GitHub Copilot: the AI pair programmer for today and tomorrow

[Music] please welcome Ryan J Salva oh [Music] yeah good morning developers data scientists creators beautiful nerds everywhere uh y'all I have I have been talking so much about co-pilot these last couple days my voice is a little bit horse but we're just going to get through this together all right yeah I think we can do this my name is Ryan J Salva I am VP of product at GitHub where I am responsible for GitHub co-pilot and it's hard to believe I know Thomas was talking about this a little bit earlier that we are nearly on the precipice of three years working with GitHub co-pilot we are quite literally the granddaddy the OG the original large language model product out there I mean is before chat gbt he held even before stable diffusion was out there we were out there helping developers build and create and do wonderful things with code one of the things that I'm really proud about as we've kind of like gone on this journey together is that we have been research L right we've LED with research about code quality about developer happiness and productivity heck we've even LED with research about economic impact but one of the things that that we've always really heard from a lot of teams is like they want to know not just how it works inside of a controlled environment where we're just testing for a single variable they want to know what it's like in the real world where things are messy where things are dirty where there's reorgs and there's hairy refactoring projects and things like that and so we actually embarked on a little journey few months ago to work with some external companies and do a high Quant study with hundreds and and hundreds of developers to see what it's actually like what impact these large language models and co-pilot have on engineering teams and so for that I want to invite on stage one of our research Partners Dan shocky from MC centry to tell a little bit about some of that research Dan come on out buddy yeah let's hear it cool welcome Dan welcome welcome thanks for having me yeah dude absolutely so all right I'm not sure that everyone here necessarily knows Accenture so can you just like maybe start by telling a little bit about what Accenture does with your business yeah sure at Accenture we help our customers build their digital core um transform their operations and accelerate their growth and we do that by bringing together teams from you know across a wide range of capabilities many locations around the world bring those together to help them serve our clients and as you can imagine that can be a pretty complex environment um where we have you know many developers um with a bunch of different you know diverse backgrounds coming together working in many different Technologies at the same time uh to bring the best solution to our client I love it I love it I kind of like I kind of think about this is the shock and a slide just a bunch of numbers with like the sheer massive size of it right so Accenture from an engineering size like how many Engineers do you have worldwide roughly so roughly about 125,000 software Engineers working for us right now holy moly dude is like massive I think that might even be bigger than Microsoft all said and done that's big that's big okay so why were you interested in kind of like pursuing this research with us around get a CO pilot so you know as a company that invests very heavily in both our people and the technologies that they use to do their their job um we're we're very interested super excited about the promises of AI assisted software delivery and really the changes that that's bringing to the way that we all work day to right so we had a lot of interest just in general in GitHub co-pilot as the the leading you know technology and that set of change um and really what we wanted to do was become what we consider a customer zero right so we like to First enable our own employees our our own Workforce with the tool and then we experiment with it at scale right let them use it then we learn the best way to do it the best you know how best is applied in our industry and then we can take those learnings back to our clients um that that we go and work with every day and of course we're also bringing that feedback back to GitHub to make sure that we're helping to drive the The Innovation with GitHub co-pilot I I like to think that we're getting better because of your feedback so thank you thank you I will actually say here that that is a um maybe a commonality that I've seen among some of the first adopters that a lot of the the organizations that adopted co-pilot here in these first what eight months that it's been available they've been some of the largest engineering teams on the planet with some of honestly some of the customers that would have been the most cautious or slow moving in the past have really jumped on AI now my hypothesis here is that because the productivity gains there actually math out to be a huge amount of well additional value delivery additional cost savings and the like that you get and so is that kind of the same thing that you see going on extension yeah of course that's the question on every body's mind right how much is this really going to save us what's going um how how does it impact our delivery um and you know we started pretty simply we just enabled a couple hundred of our internal developers and you know turned them loose with co-pilot let them use it as part of their daily work um and to be honest our developers loved it they came back to us we had a developer that called us back in about 24 hours after we turned on co-pilot for him and he told us that you know Not only was he using the that is part of his day-to-day job but it was saving him about 20% of his time just taking care of all of his unit test cases for him I I was actually I we were talking about this backstage earlier and so Dan was telling me like the developer had like a whole bunch of tests that were going to take like days to complete and he completed them that afternoon right so exactly took some time off got a coffee you know maybe browse the web that kind of thing and you know we had a few developers that had that brought us stories like that and so we knew that this was going to be a big change to the way we work and really impact the way we deliver software in general all right well let's look at the numbers here and actually see what results we saw out of this right okay so I mean first of all let me give you a little background um you know we enabled about 450 of our internal developers with GitHub co-pilot uh we didn't change we didn't give them any specific tasks to work on we basically just said continue doing your normal work um you know your day-to-day job as a developer and then we uh we observed that set of developers and their experiences over the next six months uh at the end of the six months we compared those 450 developers using co-pilot to a similarly sized group of their peers that weren't using co-pilot right control treatment group control and treatment group uh so you know we tried to be scientific but you know like you said there's a lot of factors at play here um but what we found out that overall uh the rate at which our developers were accepting suggestions from co-pilot was about 35% by the way that's actually that's in keeping with what we see with the population at large that's that's it seems pretty normal um of those that code that was accepted by our developers they retained a little more than 88% of the code that that co-pilot suggested after they did their review okay all right so that's pretty good so high retention rate so co-pilot was giving suggestions that survived edits essentially I mean our developers like the code that they were getting right um we also saw a pretty pretty dramatic increase in the number of builds that our developers were attempting um and overall not only were we seeing more builds coming from our developers but about 45% increase in the success rate of those builds so we were also seeing you know some increase signs of quality of that code as our developers were using the tool more so increase throughput increase quality now that's around builds I know you did you monitored we together monitored PRS as well right right so yeah we also Al saw an increase in the in the throughput uh poll requests that our developers completed over that 6 months so when they were using co-pilot they completed about 15% more poll requests That's crazy dude that is absolutely crazy amazing so more throughput for the team higher quality at large let's talk about what I you know genuinely care about the most though I I'm all for productivity developer happiness right how develop how do developers respond to it right so you know we of course had the same question we gave a survey to all the developers that participated in the study with us and 100% of those that responded to the survey told us that they found co-pilot useful damn all right yeah we didn't get a single negative comment nobody told us it was distracting they all thought it was useful that's almost unbelievable I'm going to be honest it is kind of unbelievable but get this 50% of them told us that it was extremely useful all right in their day-to-day job so that thought that was impressive that that that does that warms my heart so so now that we're through the research what what are you looking forward to next kind of from this week right and from co-pilot at large well two things number one of course we're going to be rolling out co-pilot to all of our developers as quickly as we can we're super excited about that that's awesome yes so we're going to be much more aggressive about that and of course we're excited to continue partnering with GitHub right and offering some co- solutions to our to our clients to help accelerate adoption in the market um but but personally what I'm excited about Ryan I I'm super excited to see where co-pilot goes when we get it out of just the code editor and bring it into the rest of the software delivery life cycle I love it we saw a little bit of that this morning i' I've got some good news though uh I'm actually over the next 45 minutes or so we're going to see a lot more detail and depth in that um in fact we're going to kind of like take a tour here through three different chapters we're going to talk a little bit about developer experience we're going to talk a lot lot about custom models because I know that every single customer and developer that I talk to asks about how we can get more customization for these models and we're going to take a deep dive into extensibility for that I actually want to invite one of the one of the Terri most terrific partners that we've had over the course of the last year um Harold a product manager from the vs code team to talk a little bit about how we've been working on developer experience in the ID together Harold come on out buddy come on let's do it thank you Ryan and thank you to all the early co-pilot chat adopters in the audience I'm Harold part of the VSCO team and I'm going to take you inside and behind the scenes of co-pilot Chad in my editor let's kick off the demo small apps are fun for showcasing AI but the in reality we all work on large and more complicated apps I work on vs code so let's contribute to its open source repo and here's our mission vs code's editor background recommend some must have shortcuts let's add co-pilot chat to that list as an essential tool to anybody's workflow which chat sty though in a typical vs code fashion co-pilot chat offers a customization the face that adapts to your coding style let's start by having co-pilot on the left but that HS the Explorer and other views I often need dragging it to the right and the secondary Side Bar lets me keep it always on on the big screen I can also use co-pilot chat in the editor area for ta to in workflow on small screens but my favorite is quick chat which I get from the command center or command shift I shortcut it provides a keyboard driven and minimal UI keeping me in the flow so I'll pick quick chat as the shortcut to recommend and to be clear this task won't be complicated but often even adding a simple feature can take hours out of your day as you try to understand the code base and figure out where things need to go copal can help me do that so much quicker so today in co-pilot ched I can explicitly mention at workspace to pull in relevant information from my code base I'm going to get a rough description of what I'm looking for and watch the magic happen wa let's hit the pause button though and delve into the Nitty Gritty how does co-pilot actually get my code AI needs just the right amount and the most relevant data for a given question copilot Chad achieves this with a hybrid approach first since my Reaper is on GitHub copilot Taps into github's Knowledge Graph a semantic sech returns relevant code Snippets and metadata and we're starting to roll this out to more users and more repositories next to make a search work in any workspace we layer in a local index which runs offline even for uncommitted files this uses a lexical search which means it's fast and more exact finally vs code's language intelligence is the syntactic sugar on top it has crucial details like function signatures parameters and even inline documentation all of these are ranked slic and summarized for the AI making copilot chat an in-depth expert of any code base so let's back to the action copil tells me the relevant code snippet that ground its answer grounding improves accuracy of responses but also makes them verifiable with links after copal sifts through the code it suggest starting in the watermark editor that's what I've been looking for let's dive in with the file open cop can now help me make a plan by default it focuses on the current file but I could also explicitly reference other files there's another Improvement at play here that is harder to show and I can't have an AI demo without mentioning the word so since recently gp4 is used in all chat conversations we still use other models like GPT C5 turbo as it speed complements GPT 4's intelligence this multimodel approach is informed by continuous experimentation to give you the cutting Aji with the feel of a lightweight responsive tool so back to my Watermark file I've got a rough plan of what to where I could do this manually but let's have co-pilot do it for me with inline chat hit command or control I and based on what I asked it it now pulls in symbol definitions surrounding code related files and more the diff view is fully interactive with linting hovers and Rich autocomplete and since we added inline chat 9 months ago its fan base has been steadily growing developers love iterating with AI directly in their code so much so that since September our early adopters have used co-pilot chat more often in line than in the sidebar staying flow is key talk about flow who here spend hours crafting perfect prompts for the AI just to get help yes Floy days is also about avoiding long wind and prompt crafting and code responses buried and long AI responses co-pilot has commands to solve that to Dem my favorite command fix I need some red squiggles let's rewind and say I had attempted to write the new entry myself and got the types wrong in any chat input here with inline hitting slash gives me a list of commands these are shorthand for chores and even longer workflows Allison demoed some of them so let me close this and show you an even faster way fix is just a click away on every red squiggle open the code actions menu or hit command Dot and fix is not just the well-tuned prompt it pulls in relative problems like linting errors looks up code from stack traces it retries if the AI missed arrows on its first pass it's a lot of power behind one command not surprisingly fix is co-pilot's most used action so let's see how our change looks I'll run V code from source and here here is our new favorite shortcut it's a good time to commit with co-pilot coming along check out that magical co-pilot button one click and I have my message ready I save time and can go right back to coding as I added to this file I do want to leave it a little cleaner so commit and this time I'm controlling inline chat using my voice hit the microphone button or assign it a shortcut manage entries as collection of key value pairs sweet realtime preview looks stunning and more importantly gives me time to double check copilot's approach so let's iterate with a followup rewrite entries to be alphabetically sorted clean up like that would have take me several minutes with co-pilot chat it's done in a few seconds without typing and back to committing as you might notice you will probably commit more often even without thinking about it but why stop making commits easier co-pilot knows about my code and what I've changed it even drafts a PR title and description for me and that is the real power of AI in your editor removing that tedious busy work we all must do ATT tax on productivity that we no longer have to pay when I get that review with change co-pilot can address that too even though AI suggestions are just a click away I'm still in charge of reviewing and iterating with co-pilot and there are many more smart actions saving you time in all areas of vs code when copad is installed in a terminal copad chat suggests commands in jupyter notebooks it fixes runtime errors it even Powers a smarter search in command pallet and settings customization accessibility and accessibility are what V code users love and AI is part of it so this was my behind the scenes of three major updates how co-pilot Chad gains expert knowledge of any code base how GPD 4 is one of multiple models it picks from and now copal infuses ai in every aspect of your developer workflow all of that everything is available today in vs codes co-pilot Chad all of these you will also find in Visual Studio at jetbrains is ready for you to preview via weight list thank you and happy smart e [Applause] coding all right yes cool so we've heard a little bit about developer experiences I promised a good Deep dive into custom models and so let's do it like I said every single conversation that I have with a developer invariably goes into how can you personalize these models how can you create them so that they're genuinely learning adapting to me and my team's preferences and styles and and all that good crazy stuff well we've been doing exact ly that for about the last year experimenting with GitHub and Microsoft's internal repositories and our internal engineering teams we've learned a few things you know through fine-tuning and just fine-tuning alone we can convince these models or tease these models into doing some pretty pretty cool things we can improve the overall suggestion quality right improve the acceptance rate and the characters retained by delivering more relevant suggestions we can introduce bias into the models so so that they show a preference or bias towards our preferred API version numbers sdks libraries and packages so that the models feel a little bit more I don't know at home for your engineering team and there are many times where teams might be working with genuinely unique code bases that are not well represented in the foundational models this comes in a couple of different flavors uh sometimes it's well it's really old code frankly sometimes it's like Cobalt and Fortran stuff that wasn't really well represented in kind of the the cloud-based repositories where a lot of the original training data comes from but can also come in for organizations that have their own proprietary languages for really specialized purposes right now you might be asking yourself how do you do this right how do you actually fine-tune these models well we've got three kind of categorical approaches that we take to this so first we take a look at your repository data and yes that means your code right but you know we all have code in our repositories that we're proud of and then there's some code that we're maybe not so proud of right it happens it's cool usually that code that we're less proud of it's code that's been lingering around for a while it's maybe code that we wrote three five more years ago and so what we do is we introduce kind of preference or bias into the models for the more recent PRS that were merged under the kind of the Assumption here that whatever you're merging today likely passes muster for your organization's best practices we also use signal coming back from the suggestions that we're presenting to you in uh in the IDE because well if you're presented a suggestion and give it the thumbs up accept it in your code well that's probably a good signal that it is worthwhile code if you pass it by or move on to the next one well it's probably a a good signal that that was not great code at least not for your purpose and turns out that that signal that Telemetry that we get from Co developer activity there's a lot more of it there are hundreds or thousands of decisions being made every hour every day and that volume of data can really really benefit the fine-tuning algorithms and last we use reinforcement learning from Human feedback right in order to essentially build an in between model you can kind of think of it like a um a sentiment analysis on suggestions that sits in between the large language models and the IDE to sort of filter out suggestions that may be good or may be bad in all of these techniques together we've seen significant uplift in our own internal experiments but you know what it's not enough to just to experiment on ourselves and we do you know we like to drink our own champagne eat our own dog food as it were but you really need to diversify the number of data sets that you work with and so for that we looked to a few trusted external Partners in order to continue our research so one of those is AMD and for that I want to invite Alex out on stage to talk a little bit about our research to together come on out Alex yeah hey I love it I love it hey thanks for having welcome my friend welcome yeah dude for having me all right well so we're here um tell me a little bit about kind of like AMD and what made you interested specifically in exploring this kind of research with us so the AMD we build the hardware that runs all your applications yeah right part of it is know Hardware design part of a software development to put together a system right what we found this copile is great it works wonderful on you know Ruby you're just flattering us now you're just flattering us now on JavaScript so basically if you're doing web applications you golden right totally but it doesn't really work well on lower level languages and if you asked me last year it didn't work at all on Hardware description languages such as verilog or system verog and this is where yeah we started talking and we decided okay let's do this and then figure out how do we solve this challenging problem right so I love it I love it and so like as we were approaching this I know that like one of the things that that I you know I've known about Hardware Developers for a long time and device driver developers is that they love their vim and they love their IDE right now fun story about this one so when we're conducting our original experiments you know we needed to work inside of an ID where we could have good kind of telemetry feedback and so we kind of required developers to use vs code specifically for this experiment what happened well I mean if you ask your friends who are die hard you know Vim supporters right switch what do you think the answer is going to be it's going to be you know hell no you'll pry it from my cold dead head right so but in this case the power of the model that we developed was actually compelling enough for some of them through that period of time to switch to vs code to use it to get the fine tun Model results which it blew my mind yeah yeah yeah I mean like shows the power of curiosity and shows the um the power of um maybe the potential of the large language models here exactly yeah and so we were going through it what was your team's experience with kind of the fine-tuning exercise what sort of results did you see so you kind of touched on fine tuning in the general process um let me give you some color like um and uh we started with you know identifying representative code and then we gave that to GitHub to do the fine tuning right so in our case the code consists of ver log system ver log and C okay um and you know what we found you know a couple weeks later after the initial I would say exercise is that AMD had a custom private co-pilot model that worked better on some of our specific languages right so just to describe the experience so you know that cop experience that you guys all love right it's autocomplete it's getting the next line it's doing documentation so imagine this now working on verog where a year ago we could dream about that mhmh so that's kind of number one and that's really early results uh we continued optimizing like team through Telemetry yeah with Telemetry and we found two very very cool things I think yeah right so the first thing is that the style of code that you know we got from the custom fine tune model was actually more in line with what our Engineers produced at AMD so more like AMD like style of code and remember I said hey we gave some C code right and you may Wonder like why like we want to enable ver log so why why give C code well co-pilot current you know product in the market support C so it gave us a good Baseline so through internal polling through some of the surveys we found there's actually significant preference to go to the Fon model with even with c um than the Manila model was put that way U so that was that was great and I guess if I were to maybe unfold it a bit further the potential here is that we can have multiple co-pilot models they are domain specific per team and their style of you know design right maybe it's a firmer team maybe it's a soft team maybe it's a web development team right and that could be tailored to them as you know the best possible co-pilot experience so that to me was very powerful I love it and in fact over the last several months a lot of what our team has been working on is really re architecting the backend of co-pilot because we foresee a world that is not a single model world it's not a single generic foundational model for everyone but is rather a multimodel world where each organization n even each team within an organization if they've got different set of uh onboarding docs or a different set of kind of standards and practices that they work against it will be a multimodel world where we're each matching the model against the job to be done or the task that we were doing that day right yeah and even this early kind of results they show that already I love it at least in our experience now the second thing actually blew my mind and you see a quote up on the screen uh it was made by Thomas who's one of our Engineers uh it blew his mind right and when we're talking to him um you know he said hey this is really cool so in order to introduce I'd have to explain a little bit and draw a parallel between software developers and Hardware developers okay and guys let me know I'm assuming mostly the software developers here you want to run quick you want to iterate you want to get to your minimum viable product you want to prototype you want to just rebuild you have your cicd set up right does that sound all right yeah yeah all right when you make a maybe a introduce a bug what's really the cost of that well just fix the bug spend a couple minutes rebuilding redeploy boom off you go right for Hardware design it's a little bit different you know the logic that these guys do it's you know tiny wires smaller than your hair printed on a piece of glass going through a factory for months so at that stage if you make a bug you can just hit rebuild right your rebuild means that months are lost and millions of dollars are wasted on a respin you got to go back to the fact re exactly back to the factory right so that's all to say that the hardware guys are really really good at writing requirements and writing specs because they don't want to you know they don't want to do this whole Factory cycle thing right um so in AMD process often times you know they write plain English specs right so on this screenshot you actually see a spec for usb4 router state state machine right but it's not just a spec this is actually a co-pilot prompt right this is actually the cool thing about it so our Engineers took this verbum copy pasted into into the ID uh added the first line saying hey generate me a more State machine and added one more word in the middle says this is where transition start the rest of it is absolutely unchanged from the spec and what came out if you want to fast forward to the next one is a pretty accurate set of verlock code right so imagine this right so a product development cycle now you can automate part of this translation from play English into ver log um and shorten directly our product development cycle so that was just unbelievable that is trick that is Trick man well I love it I love it well I thank you so much for doing this research with us and um thank you for advancing it really for everyone here uh so private models are available via weit list on github.com we do hope that more Partners will come and research with us as we bring it to Market here in the the coming year now fine-tuning is one way that we deliver customization but there is a whole world of additional data that exists outside of github.com it is in your Splunk and data dog instances it's in your launch Darkly instances it's all these other places all these other tools that you are using for development where co-pilot is essentially blind until today until today so for this next portion I really want to invite onto the stage Shen to talk about how we were expanding the GitHub platform to invite all of you to integrate and develop and to create an AI based economy with us with that Shen come on out thanks thank you thanks R and Alex how amazing is it now co-pilot can customize co- suggestions tailored to your co-base imagine how much productivity boost that can unlock for our developers really exciting what if we extend that even further not only we customize for you but also invite you the entire developer Community to build and customize co-pilot together with us that is what we envision the future of co-pilot to be and eal system with Marketplace of plugins integrating with thirdparty developer data tools and services together we can innovate reach millions of developers all across the world and transform the way the software is being built and elevating GitHub co-pilot the AI PA programmer to assist all aspect of developers workload so deeply engaging with the developer community and Building Product together has been our success recipe and we would like to use that recipe to build our Collective success and maybe the greatest one yet together so across copilot ecosystem as one girls we all grow so what do we mean by extensibility and how do we build the uh ecosystem together so so building on the fine tune capabilities that Ryan talked about the co-pilot extensibility enable customers to further personalize your co-pilot experiences through two high value scenarios expanding co-pilot context and knowledge and extending capabilities so today I will work you through a few integration examples we hope by sharing these examples um we we can you know uh get you energized uh unleash more creative ideas that we have yet to imagine so together we can innovate and unlock the most value for our developers community so first expanding co-pilot contacts and knowledge by bringing your own data so in the broader world and every business is software business that's built by developers so within each business organization ation developers need to build with internal guidelines documentations or business specific context and I'm sure like a folks in this audience all have experienced the pain of spending time searching through documents scattered around multiple places just in order to find the most relevant and upto-date information with ability of bring your own data the pain will go away and now developers can access all the relevant information from the singular Place co-pilot chat so let's work through how this will work let's say I run the internal GitHub engineering system so there are set of documents created over the years that cover the best practices like common engineering patterns or incident review processes and I want every developer in my orc to get co-pilot suggestions that are tailored to this knowledge so how do I bring this private info to co-pilot so as a GitHub customers platform customers I can just log in to github.com and follow the onboarding steps so first create a doc set and give it name and set the uh visibility to private so only people in your organization could access to it and second select the list of repos where the docks are being stored it could be single repo or multiple repo and just like that and now I have registered a private doc set for co-pilot to reference so for developers in my engineer Arc they can just go to copilot chat to Runing up on internal GitHub engineering system and asking questions like what is kubernetes which is a internal GitHub tooling system for managing and provisioning kubernetes and also asking question like special best practices like how do I prevent csrf happen in.com and copilot answer given to me is very GitHub Centric so a very straightforward process is for a very powerful scenario right so what happens behind the scene is the creation of the dock set kick off indexing pipeline that create the embeddings and store them in the vector database so when the user queries come in co-pilot retrieved the required uh snippet and use it to Grant co-pilot answers so what's being played here is a very popular generative AI tag a retrieval augmented generation rack I'm sure like folks in this audience know oh know about that so besides rag the co-pilot platform also does the ground work to optimize the storage and cury performance so ensure co-pilot responded users uh in a prompt way so that is one of the two high value scenarios of copile extensibility expanding copilot context and knowledge by bringing your own data the other high value scenario is extending co-pilot capabilities through building plugins so developers today they have to work across variety of tools like editors to write code database tool to manage data CS to to manage deployment observ uh observability to to like monitary service health so there's a lots of contact switch just to get the job done so with co-pilot plug co-pilot plugins this will change so for integrators by building plugins you bring you tools and services to where the developers are for developers by accessing your tooling need from the singular place co-pilot chat you get further productivity boost so there are two type of plugins n skill and agent so integrators can register them with co-pilot platform and have the choice of making them public for everyone to access or private only available for people in your organization so let's first demonstrate the new capabilities that the skill enables uh let's say I'm a developer working on kin's classroom project and during unit test I find a bug in the code and I can just ask a co-pilot to create an issue for me and co-pilot has the intelligence to summarize the issue based on the editor context like arrows and relevant files and prepopulate the GitHub issue template with a summary and all he left for me to do is to review and click uh create it's a pretty magical experience to accomplish all these steps with just a few simple natural language instructions so uh what happens behind the scene is when a user Make a request in the chat it hits this intelligent layer that does the internet detection and push in the force party and third party skill plugins as necessary so in this particular case it identif I users wants to create a bug so is Select and invoke the first party create GitHub issues and summarization issues to complete the task it's also worth noting that the created GitHub issues is a example of a first-party skill plugins that we're building to expose the rich GitHub capabilities so out of a box all of the copilot users will get access to these first party GitHub plugins that are naturally composed with third party ones so that's skill plug-in so for integrators who wants to have a greater control over the user experiences and you can do that too with agent plug-in so let's see how the agent integration experience looks like so data Stacks is a database service provider through the agent plug-in data stack Developers can now directly access all the application specific data from a singular place the co-pilot chat and they can ask a questions like tell me more about my database what tables are in this database and how do I query this database so they don't have to go to a separate website to manage the data like what they do today so again really keeping developers in the flow meaning like further productivity boost another like magical experience enabled by this agent uh plug-in so how does this experience work so our partner at data Stacks they develop this agent plugin that connecting the co-pilot chat with the Astro DB web service and register with co-pilot platform so when the users issue the query co-pilot platform invoke the data Stacks agent plug-in which knows how to interpret the user's natural language queries and cause the right API to complete the task so a very simless integration enabled by the agent plug-in it gives integrators the full control over the user experiences and also the flexibility to define the type of functions and capabilities to bring to co-pilot so to recap uh we just went through a few integration examples of co-pilot sensibility expanding copilot context and knowledge by bringing your own data extending capabilities with plugins so hopefully by sharing these examples we get you energized by the endless opportunities ahead of us so as we embark on this new journey together we also must acknowledge the potential risk that come with it powerful new technology require us to be more thoughtful and proactive so at GitHub not only do we committed delivering high quality product but also we stay true to the safety security and privacy core principles so we'll be publishing guidelines best practices to our integrators and review the plugins when they are brought into co-pilot so want to take the flight with us be sure to apply for the GitHub co-pilot partner program let's shape and build the future of co-pilot [Applause] together awesome well thank you so much for joining us on this tour uh and deep diving into developer experience into custom models and into extensibility with us building the future together is such an important important piece of this but there are more people than just the folks here in this room and the folks in the audience for us to thank here today Shen yeah so like before we end today's presentation we would really like to give a big shout for every co-pilot team members folks in GitHub Microsoft VSS vs code and Azure open AI is truly a privilege to represent everyone's hard work on the stage we will also like to take a moment to to thank you the entire developer Community listening to your feedback seeing your post about what you love what features you're hoping for what could be better this has been the through line of what motivating the team so on behalf of entire copilot team we hope you enjoy using co-pilot as much as we are building it thank you thank you

Channel: GitHub

Views: 52,104

Keywords: github, git, git and github, github tutorial, how to use github, github for beginners, code, coding, programming, developer, software, software development

Id: AAT4zCfzsHI

Length: 46min 56sec (2816 seconds)

Published: Thu Nov 09 2023