How GitHub operationalizes AI for teamwide collaboration and productivity
[Music] please welcome Kyle [Music] dagel Hello Universe I'm Kyle dagel and it's so good to be here with all of you I'm a developer myself a longtime hubber and I'm now proud to be the Chief Operating Officer of GitHub thank you this is my ninth GitHub universe and I can say that I've seen a few things I've seen us go from releasing poll requests to announcing the first version of GitHub actions and even the Arctic code Vault where we stored the world's code for a thousand years and now we're in a moment where suddenly The Impossible seems possible today I want to start by sharing a story about just that a few months ago Thomas and I were discussing how awesome it would be if we could get attendees at Universe programmable badges we did a trial run last year for a small number of folks but we wanted to go bigger the problem was there's no off-the-shelf solution to do what we wanted to do at the Quality that we wanted to do it so like some of the best problems we found a super smart hubber to help us make this dream of reality we went to the incomparable Martin Woodward and team he took love of 3D printing raspberry pies I think the man owns more raspberry pies than anyone I know and he set out to create a way to get over 1,000 of these badges in the hands of our all access passholders how did he do it well it wasn't his full-time project he's running all of devell for GitHub and helping to prepare for this event too but not only did he build a custom platform to enable this work he made a custom flashing software so during registration the team could pull a badge from a bag flash your information onto it and hand it to you he dreamt it built it and he showed it to me last week and I asked him I said holy crap Martin this is amazing how did you do it and without a whiff of irony he says honestly co-pilot I could have never gotten this done without co-pilot in partnership with Martin's creativity attention to detail and immense Talent co-pilot helped fill in the gaps teaching him the parts about the uh the different embedded technologies that he didn't know and now we all get to use and hack on these awesome badges with AI he went from a crazy idea to something achievable in just two months and we're talking about physical Hardware which we all know is even harder to work with Martin and team set out to make the impossible possible and he had the help of an AI pair programmer along the way let's take a second and give Martin and team a huge round of Applause see progress is doing small things in shipping them not perfect things but ideas projects and features were bold enough to cast out into the world so think back to 2005 when git was created we were all already using subversion or maybe CVS or something even older but then we started using git more and more and more and more open-source projects started following linux's lead in adopting git and now git is used by 93% of developers globally and GitHub is used by more than 90% of the Fortune 100 in this era of GitHub we've refounded ourselves on co-pilot in the same way we were founded on git all those years ago now while a lot has changed over the years we've kept this Spirit of shipping to learn with everything we do it's a core value here at GitHub we build build something we get it in the hands of developers we see how they use it adapt change and sometimes we just start over but it's how co-pilot could even come into existence in the first place and with this approach we found ourselves at the Forefront of what's destined to be the greatest chapter yet the age of AI the first GPT model was released in 2018 two years ago in 2021 GitHub started sharing our work on a new tool called co-pilot and last year we shared co-pilot with the world and now developers have already adopted AI at scale in a survey we ran earlier this year 92% of developers said they're already using AI in their dat to-day work and that number's almost certainly grown since then what a miraculous adoption curve over the course of only 5 years developers grappled with Transformer models brought their impact to the mainstream via llms found new ways to embed it across developer workflows and harnessed it to help you focus on what you love most which is code with co-pilot you're writing code over 55% faster no more looking up functions and methods no you get to focus on the essence of being a developer you get to spend more time solving the hard creative problems not the ones that you've already encountered and perhaps most importantly you're spending time in the flow co-pilot can sometimes feel like it's just completing the thoughts in your head getting down on paper an idea you've only just had or other times co-pilot chat can act more like that expert non-judgmental friendly developer that you can ask any time of the day or more frequently at night this notion of being in the flow isn't an arbitrary concept it's what enables you to do your best work and focus in a world that demands so much of you every single day because we know developers aren't monolithic yes you write code but you also manage projects work with other developers schedule meetings fix broken builds give advice and deployment strategies while coding is one of the most important things you do let's face it it's not all you do case in point we asked 500 developers what they spend the most time doing on any given day anyone here think it's 100% coding of course not you're writing tests you're waiting on code reviews you're chatting with colleagues about potential Solutions and only about 32% of your time is actually spent writing code and so while co-pilot is helping you work faster helping you write code faster than ever before and keeping you in the flow what about all those other things you spend your time on are you in a flow State there too are you happy about asking say about an expense policy so you can know if the Blue Bottle Coffee you're getting while visiting universe is covered does it spark joy to open up three tools to monitor your deployment after you ship something thing as the COO of GitHub it's my responsibility to enable hubbers to do their best work so when I took this job I set out to ask how can we take what we've learned from using co-pilot internally to both help developers at GitHub but also to help every hubber no matter what their role is what I'm saying is it's not enough to operationalize AI in just one area of the business to fully harness the potential of AI we must pull it across all all of our existing workflows not just one because if we want to continue scaling the developer experience we have to solve for coding and collaboration and all the other things developers spend their time doing developers are multifaceted so our approach to AI adoption must also reflect this so we should start by listening to them as early adopters of AI developers are bellwethers for Innovation while much of the world grappled with this technology in use cases you're already adopting them at scale and providing the world with a blueprint of what's possible we've watched more generative AI projects emerge on our platform over the last several years between 2010 and 2018 we jump from little to no AI projects to just under 10,000 we saw an increase of about 8,000 projects as we approached 20120 and then in 2022 there was a steady incline and then boom look at that that's a hockey stick if I've ever seen one in less than a year we more than tripled the number of gen projects on GitHub even just halfway through the P this past year we saw over twice the amount of geni projects in 2023 compared to all of 2022 and this is a global phenomenon developers all over the world are experimenting with AI in big numbers building new applications with Foundation models and working with one another to build what's next but it's not just about the number of projects on GitHub it's the types of projects that we're seeing emerge AI projects have progressed from more specialist oriented work in research to more mainstream adoption with developers increasingly using pre-trained Foundation models and apis to build gen powered applications now when you look at these graphs and these projects it can be easy to dismiss them and the speed at which they've been adopted as height but when you dig in deeper you ask questions it's easy to see that it's not just exciting these projects are having an enormous impact on the day-to-day lives of developers just this past year alone we saw developers add hundreds of AI powerered actions to the GitHub Marketplace that do everything from code review to Security checks that shows how developers are already organically infusing AI across more of their workflows it's remarkable but it also gets to a central question question that we've had over the last year how are developers solving real problems with AI and we've learned a lot in the process developers in Enterprise settings work with an average of 21 other Engineers on their projects and they want collaboration to be a top metric in performance reviews over code quality experience developers know that power the the power that collaboration can bring to their projects and their team and what I found most interesting is that developers believe AI will actually help solve the collaboration problem more than four out of five developers expect AI coding tools will make their team more collaborative just think about that for a second more than 80% of developers think AI will improve developer collaboration when we fielded This research I expected the opposite maybe AI means more people can code on their own with co-pilot and do more of their work independently crank out those big Ideas put at the pull request on a pile for review walk away but the opposite is proven to be true half of the developers we surveyed see AI removing the mundane tasks right away and leaving them able to focus on the higher order problems what we're seeing with this first wave of AI tooling is that AI is quickly removing the tedious and repetitive work and that makes sense the first work to go is the work of making Library upgrades switching to the new aray sytax and Ruby and all those things that you love to hate when you're pulling your next GitHub issue to work on but our research shows that the next step the next wave in AI is becoming a second brain we're not talking the jump to the Industrial Revolution but a jump to the information age in a course of just a few years not several decades and finally where developers go everyone follows what you adopt the world adopts so it's absolutely no surprise that 70% of developers say AI coding tools are offering them major benefits at work like improving code quality faster outputs fewer production level incidents and they're simply enjoying their work again in a way they haven't in a while if implementing AI means increased productivity collaboration and Innovation on engineering teams when they're writing code 32% of their time there is an enormous incentive for organizations to bring AI to the other 2third of the Vel opers day and across their other teams too think about it if customer teams find what they need faster they can shorten response times if it teams can automate and triage frequently asked questions they can focus on solving the bigger issues at hand if you can get the answers you need you can get back to what you're uniquely capable of if we follow in the footsteps of developers this incredible Flow State they achieved can be a reality for all team teams at GitHub we always think developer first that means thinking about what tools and processes developers thrive in it means we use GitHub for so many of our own internal processes want to have a document reviewed by legal open an issue for that have an idea for some custom swag at Universe open an issue for that want to update our latest octe report you can open a PLL request for that in a lot of ways everyone at GitHub and and increasingly more and more people in the world are developers of some kind ship to learn have a growth mindset and take steps to work better together because our community of a 100 million developers is powered by a little over 3,000 insanely talented hubbers to sustain this incredible momentum and continue building what's next for developers we must fully integrate the very technology that we've refounded ourselves on we want to make make hub's lives easier in the same way we make all of your lives easier and of course that starts with AI so I'm excited to show you the various ways we're using AI across the GitHub business to not just get more done but help all teams here at GitHub do their best work I'll show you a few things that we've been working on share some of the lessons we've learned and while giving you a sneak peak at what's to come just like our community spans the world so do our employees our geod distributed Team Works async to bring the products and tools and features that you rely on to life so let's say you're a hubber based here in San Francisco managing projects for the GitHub CLI team you're working to meet a critical ship date and you need to get the band together to see it through the only caveat is one person's based in New York and the other is in the UK and two more in France in Spain working remotely takes investment and so we've looked at tools like rewatch to help take our meetings and videos to the next level thanks to their in the- flow AI powered functionality like most remote companies we're always navigating time zones but with tools like rewatch we can take a more thoughtful approach to collaboration with the click of a button we can record our meetings in teams or zoom and share them with everyone the videos are uploaded automatically to rewatch and that's when the magic happens I travel a fair bit for work so sometimes I can miss internal meetings but now I can go to rewatch find the key points via their summarization of the meeting see key moments in the video and watch those in a high information density way and in all of this this is my favorite part it can actually identify action items coming out of the meeting the reason I find this so interesting isn't because transcribing videos is particularly novel but because we upload all of our videos into rewatch and we get this increased time to answer by using rewatch I can upload my staff meetings our internal company get togethers and you don't need to rewatch the entire meeting to understand what happened and when I have employees in Australia and San Francisco and Berlin and Tokyo there's no perfect time for a meeting so this helps in what feels like a small way in a very big way but with a new bit of functionality from the rewatch team we can go even a step further so see these action items why not link them to a GitHub issue you can choose an existing issue or in our case create a new issue Define the title set the body use the existing information and what repository you'd like and finally who you'd like to assign it to it'll create a new issue to help you coordinate your work with all within your normal GitHub process and flow then when you close out the GitHub issue when it's complete the state will even sync itself to rewatch now you can have all of your action items in GitHub in any anyone who comes after can see the progress you've made right next to the video using AI like this automatically an in existing behavior is a great example of starting small when it comes to AI don't overthink it start small with existing workflows measure the success and then scale up to reap even greater rewards we know that meeting notes are helpful but it's just one of the many things that hovers do on any given day like you they're making and receiving requests for it and we've found some ways to operationalize around that too using a bot to send reminders is already an effective method in the open source community and it's something that we've done for years at GitHub with something we call hubot which you can see here for a long time we've known that people respond better to feedback from robots particularly when they're telling people that they've done something wrong or badly we've given feedback about code reviews brok broken builds and required fixes from hubot for years and now we're working to bring it into our it workflows at GitHub every hub's interaction with our it team actually begins with a chat and as you can imagine there are a ton instead of bringing in a help system as the way hubbers had to interact with it we brought it to where hubbers were already asking their questions and for us that's in slack and it's a lot of questions and help requests over 5,000 in inquiries every quarter to be exact that's an average of over 625 inquiries per it team member per quarter as coo I get to lead our insanely talented it team and if I'm honest this is not where I want to them to spend all their time we need them focused on our strategic efforts bringing Technologies to Hubers to make it easier and faster to get their work done to let them collaborate with each other and of course have the tools that they need to do their best best work not responding to the same how do I reset my two-factor authentication question again and again and so of course the team saw the power of co-pilot that it was providing to our developers and thought we need this for our company processes and workflows and after partnering with a company called move works we now have a friendly neighborhood bot that can keep all Hubers in the flow and we call them octopot powered by AI octobot now handles over 30% of these inquiries ensuring that hubbers found the right answer and in some workflows octobot can even take actions on behalf of Hubers across our systems I'll show you that one in a second with octobot we're able to give our it support team 25 hours back every week that comes out to each person getting over three hours back to do more meaningful work like building the rewatch integration I just talked about it customer satisfaction score now stands at 98% that's up 12% from where we began that's not just the number it's a reflection of the exceptional support our team is providing not only do we have octobot solving these requests but satisfaction is up along the way let me show you what this looks like in action we'll start a conversation with octobot and slack let's say this time I'm a new hubber it's my first week and I need help with a few things like how I should go about setting up my laptop all the details I need are right here I can click the link that applies to me then I'm connected to our internet for instructions now we found that linking to existing documentation instead of generating new language from llms as a slightly more immediately actionable behavior for Hubers and they can get the rate they can rate the response and give valuable feedback at the end now that that's done I also need to set up one password let's ask octobot to help me with with the setup instructions it even understands if I mess up with a tyo halfway I see my prerequisites here and now I can click out to further instructions no need to Ping anyone else and I can get the answers I need fast and get it working on my computer easily now let's say I've been here for like a month or two and now I'm working with a customer who also uses slack it's easier to work with them in the same space but I need to get them added and now I have what I need to add the vendor for quicker collaboration but we can also build Out full workflows starting with proactive reminders from octobot so I'm going to show you an early example every four years hubbers get a new laptop refresh it's something most people look forward to but it's easy to forget and can be a manual process in many companies done over email and including multiple people at GitHub it all starts with a simple slack message telling me I'm eligible for a refresh this is our version of starting small we alert the hubber about their eligibility and in the future we can walk walk them through the workflow in choices of laptops all within slack using moveworks I didn't have to wait on someone to reach out or hope that they'd remember we reached out to them kicked off a workflow and I'm back in the flow and on to my next project and we're not stopping there we're bringing octobot trained on our internal processes and handbook data to help answer questions about expense policies travel policies and this year we even trained up to help answer hub's questions about GitHub Universe this type of seamless interface where AI is already embedded into your workflow is what keeps not just your developers but all the teams across GitHub focused they don't need to Shell out to another Search tool they just use the same behavior they're used to asking someone for help in slack everyone the hubber and the team usually answering the questions get to stay in the flow and do their best work focused on more strategic work focused on staying in the flow and focused on on building and supporting the platforms products and features you love my last demo here is a little bit more personal just a few weeks ago I got a ping on a code base I developed here at GitHub over 10 years ago and I assure you I am no longer the right guy for the job it was related to web Hooks and our apis and it's something I'm still very passionate about but I've lost touch of bit I simply don't know what the fix needs to be and on top of that I'm jumping from meeting to meeting answering messages from Hubers this code needs to get fixed but there's no time for me to dive in when I talk to developers and customers we often talk about how powerful co-pilot is for junior developers and our research has shown that to be true but co-pilot is also incredibly powerful for senior developers particularly when they're dropped into a codebase they're not as familiar with or in this case the code base is 10 years old so in this example I'm using a repository that I built over 10 years ago called hook hook is a small application that receives web hooks to make it easier to test something like github's web hook system I honestly can't even remember what the core parts of the application are since it's so old so I'm going to start by asking co-pilot chat to explain the key parts of the project to me okay so it's using mapper the hook delivery model and Erb great next I know the main issue is that the underlying data store is no longer supported where I'm hosting this so I need to move it this is still using mongod DB and I need to move it to my SQL it's been so long since I've worked on mapper so I'm going to ask GitHub co-pilot chat and see what it says it can give me a step-by-step set of instructions about the changes I would need to make to go from mapper to my SQL 2 okay finally further down in this uh demo it looks like there's a method that's a little bit suspicious clearly not my best work so I think that there's a method that does this in rails or Ruby I can never remember so let me ask co-pilot chat in line about it so I'll select the code I'll trigger the inline editor and I'll ask if this is the best way to do the code okay i' always forget if the pluralized method is standard Library Ruby or rails particularly if Ruby has gotten this in the standard library in the last few years so let me follow up and ask about that no it is not okay so now I know all I have to do is import is it action support action view uh and I can go on to cleaning up this method now this is a simple example but you can can imagine this workflow with dropping into a scholar project that you're moving to Java or a Cobalt project that you're working to modernize or like I did here just an older project in the language of your choice with the more modern equivalent I spent barely a few minutes paging in the context of the repository and asking questions about it before I ever started coding but I want to show you one more example of a developer who can show you how co-pilot and co-pilot chat are working for her dayto day and her role at infosis I had this existing code and I just had to implement a feature so I was still processing it in my head I was like okay this is the logic I want to do uh these are the lines of code and in the meanwhile I thought let me just you know give co-pilot chart a try I gave a prompt and the code I didn't even know I need it yet was already there on my screen I was like wow when anyone imagine developers or a coders system it's always multiple windows open so you can see a shift now where all of this is happening in one window so all I have to do is I need to implement a piece of code I ask copilot chat so this is what I want to do it gives me a response and then I am like okay I don't know what this function is doing I need some explanation for it again cellet chat I'm working with something new so that I haven't really gone into it again copilot chat all of my smaller tasks I can do all of it in one screen which not only saves time which makes it more efficient so that way I am more happy Satisfied by the end of the day the code I [Music] produced what co-pilot got so right was keeping developers in the flow and introducing as few new behaviors as possible sure it'll help help generate code for me but I can ask a question like I would any developer but it's at my speed and I don't need to take the time of my team until we get to a real novel idea this tool wasn't created with just one type of developer in mind whether you've been coding for N9 months or nine years co-pilot can help us do so much more I'd like to take a second to give a huge thank you to our it teams and operation teams whose work I got to show off today they're all watching remotely keeping everything running for us while we're at Universe let's give him a big round of applause too thank you everybody AI can help us all do more achieve more and ultimately at GitHub that's our goal we want to help the world accelerate human progress through developer collaboration and that starts by enabling developers to do more satisfying work more creative work work that makes them happy and with co-pilot and some of the other tools I showed you today we're well on our way at GitHub for hubbers and I believe you can have it too by following a few simple principles first start small but start now with so many opportunities to adopt AI tools right now in every company Under the Sun suddenly being an AI company it could feel like the best path is just to sit back and let it all play out or go into the fancy boardroom you have at work and come up with the council to plan for AI tools adoption across the Enterprise for the next 5 years this technology is evolving so fast you need to start start small in measure and adopt quickly the technologies that work for your team I don't want to downplay that particularly for some of our largest customers there are compliance and Regulatory obligations to meet but I've seen some of the most complex Industries adopt tools like co-pilot and quickly so I know it's possible a recent study from IDC said for every dollar of investment into AI there is a return of $3.5 on that investment that's a hell of a return the only people not experiencing the benefits are the people still planning to take advantage of it second no one wants to use an AI tool that requires you to Shell out to a new page to a new app or the only that only works by clicking an AI button the best AI tools are in flow where they're helping people while they're already doing what they normally do like code in their IDE or asking for help within slack this is what fueled GitHub co-pilot's fast adoption it simply helped you while you were already doing something that you did every day code then developers are very used to asking a fellow Dev what's the method name to call that API so chat also made total sense and as you heard this morning as we continue to bring AI through the entire GitHub platform we're focused most on having AI help you do the work you're already doing finally AI isn't just an amazing productivity tool I think it's also our best shot at true learning on the job and it's helping both Junior developers and the most senior developers learn new skills and get comfortable with unfamiliar codebases quickly find AI tools that have compounding impact by building shared understanding and personal development while you're using them every big open source project or company has code that very few people know the ins and outs of with tools like co-pilot chat and rewatch summarization another Dev can go in ask questions about the codebase what life was like back in Ruby 26 and even find a video of me speaking about it in a team project all impossibly faster than it would be if they had to learn all of that on their own so start now keep it in the flow and focus on the compounding impact of learning this momentum and productivity and honestly Joy will compound over and over until the world is forced to pay attention with each small step we'll build a future where innovation has no bounds together we'll make the world take notice just as we did with Git just as we're doing now with AI for everyone and will accelerate Humanity's progress together thank you so [Applause] [Music] much
2023-11-14 08:41
