job amazing   projects if you won please meet us outside I want  to hand you your prizes we have them with us so   please meet us outside so we can take pictures and  give you your prizes thank you thank you Intel all   right next up AWS come on stage AWS we've got  Rohit Teri Kevin Lou and Brandon Middleton and   that's what they're going to tell you go ahead  Rahe howdy there can you all hear me yes awesome   well hey uh thank you so much Sky Deck team  for having us and Cal hacks this has been an   extremely impressively run operation and we're  really excited to be partners and sponsors of   this hackathon uh today we have three different  prizes actually let me introduce myself first um   we have Brandon Kevin and Rohit we are members of  the generative AI organization at AWS uh we work   with top generative AI startups like anthropics  stability uh perplexity and others uh in the   development of their large language models as well  as our overall kind of inorganic and inorganic and   organic growth strategy including Investments  as well um today we have three different prizes   uh we have four of the teams that we have uh  chosen to uh give the prizes out to our first   place prize is for $10,000 uh in AWS credits and  uh we have two other prizes one for climate Tech   and then one for responsible AI which are 5,000  I did want to say that uh we talked to so many   impressive um startups and uh founding teams  today and and hackathon teams today um I wish   we could give prizes to all of them we did want  to recommend that uh those who we spoke with   and I think we have these conversations with  you already to go ahead and apply for the AWS   generative AI accelerator uh as well as our AWS  activate program to receive additional credits   um I'll go first I'm going to be announcing  the climate Tech uh we're going to give the   prize out to uh disaster Aid is disaster Aid  in the in the room today yes good good job guys   Kevin and then uh for responsible AI we have a  two-way tie so we're splitting that prize into   2.5k for each team in credits and that's GPT  ethics and DP ancestry they're in the in the hall all right and I'll round us out um our  grand prize kind of the most impressive thing   that we heard and saw uh today is going to  go to safeguard so Safeguard team if you're   in the building uh stand up real fast let's  give you a round of applause I don't see them   but God bless you and keep doing what you're  doing thanks so much guys thank you thank   you Intel our next amazing partner  reach Capital please come out Tony Juan oh out of order let me see if I can find you  there you are all right okay oh Mike awesome thank   you so much um thank you to C hack thank you  to Sky Deck it is such a delight and pleasure   to be with all of you today and thank you to  everyone for being here from across the country   from across the world my name is Tony and I'm from  reach capital and let's just cut to the chase cuz   there's no drama here we want to congratulate  Frodo for winning our Ai and education prize   Frodo Aman Kush and a team if you are here please  stand up please stand up all right you are right   up front you are in the right place thank you so  much you've uh you've won the one ring as they   say or at least or $1,000 cash prize so please uh  let's meet up afterwards um reach Capital we're in   early early stage at Tech VC firm investing  in attech we invest in education across K12   higher ed and Workforce in tools that expand  access to education and economic ability and   many of the companies at a portfolio were founded  by students themselves because you know like what   better place to find great ideas and great talent  than to go to places like this where students are   living that experience so if you're building  Adventure NCH please reach out thank you so   much thank you Tony next up we have u.com we  have Rex come on out Rex and tell us about the prize Applause please for our thank you so much  yeah so we wanted to announce um I'm Rex we're   from u.com this is Oliver um as as you know  u.com brings transparency and um web access   to our llms to make sure that they're as accurate  as possible so we wanted to give an award for the   best use of us u.com apis um to transpar ify so  congratulations if you guys want to stand up if   you're here there you guys are yeah thank you  so much transpar IFI did an incredible job by   um they were live streaming videos factchecking  them as they went using sources from the web um   and u.com search apis it was really incredible and  Powerful um and Oliver will talk about our custom   assistant yeah so for our best custom assistant  we'd like to give that to events. uh with Oliver   and deves so Oliver and deves can you please  stand up if you're in the room congrats over there yeah so we were particularly impressed  by the what they've built uh essentially they   handled booking searching and uh even talking  with customer agents on the phone and they used   u.com uh in a way to actually find these  events so we were incredibly impressed by   them and can't wait to see what they do in  the future yeah uh come find us after and we   will give you your um Awards thank you Hume  or you thank you all right I think we're   going back to Hume now with Zach great  house welcome hum nice round of applause please hi so first just a huge thanks to  Sky Deck and C Haack for organizing this   event and inviting us back and to all  the staff for running such a memorable   event um so I'm going to be announcing  our three categories for our track we   have our um best AI our best empathic AI  for social good best empathic AI for just   most Innovative in pathic Ai and then just  the best overall as you can see the team's   here we've chosen scam scanner can scam scanner  you're here can you stand up all right big cent Applause for most Innovative we have Bloom buddy  where's Bloom buddy he stand up yeah okay great   job you guys talking to plans uh and then best  use of empathic AI overall we chose lock in it's   a personalized uh learning assistant uh are  you in the room where are you yeah there we   are okay congratulations you guys uh come  meet us after outside uh we'd love to chat   take pictures and uh thank you so much thank  you to all the participants uh yeah maybe see   you next year so take care all right thanks  hum all right and our last special prize is   there they are please welcome Jose Menendez hey everyone uh very nice to be here  for those of you who haven't heard about Groq.   groq.com experience the fastest inference  on Earth period all all I have to say about   Rock right now but our special rockar award  today goes to scam scanner where are you guys so these guys have a product that I want my  mom to use today right monitor your call for   potential scams who doesn't want that for  your mom your uncles and in the whole thing   um now they get 1,000 credits on Groq Cloud  which is many many millions of tokens um   there's two special mentions I have to read so  I don't screw up three brown one blue where are   you guys another awesome solution these guys are  generating on the Fly incredible videos for math   uh something that I would use right now as well  and trans verify are you guys around here you   you've been mentioned as well trans verif five  very cool who doesn't want to hear a podcast   with instant fact checking right am I right um  now my special surprise for the day I want to   make a very special mention of Nathan Bog are you  around Nathan all right Nathan didn't use groq so   I'm going to give a special technical Excellence  award to Nathan for for a model he trained and   ran on his CPU uh for doing very interesting D  operations Corrections on the Fly for frontend   uh not only that Nathan is invited officially to  present his work in the groq uh HQ uh as soon as   he can uh that's it guys I'm very impressed  with all the work we saw thank you very much   congratulations thank you for rock all right our  esteemed judges are back with their determination   please come back judges come back so we can so we  can all enjoy the grand prize are you guys ready   do you have a guess is there a voting do we have  voting tally taking bets everybody I want you to   guess the top your top two choices for grand  prize and and then I'm going to ask who got   it right okay so as our wonderful judges take  their seat all right we we got some shout outs   going here any other shout outs okay all right  this audience is into it so as a reminder the   grand prize is a $25,000 investment from the  Berkeley SkyDeck fund also a golden ticket to   our pad-13 program at SkyDeck and a special  prize we are happy to announce that open AI is   providing $2,500 in credits for this winner so  I think we're ready for the drum roll take your   guesses only the judges know I don't know we're  all about to find out it's Dispatch AI dispatch   where are you come on come on up there stairs  right there come on come to the front stage there   you go thank you judges I want to invite while  we invite dispatch up I want to thank all of   you for coming I want to invite Spike spike yes  from Dispatch yes okay oh here's the team there   we go Dispatch AI Grand Prize winners well done  well done I'd like to invite the SkyDeck staff to   come out and the Berkeley hackathon staff to come  out come on out they've been working all weekend I   think some of them did not sleep at all please  give everyone who joined to make this a huge   round of applause thank you everybody thanks for  joining us we will see you next year [Applause]