Transcript

字幕

(00:00) Hello welcome to the devrel survival guide in this episode we are going to look at how AI tools can help you in your devrel practice my name is Matthew Ravel and I've been speaking to devrel practitioners and also people who make AI tools to see what they can do and maybe even if they're a threat to us as devrel practitioners before we get into this episode I want to thank common room for sponsoring you can head to commonroom.

(00:41) io to learn more about the platform that brings together your CRM product and Community data and is used by some of today's fastest growing companies that's commonroom.io thanks for sponsoring let's clarify some terminology there are lots of different types of AI and what's been in the news lately are large language models or llms so let's look at what they are my name is Logan gilpatrick I lead developer relations at open AI focused on helping support developers building with our API AI as well as building chat TBT plugins

(01:13) which is our sort of emerging developer space so the the general take is that large language models is essentially a machine learning model that's been trained in the large Corpus of text in the the whole intent of a large language model is just to predict the next word or the next token in a sequence so you you know give it a a sentence and it's literally just trying to predict like what is the next sequence of characters that's most likely to come after this and you throw in a little bit of Randomness for the the sort of outputs

(01:44) to be um more more Broad in in general but that's that's the basic take is lots of content that it's been trained on and is trying to predict the the next piece of content that's coming after this is as an aside this is really important to get a sense of like how the machine learning technology works because you start to understand that like uh Chachi BT for example has no world view like it doesn't have a lot of these um you know constructs that we as humans have because like it's truly just trying

(02:15) to do that one specific task it can do it really really well which is why it's able to generate content that's you know across all these different domains but like fundamentally it's like super constrained in the um if you want to think about like the logic of how it approaches solving a problem like it solves them all in the same way using that same sort of next token prediction process so then how have these AI tools been helping people with their devrel work let's say from three people who've been putting them to

(02:44) the test my name is Tessa Merrell and I'm head of developer relations at operate so I've recently asked different team members including members from other teams what kind of tools they're using and how we're using it and I realize it's for a lot more areas than I thought specifically for a team we're using it for content mainly for social media and content all four presentations we have support Engineers on the team so we're using it to help the support um it's helping us with evaluating

(03:20) analytics and market trends and data and research in that area our Engineers are using it for debugging I think that's like the main the main areas yeah my name is Matthew Groves and I work for a company called couchbase it's a nosql database company I've been there for seven years my main concern is helping developers and helping them to succeed and helping them learn about couch base right now I'm very much focused on uh building an example application trying to build as much as I can with the llm

(03:53) tools that everyone is aware of seeing how they work seeing where they are where they Excel where they are not doing so well and just trying to put together a story about how these tools can be effective in in coding column Doyle who heads up devrel for intercom today we're pretty pretty light about it I think the biggest thing is like like the speed at which you can create code samples has like gone through the roof so it's like whereas previously creating code samples say illustrator point would have taken that I don't know

(04:28) like a couple of days you know what I mean it's a bit like a decent kind of end-to-end uh code sample now it's like should be seeing co-pilot are writing a bunch of it which means like when we have a new API endpoint or um we want to support a new language uh it's much easier for us to just kind of jam out a code sample and because we have all the context of her API works we can kind of see when the the llm is screwing it up you know what I mean and we're like oh yeah well that doesn't make any sense that's wrong but uh but

(04:58) it's a slight tweak so you're more kind of editing existing code rather than generating it fresh the company I work at intercom has a has a support bot which is powered by various llms but we haven't turned it live on our docs so we haven't turned it on yet as like a layer above uh our documentation but like that's definitely interesting so what tools are Tessa Matthew and column using day to day yeah so I'm focused on uh three tools mainly at this point for coding as chat GPT of course

(05:29) uh for both three and four uh GitHub co-pilot chat and a co-pilot in general and then Google bard so I've looked at all three of those tools and I've spent time with them I actually have a live coding stream on Twitch where I explore all those tools as I build this real world example application uh so I've I've tried all three of those and uh I've got all kinds of interesting things I've learned about them along the way the sort of the end goal with this project is to create a real world

(06:01) example application that can help developers get up and running quickly with in my case asp.net and couchbase but we're exploring other ones as well like spring and node and things like that so that's just kind of the short-term goal is to get those example projects available for helping developers of course the long-term goal is for me to understand or try to help understand how effective these tools are in building applications and the goal there is to come up with a plan or roadmap to build future sample

(06:33) applications for other languages and platforms you know we're a database company and we serve uh you know 10 to 12 different language communities so using this to kind of help to learn if we can scale those example applications out to all those different communities instead of just focusing on one at a time that's a goal here is basically to help people write more code with couch base right and this might be a way to help them write more effective code or help us to write more effective example applications faster you know I'm focused

(07:08) on building this application in a language and a platform that I know well right so if there is something that's generated by chat CPT or co-pilot that jumps out at me as wrong for security reasons or wrong for performance reasons you know that's something I'm able to understand and address immediately right that's more challenging if I'm working in a language that I'm not familiar with if I try to build something in Rust for instance I might have a hard time with that so uh it's not uh that I'm building them

(07:38) completely with these applications just copying pasting exactly what they are um generating but I'm also using them to help understand uh what is being built and why it's being built so that's something that that these tools can also be used for so people focus on the uh code generation a lot and that's it's pretty cool but it's also useful for reading code and as developers you know we spend a lot of time a lot more time reading code than we do writing code so this can be a very effective tool at

(08:10) understanding a legacy code base for instance you know put in a function that's poorly named or badly documented and explain what this code does help me out so I can you know make changes to that code or or write better tests for Tessa and the team are upright AI tools have made their way into almost every area of devrel operations we use it for improving our email campaigns and evaluating data with that our operationless team uses it or we're like improving their emails being sent you can use it for our travel

(08:44) itineraries like we're going to a conference in two weeks can you help build out an itinerary I think googlebard can also do a lot of similar things as well we also have like a developer experience engineer on our team he uses and I had no idea and he was using notion AI it helps you like automatically like finish sentences it helps you format things make it look nice and the amount of content AI tools out there is just growing very quickly what you know copy AI for improving content there's like an AI tool called Genie

(09:25) um I was just reading about today that one of our team members uses for and that specializes in summarization of text just interesting I still haven't compared it to chat GPT but he says I do well it works better with that um right now we're experimenting with different video AI tools and there's several of them popping up here and there are those AI tools to create present like a slide deck for you um I wouldn't say that's like the greatest technology yet there's still a lot of work that needs to be done

(10:00) I feel like creating presentations is more of a a creative process especially in the Deborah side of things I know there's like AI image type of tools but that's like a dangerous area to explore especially with um working for a company as there's like no milk regulations around using certain tools there's also an interesting tool I've been using as well called Jasper and Jasper specializes in creating blog posts but I'm still not a fan of exactly how these AI tools generate blog posts every

(10:40) time I read someone else's blog posts I start like identifying yeah that was written by AI that was written by AI I could have just taken the topic title and write this myself and then read it on chat GPT rather than going to those blog sites and reading it there so the kind of content I'm actually really valuing is listening to people's perspectives hearing about their opinions so if you're able to take AI generated content and spin in your opinion and also train it to speak like you there's a lot of

(11:15) work you can do so it takes a little time to get used to understanding what prompts to use when working with AI to be able to get the results that you're looking for it's clear to see that AI tools are already having an impact on how we do devrel on a day-to-day basis but what about the longer term approach how does column doiler intercom see the next few years and maybe the changes that they need to make at their company to take AI tools into consideration yeah uh to change our approach uh I'd might

(11:51) be generous um I think or like I guess what it is is like my experience of building both the samples and like I had to build something recently internally um and I'd say I don't even know what percentage of the code I actually wrote myself but because it was like just a scrappy internal until it didn't really you know it wasn't too heavy but as I wrote those things like all I was constantly thinking was like right if everyone's going to be writing these things like this what do I need to change about my

(12:21) documentation approach in order to make it so that the likes of GitHub copilot or or Chachi BT can like easily ingest the content we're creating right so first of all that means there needs to be a lot more text um I love video as a format I think it's it it accommodates you know a really popular learning style was kind of useless to bots right unless you can feed it the transcript but visual stuff is kind of useless to these Bots particularly the popular ones is my understanding anyway um so it immediately puts more of an

(12:55) emphasis on like right we need to write tutorials we need to be pretty um Direct in the tutorials like we need to explain things um in in quite plain language and which we should probably be doing anyway if I'm honest like uh of acronyms and stuff like that it's it's bad practice anyway and then more of the code samples because obviously co-pilot is trained on on corpuses of public repos uh so there's a basically there's more of an emphasis away from video uh towards stuff that llms can consume

(13:32) because like I said if a lot of these apps are going to be written using um using copilot or you know people Consulting gpt4 like I've never used as much red I say I've used more regex in the last three months than I have in my entire career combined because regex is terrible to learn and that's a really really good at it so it's that kind of thinking it's like and like I am constantly thinking like how do we need to change our docs and I don't know the answer if I'm completely honest and like I I really need to talk

(14:03) to the ml team at my current place to understand it better but like what do I need to adjust about our documentation to make it more llm friendly there's some really interesting I wouldn't say problems but like challenges um to that and the two big ones that stick out to me are the first one is where do you point one of these Bots at a web page right you have to be relatively smart because web pages are busy they have lots of different things going on at any one time right so if you think of like the guardian website or

(14:34) the Wall Street Journal or whatever if you go to their website right you're going to have like a central pane of content then you're gonna have like on the sidebar you're probably gonna have um links to other articles or like a byline or something like that and and you're gonna have like ads and you're gonna have all these different things going on so like um processing text is expensive for these llms so they try to be smarter some companies try to be smart about right what is the essence of this web

(14:59) page so I'm only feeding the essence to it and in the example I gave elect Guardian The Wall Street Journal it's usually a safe bet that like the central piece of content contains and the vast majority of what we eat so um a lot of scrapers will take just that Central piece of content and they'll discard the rest and that works great to an extent for like Wall Street Journal or other Guardian but that kind of completely falls apart with a lot of documentation websites right because you've got the

(15:25) central pain which is the story so you've got the central pane which is explaining the core concept and then on the right you might have code samples on the left you might have like version Pickers or or and different languages and all this kind of stuff and usually in Docs the whole the whole viewport is is kind of important so if you slice out just the core central part you're going to lose a bunch of context the other one that folks have talked to me about is um llms are just really bad like versions

(15:53) uh the way a lot of dark sites are structured there's usually like a quick picker to pick between between versions and it's hard to explain the notion of API versions to a bot right so if you ask it how to do something so like if it's slack it's like get me uh how do I get all the you know all the messages from a given channel right if there's multiple different versions about to do that the bot's not really going to know which one of those it should give you um so that's a bit of a challenge that I

(16:21) haven't quite figured out if I'm honest and that's part of why we haven't turned on and the support bot on our own documentation is because uh the existing Technologies we've used and I have have trouble with it right they did not they have trouble but they don't give us the results that we're comfortable with so we wouldn't ship it to customers and therefore we won't use it ourselves I don't know if I'd go as far as changing our versioning strategy to accommodate other lands but like

(16:46) thought about it clearly these tools have their weaknesses so what else should we be on the lookout for I guess one of the most obvious things is that some of these tools are limited to 2021 and earlier as we're recording this anyway and so sometimes it's generating uh code or patterns that's out of date because we're you know two plus years past that this point so you know these these tools are still learning right so they're still um growing and learning more things as they get more input so I I find it to be

(17:24) helpful to be as specific as you can right if you wanting to use a specific SDK version tell the chat bot that say I want to use couchbase sdk3 because otherwise it might imply based on the year that it has cut off that you want to use the most popular SDK back then which might be version 2 or 2.

(17:47) 5 or whatever so it helps to be very specific uh when you're talking to these tools and understanding that it may be giving you outdated information these AI tools is that'll help us to be more effective we'll get more done or we'll save time but if we need to babysit them and constantly check that they're not hallucinating or otherwise making mistakes are they actually saving time oh it's 100 saving time it's like you can focus your energy on on the bits that are specific to your API they're like your knowledge right so

(18:23) if you're writing a typescript sample right you know you could have a code sample that would have like switch statements and different control statements and all these kinds of things and like I said earlier regex all right and you can use all these things to prove it right but like these are kind of like Road knowledge and they're coming to all these kind of different platforms and of course like for an experienced developer you don't think a lot about writing these things but like with the with the likes of copilot is

(18:44) like you literally don't have to think about it you could just like start to type that or like type a comment to give it a prompt then you can just get return return return until you get to the bit which is specific about the the concept you're trying to demonstrate in your API or the or the programming Concepts you're trying to you're trying to emphasize which means you can focus in on that right and because you haven't had to think about all the other so if you're naturally saving time

(19:09) which you then spend the same amount of time maybe polishing and refining your code maybe but in my experience it's it's just uh it's it's just like like I said about refining the concept that you're trying to talk about it I guess it leads to better content because you've been able to spend that just a little bit of extra on the bit you care about as opposed to like assist the most efficient you know control statement for this thing or you know am I importing this right using whatever language best practices I don't

(19:42) know if it's there yet but like you know you can imagine a world uh to speed these other items Advanced you know a world that rise very quickly where you actually write your code sample in pseudocode and then a bot you just have a build pipeline that like spits it out in lots of different languages like that's going to be really powerful for for developer Advocates with so many new AI based Tools around then presumably there must be some AI tools that are built especially for the needs of devrel teams

(20:11) and there are in this interview recorded in the summer of 2023 Contender CEO Lilly Chen describes an earlier iteration of their offering yeah my name is Lily Chen I am the founder and CEO of contenda contenda is a Content Catalyst platform for developer Advocates specifically it transforms uh videos into written tutorials with code samples and all those nice things one of the big issues with the video transcript is that it's not meant to be read so as a human if you give me a transcript it's a very very painful to read it's not a good

(20:46) experience so what we do is we take that and then all the information that was extracted on the video and we asked GPT hey like could you convert this into something that's more readable into more of a blog post and then we do some checks on our sides and make sure that all of the content is accurate is correct it matches the original author's intent and voice uh and that's what we deliver if Contender uses AI to help us generate content what about the community side of devrel he is co-founder and Chief

(21:16) Architect at common room Tom Klein Peter yeah comic room uses AI in a number of different ways we have kind of sprinkled it in in all the places we thought that make a lot of sense so we'll do things like you know kind of basic starting with things like sentiment analysis where you have content flowing into your community it's nice to be able to tell what is positive what is negative and to keep metrics on those over time and to kind of it's uh just to be able to sort of see where things are trending in your

(21:52) community that's kind of one of the more basic things we do with it we also start to categorize messages that come into a community where you can tell what just show me all the product appreciation we've seen recently or show me all the bug reports to where it's it's faster to go and filter the signal out of the noise you have all this content flowing into a community you have people and you have content and llms they can understand language give us an opportunity to extrude all of this into repeatable shapes where we can

(22:26) standardize things and then have metrics and Reporting and sort of know you know the sentiments training over time we've got more bug reports over time or we release some new feature and Bug reports went down so the first big thing we get is just better consistent repeatable understanding of the content that's happening inside of a community I think a second big thing we do with it is understanding that people in the community more we use an ai-backed merging system where we can identify that different accounts

(22:58) in the in your community are actually the same people which gives us just a better understand or gives the whoever's running the community just a better understanding of who is in their community and if you want to correlate some you know happy comment on Twitter with some more report that got fixed on GitHub by being able to merge people together you know we can do a better job of that and merging is incredibly complicated there are tons of signals and different ways you can look at this and one of the things AI is really good

(23:28) at is just dealing with more information than humans can't and so we built our merge algorithm on top of that and we're pretty happy with it so that's content and Community what about code I'm Martin Woodward I'm the VP of developerations for GitHub GitHub co-pilot is an AI pad programmer we call it so it's a coding assistant that lives inside of your code editor and helps you do your coding with code completion or you can also kind of ask it you know do like chat functionality and things like that you

(23:59) know as a devrel person you tend to be coding a bit but you're not coding all the time but you know how to code you know you know programmatic thinking but you might be having to use a lot of different languages or maybe a lot of different Frameworks to then show somebody how to code from where they are into your thing you know that's that's a lot of what other real work is and that's creating those code samples and things that's actually what copilot is awesome uh to help you with um in terms of you know code assistance

(24:36) but also you know maybe your chat say Hey how do I do this in Python I kind of know how to do it in Java but how do I do this in Python and it'll give you an example get you the framework the skeleton then you do the thing that you want to do and it can assist you there so really it's it's it's it's what it says go by is a co-pilot it's not the thing that's driving um and so for somebody who knows how to code then you can be incredibly productive with it and it allow you to write a lot of stuff and do stuff and so

(25:07) for Deborah where we're flipping languages that helps a bunch around AI I find while codepilot will help you in writing code what about creating Integrations here's Martin Davis from superface.ai uh so at the high level super face uh abstracts apis into the business cases that you need to achieve when you're building applications as a developer and those abstractions are turned into what we call com links short for communication links um these Comics they are structured representations of how an application

(25:47) needs to communicate with an API in order to achieve that particular use case which might be sent them SMS or sending an email or you know get me this list of users or add people to this list uh in a particular platform anything that you might need to achieve on a kind of everyday or repeatable basis um you're thinking about it in that use case form super face kind of breaks that down turns it into structured tools that you can then use and you know continue to use inside of your application so we've got developer specific Tooling in

(26:21) the form of a CLI SDK that make creating and working with those comments actually possible but ultimately it's AI um that's under the hood that's responsible for figuring out what that communication needs to look like and and it can do it simply by looking at you know the documentation for an API or an open API specification and then it forms a plan that is represented as this com link that then gets turned into code that you can use directly in your application via the SDK so that the aim is to create like this unified developer

(26:56) experience for you whoever's implementing um these use cases uh and an interface that stays the same regardless of the provider you use so if you've got multiple different email providers and multiple different Communications providers your interface through superface would be exactly the same so your inputs would be the same your outputs would be the same but you can swap out the providers uh just by generating your new use cases so it has an idea of how that API needs to be communicated with the idea here is that

(27:27) you it's less labor intensive to get up and running and you go much faster time to hello world whilst presenting developers at the end with code they can control not a chat bot that spits out an amalgamation of code that somebody else wrote at some point in terms of how the super face uh could help devrel folks I think we're offering another integration Channel effectively so you know especially for those that API first devrel folks those you know that love their open API specifications superface is a great tool to help

(28:02) demonstrate those apis with a use case first mentality which can really help especially when you're thinking about documentation and example creation like super face could very well sit alongside your node.js SDK Ruby SDK as a pathway for developers to consume what you do in their own applications so it's uh I think superface is going to be a good tool for developer relations practitioners to uh to employ and to deploy at some point and then there's the intersection of it all between code content community and that is support

(28:41) how do we offer better support more timely support more appropriate and relevant support to developers using uh apis and other tools through AI here's Deepak Kumar from Doc e to talk about their Tool uh we help companies grow their developer adoption so I have a dream that one day developers will not have to wait through pages and pages of documentation to be able to understand an API with docky we are taking one step closer to a world where I can get as a developer the precise accurate information as and when I need

(29:29) it to build the software that I really want to build and we leverage AI plus human control in a big way to deliver that to our developers docky can help you at three places number one add answering your community questions which generally is outside of you know like the the group where your company is operating so this helps you with the expansion of you know addressing different concerns that people have as we answer lots of questions hundreds and thousands of them at the end of the month we generate a developer pain points report

(30:16) which essentially is your advocacy to your r d and product team as to where your developers are struggling and third we not only do advocacy but we exactly help you create the content which helps the team in addressing those gaps because if people are asking questions there is certainly something missing right a person would have gone on Google and other places searched for at least 15 minutes they could not get help that's why they came to a devrel uh person or a community so we start with support do advocacy and

(30:56) then help you with exactly filling the Gap that we figured out in that case our model is to bring help where developers are since developers are in slack and Discord so B bring you help where developers are as a bot for slack and Discord so we've heard some of the good things about AI tools with devrel but what about some of the trickier aspects if you've spoken to someone who does developer advocacy on behalf of an AI tool you might have come across the idea that sometimes it can feel like you're

(31:32) working with a moving Target here's Martin Woodward from GitHub to share his experience could be a nightmare to do devrel for if I'm honest and the number of times you sort of try and get it to redo the same thing and do it this in the same way and it doesn't because um these coding assistant tools are non-deterministic um they have a certain amount of Randomness in them and remember they also are learning from what's around you and what options you're picking and so sometimes it's quite hard to get it to

(32:03) do exactly the same thing um the reason why is really interesting actually because the way that these large language models work is um they basically are predicting what's the next thing you're about to type the next thing you're about to say um it's a bit like that Meme on social media where you you start typing something like you know this podcast is awesome because and then you hit the first word the the the the autocorrect gives you every time and then you get an amusing sentence sometimes but it turns

(32:35) out actually when we've been doing you know research with the um llms that actually it the most interesting sentence or the best result isn't when you pick the most likely response every time but if you pick a response that's you know maybe occasionally you'll pick not the most but maybe the second most or the third most likely response every time and that's what they call that Randomness um that entropy that they call that temperature um in a lot of these llm models so you set a temperature setting and about 0.8

(33:07) usually gives you a good setting of a kind of the usual kind of you know written prose type stuff so that's why it's non-deterministic but oh my goodness you're only trying to create a video you get co-pilot and you're playing with it and you cut yourself coming up with a good demo scenario and then it just gives you an amazing response oh that's fantastic and so you like right I'm going to record that and so you rewind your demo you try and do it again and it it either gives you a completely different response or

(33:35) sometimes when you're doing devrel for Copart you're trying to explain to people how how it learns from you and you know how to like prompt it better so you maybe get one response which is okay you say oh yeah but how about I give it a better prompt and so you give it a slightly better comment or something like that and then it gives you exactly what you want next time you come to then give it kind of what you want it'll often just shortcut straight to the answer you definitely wanted because you

(34:02) know it's learning and so that's that's the um that's some of the hardest things about doing a sort of co-pilot with um with uh you know for devrel I want to create actually a faux pilot um where it uh where it just kind of mocks where it records real responses back from the the service and then I could repeat my keystrokes kind of thing and it plays it back I keep meaning to do that so I can kind of Replay some of these but uh yeah otherwise it's good fun and it's certainly speeds you up and

(34:30) you get a lot of those oh my goodness I can't believe it just did that moments all the time you know it's amazing let's wrap up then with some thoughts on what you as a devrel person might consider when it comes to AI tools but also what the future for these AI tools holds so you can start to plan over the next 12 to 24 months how are you going to incorporate them into your programs one question that frequently comes up with really any new technology is will it take my job here's Lily Chen from

(35:04) Contender with her take I would say that develop people should not be worried about the growth of AI and what it means for them the value proposition for devrel internally is that you bridge the gap between engineering products and then the user base writing tutorials is part of the way that you're communicating to the user base but it's actually not the end job for you there's lots of other things that Deborah people do that cannot be automated by an AI I believe communication you know open AI put out a report about what are the what

(35:34) are the skills that are threatened by Ai and what will become higher in demand one of them is critical reading and communication let's wrap up then by considering what these tools might mean for the next 12 months and we'll leave the final word with column Doyle of intercom devil people tend to be on the bleeding edge of things they tend to like to adopt adopt new new things right but but uh that sometimes it creates like a um uh adjadedness in US you know what I mean oh this is a flash in the pan it'll

(36:06) go next I think like the only advice I'd offer is like the the llm things are like not a flash in the pan like they're I I've people I respect have been like no this is like on the order of the iPhone in terms of a technological shift um so I guess my advice would be to not treat it like a Flash and ignore it and give it the intention it deserves like think about do an audit of your docs like pointing out of them in your docs and see how they do just go into copilot in gpt4 and try to write something with your apis

(36:39) right now and see how it responds and because if it responds well you're doing something good and you should double down on it if it responds poorly then you might be missing at an opportunity [Music] foreign foreign

(00:00) デブレル・サバイバル・ガイドへようこそ このエピソードでは、AIツールがデブレル実践においてどのように役立つかを見ていこうと思います 私の名前はマシュー・ラベルです デブレル実践者、またAIツールを作っている人たちに、AIツールで何ができるのか、そしてもしかしたらデブレル実践者としての私たちにとって脅威となるのかどうかを確認するために話を聞いてきました このエピソードに入る前に、スポンサーになってくれたcommon roomに感謝したいと思います。

(00:41) ioで、CRM製品とコミュニティデータを統合し、今日最も急成長している企業のいくつかによって使用されているプラットフォームについての詳細を学ぶことができます。

(01:13) これは、私たちの新興開発者スペースのようなものです。一般的に、大規模言語モデルとは、基本的に、大規模なテキストコーパスで訓練された機械学習モデルのことです。

(01:44) 一般的にはもっと幅広いですが、これが基本的なテイクで、学習されたたくさんのコンテンツがあり、この次に来るコンテンツを予測しようとしています。

(02:15) その1つの特定のタスクをこなすために、それは本当に本当によくできるのです。だからこそ、すべての異なるドメインにまたがるコンテンツを生成することができるのですが、基本的には、同じような次のトークンの予測プロセスを使用して、すべての問題を同じように解決するように、問題を解決するアプローチ方法の論理について考えたいのであれば、それは超制約されているようなものです。

(02:44) 私はテッサ・メレルで、operateでデベロッパーリレーションの責任者を務めています。最近、他のチームのメンバーも含めて、様々なチームメンバーにどんなツールを使っているのか、私たちはそれをどのように使っているのか聞いてみました。

(03:20) アナリティクスや市場動向、データ、その分野のリサーチなどに役立っています。エンジニアはデバッグに使っています。

(03:53) みんなが知っているようなツールを使って、できる限りサンプルアプリケーションを作ることに集中しているんだ。ツールがどのように機能するのか、どこが優れているのか、どこが劣っているのか、そして、これらのツールがコーディングにおいてどのように効果的なのか、そのストーリーをまとめようとしているんだ。

(04:28) それはつまり、新しいAPIエンドポイントができたときや、新しい言語をサポートしたいときに、コードサンプルを詰め込むのがずっと簡単になったということだ。

(04:58) それはちょっとした微調整で、新しく生成するよりも既存のコードを編集するようなものなんだ。僕が働いている intercom という会社には、様々な llm を利用したサポートボットがあるんだけど、まだドキュメント上でライブにはしていないんだ。

(05:29) あー、3つも4つも......GitHubのco-pilotチャットと一般的なco-pilot、それからGoogle bardね......だから、これら3つのツールは全部見てきたし、時間を費やしてきたわ......実際、Twitchでライブ・コーディング・ストリームをやってるんだけど、そこでは、この現実世界のサンプル・アプリケーションを作りながら、これらすべてのツールを探求してるの......だから、これら3つを全部試してきたし、あー、その過程で学んだいろんな興味深いことがあったわ......このプロジェクトの最終目標は、現実世界の

(06:01) このプロジェクトの最終的なゴールは、開発者がasp.netとcouchbaseを使って素早く開発できるように、実際のアプリケーションのサンプルを作ることです。

(06:33) 他の言語やプラットフォーム用のサンプル・アプリケーションを将来的に構築するための計画やロードマップを立てることだ。

(07:08) 自分がよく知っている言語やプラットフォームでこのアプリケーションを作ることに集中しているんだ。だから、チャットCPTやco-pilotが生成するものの中に、セキュリティ上の理由で間違っているとか、パフォーマンス上の理由で間違っているとか、そういうものがあれば、すぐに理解して対処することができる。

(07:38) これらのアプリケーションで完全に、生成されたものをそのままコピーペーストして作っているわけではなく、何が作られているのか、なぜ作られているのかを理解するのにも使っています。

(08:10) レガシーなコードベースを理解する。例えば、名前付けが悪かったり、ドキュメントが不十分だったりする関数を入れて、このコードが何をするのか説明すると、そのコードに変更を加えたり、より良いテストを書いたりすることができる。

(08:44) 2週間後にカンファレンスに行くんだけど、旅程を組むのを手伝ってくれないかなとか。googlebardも似たようなことができると思うんだけど、うちのチームにはデベロッパーエクスペリエンスエンジニアがいるんだけど、彼が使っていたのは概念AIで、全然知らなかったんだけど、文章を自動的に仕上げるのを手伝ってくれるんだ。

(09:25) ええと、ちょうど今日読んでいたんですが、私たちのチームメンバーの一人が使っていて、テキストの要約に特化しています。

(10:00) プレゼンテーションの作成は、特にデボラ方面では、より創造的なプロセスだと感じています。AI画像タイプのツールがあることは知っていますが、特に会社勤めをしていると、特定のツールの使用に関するミルク規制がないため、探求するのは危険な領域です。また、私が使っているJasperという興味深いツールもあります。Jasperはブログ記事の作成に特化していますが、これらのAIツールがどのようにブログ記事を生成するのか、私はまだ正確には好きではありません。

(10:40) 他の人のブログ記事を読むたびに、「ああ、これはAIが書いたんだ」と特定し始める。

(11:15) できることはたくさんあるので、AIと一緒に仕事をするときにどんなプロンプトを使えば自分が求めている結果を得られるかを理解するのに慣れるには少し時間がかかります。

(11:51) 気前がいいというか......サンプルを作った経験と、最近社内で何かを作らなければならなかった経験があるのですが、実際に自分で書いたコードの割合がどれくらいなのか自分でもわからないんです。

(12:21) GitHubのcopilotやChachi BTのようなものが、私たちが作っているコンテンツを簡単に取り込めるようにするためには、私のドキュメンテーションのやり方を変える必要がある。

(12:55) 「チュートリアルを書く必要がある」「チュートリアルではかなり直接的である必要がある」「チュートリアルではかなり平易な言葉で物事を説明する必要がある」「正直なところ、頭字語やそのようなものはとにかく悪い習慣だ。

(14:03) 僕は常に、僕らのドキュメントをどう変更する必要があるのか考えている。

(14:03) 今の職場のMLチームともっとよく理解するために話す必要があるんだけど、私たちのドキュメントをもっとLLMフレンドリーなものにするためには、何を調整する必要があるんだろう？

(14:34) ウォール・ストリート・ジャーナルとか、そういうウェブサイトに行くと、中央のペインにコンテンツがあって、サイドバーには他の記事へのリンクとか、傍線とか、そういうのがあって、広告もあって、いろんなことが行われている。

(14:59) だから、私はエッセンスだけを送り込むんだ。私が挙げた例では、ガーディアンとウォールストリートジャーナルを選んだ。

(15:25) 中心的な痛みはストーリーであり、中心的な概念を説明するペインがあり、右側にはコードサンプルがあるかもしれない。

(15:53) ええと、多くのダークサイトが構造化されている方法では、通常、バージョンを素早く選べるようになっていて、APIのバージョンという概念をボットに説明するのは難しいんだ。

(16:21) 正直に言うと、私はまだよく分かっていません。そして、それが私たちが私たち自身のドキュメントのサポートボットをオンにしていない理由の一部です。

(16:46) 考えてみると、明らかにこれらのツールには弱点がある。

(17:24) 特定のSDKのバージョンを使いたいのであれば、できるだけ具体的にチャットボットに「couchbase sdk3を使いたい」と伝えると便利だ。

(17:47) 5とかでもいいんです。だから、こういったツールに話しかけたり、ツールが古い情報を与えているかもしれないことを理解したりするときは、とても具体的であることが役に立ちます。

(18:23) もしあなたがタイプスクリプトのサンプルを書いているのなら、switch文やさまざまな制御文や、さっき言ったような正規表現など、あらゆる種類のものを含むコードサンプルを書くことができますよね。

(18:44) 文字通り、考える必要がないんだ。ただ入力し始めたり、コメントを入力してプロンプトを出したりするだけで、APIで示そうとしているコンセプトや、強調しようとしているプログラミングの概念にたどり着くまで、リターン・リターン・リターンすればいいんだ。

(19:09) そうすれば、同じだけの時間を使って、コードを磨いたり、洗練させたりすることができる。

(19:42) それがまだそこにあるかどうかはわからないが、あなたが想像できるのは、あー、これらの他の項目をスピードアップさせるために、あなたが実際に擬似コードでコードサンプルを書いて、ボットがそれを多くの異なる言語で吐き出すようなビルドパイプラインを持つだけで、非常に迅速に立ち上がる世界を。

(20:11) そして、2023年の夏に収録されたこのインタビューでは、ContenderのCEOであるリリー・チェンが、彼らが提供する初期のイテレーションについて説明している。

(20:46) ですから、私たちがすることは、ビデオで抽出されたすべての情報を取り込み、GPTにお願いして、これをもっと読みやすいものに変換し、ブログ記事のようにすることです。

(21:16) コモンルームのチーフ・アーキテクトであるトム・クライン・ピーター（Tom Klein Peter） コモンルームではAIを様々な方法で使用しています。

(21:52) コミュニティで行う、より基本的なことの1つです。また、コミュニティに入ってくるメッセージを分類し、最近見た製品への感謝やバグ報告をすべて表示することで、ノイズからシグナルをフィルタリングすることができます。

(22:26) 物事を標準化し、メトリクスやレポーティングを導入して、時間の経過とともにバグ報告が増えたり、新機能をリリースしてバグ報告が減ったりすることで、感情の変化を知ることができます。

(22:58) コミュニティ内の異なるアカウントが実際に同じ人たちであることを識別することができるんだ。そうすることで、私たちはよりよく理解することができるし、コミュニティを運営している人は、コミュニティに誰がいるのかをよりよく理解することができる。

(23:28) AIが本当に得意なことの一つは、人間にはできないような多くの情報を扱うことです。だから私たちはその上にマージアルゴリズムを構築し、かなり満足しています。

(23:59) Devrelの人間として、あなたは少しコーディングする傾向がありますが、常にコーディングしているわけではありません。しかし、あなたはコーディングの方法を知っています。プログラム的な考え方を知っています。しかし、あなたは多くの異なる言語や多くの異なるフレームワークを使用しなければならないかもしれません。

(24: 36）だけでなく、チャットで「Pythonでこれをやるにはどうすればいいんだ？ コードを書く方法を知っている人にとっては、それを使えば信じられないほど生産的になれる。

(25:07) デボラにとって、言語を反転させることはAIの分野で大いに役立つ。codepilotはコードを書くのに役立つだろうが、ここではsuperface.aiのマーティン・デイビスがインテグレーションを作成することについて説明する。

(25:47) 特定のユースケースを実現するために、アプリケーションがAPIとどのように通信する必要があるかを構造化したものです。例えば、SMSを送るとか、Eメールを送るとか、ユーザーのリストを取得するとか、特定のプラットフォームでユーザーをリストに追加するとか、日常的あるいは反復的に実現する必要があるようなことです。

(26:21) CLISDKのような形で開発者専用のツールがあり、そのようなコメントの作成と作業を実際に可能にします。

(26:56) これらのユースケースを実装している人のための、統一された開発者エクスペリエンスと、使用するプロバイダに関係なく変わらないインターフェイスを作ることです。

(27:27) 立ち上げと実行に労力がかからず、ハロー・ワールドまでの時間が大幅に短縮される一方で、開発者には、誰かが書いたコードの寄せ集めのようなものを吐き出すチャットボットではなく、彼らがコントロールできるコードを最後に提示することができる。

(28:02) ユースケース・ファーストの考え方でAPIを実証することができる素晴らしいツールです。superfaceはnode.js SDK Ruby SDKと並んで、開発者が自分たちのアプリケーションであなたが行っていることを利用するための経路として、特にドキュメントやサンプル作成について考えているときに本当に役立ちます。

(28:41) どのようにすれば、より良いサポートを、よりタイムリーなサポートを、より適切で適切なサポートを、開発者に提供することができるのでしょうか。Doc eのディーパック・クマールが、彼らのツールについて語ります。

(29:29) 私が本当に作りたいソフトウェアを作るために必要な、正確で正確な情報を、開発者としていつでも得ることができる世界に一歩近づいているのです。dockyが開発者にそれを提供するために、私たちはAIと人間の制御を大いに活用しています。

(30:16) これは本質的に、開発者がどこで苦労しているかをRDや製品チームに提言するものです。第三に、私たちは提言するだけでなく、チームがこれらのギャップに対処するのに役立つコンテンツを作成する手助けもします。

(30:56) それから、私たちのモデルは、開発者がslackやDiscordにいるので、開発者がslackやDiscord用のボットとして開発者がいるところにヘルプを届けるというものです。

(31:32) 動いているターゲットと一緒に作業しているような気分になることがあるというアイデアに出くわすかもしれません。GitHubのMartin Woodwardが彼の経験をシェアしてくれますが、正直に言うと、Devrelを使うのは悪夢かもしれません。

(32:03) 全く同じことをさせるのは難しいこともある。その理由は、これらの大規模な言語モデルが機能する方法は、基本的に、あなたが次に入力しようとしているものは何か、あなたが次に言おうとしているものは何かを予測しているからだ。

(32:35) 私たちがllmを使った研究をしていてわかったのは、実は最も面白い文章や最良の結果は、毎回最も可能性の高い回答を選んだときではないということだ。

(33:07) 通常は、通常の散文的なもののようなものをうまく設定することができる。

(33:35) CopartのためにDevrelをやっていると、どうやって学習するのかを人に説明しようとすることがある。

(34:02) それは学習しているんだ。だから、それが......コ・パイロットのようなものをするのが難しいところなんだ......Devrelのためにね......フェイク・パイロットを作りたいんだ......モックみたいなもので、サービスから返ってくる実際の応答を記録して、自分のキー入力を繰り返すことができるような、それを再生するような......。

(34:30) あなたは、ああ、すごい、信じられない、こんなことができるなんて、そんな瞬間がいつもあるんだ。

(35:04) ContenderのLily Chenさんの意見です。開発者は、AIの成長とそれが意味するものについて心配すべきではないと言いたいです。社内でのDevrelの価値提案は、エンジニアリング製品とユーザーベースの間のギャップを埋めることです。

(35:34) Aiに脅かされ、需要が高まるスキルの1つは、批判的な読解とコミュニケーションだ。それでは最後に、これらのツールが今後12ヶ月の間に何を意味するかを考えてみよう。

(36:06) 次へ。私が唯一アドバイスしたいのは、llmのようなものはフラッシュ・イン・ザ・パンじゃないということです。私が尊敬している人たちは、技術的なシフトという意味では、これはiPhoneのようなものだと言っています。

(36:39) 今すぐgpt4のcopilotに入って、自分のapisで何か書いてみて、それがどう反応するか見てみよう。もし反応が良ければ、あなたは良いことをしているのだから、それを倍増させるべきだ。