Cloudflare is on fire with AI 🔥 Interview with Craig Dennis

yeah we are here in this beautiful uh office cler's office uh in San Francisco we are joined by Craig Craig you're developer educator here at Cloud right yeah I'm very happy that you agreed for this thank you for Aging for this interview I'm very excited that you're here again yeah because we've been here for hackaton and now we are here for the interview uh we are extremely happy because we both are super fans of of cloud Flur so I just want to mention that to the our viewers that this could be bias a little bit maybe yeah we are all about technology so I think it would be it would be good um so before we start uh I would like to ask you Craig how did you get uh how did you start programming and what's your story yeah how did you get involved with AI yeah so so two two stories there uh one uh when I first got started programming I I come from like a it's a very place of privilege right so my my dad worked at a Y2K company so the first time I app applied for a job I applied to KFC Kentucky Fried Chicken you know that and then a Y2K company and KFC didn't call back so way back in the day I was a a youngster and I was putting stickers on tapes and so uh uh I had a lot of time and I just learned I learned I like kind of did some self selftaught stuff of of like how do I how do I automate some of this work that's happening here and oh look batch files and then from there I just started going automation is usually the starting point right when you you're thinking about programming and when you want to go get get into programming you thinking about how to automate things and that's usually the the beginning for many for many people and which programming languages did you start with yeah so that that's a you know this a this a long time ago uh so it was like a batch file a Windows batch file right so so like just like that was the the level of automation we needed we were setting up laptops for people and we were actually like putting discs like dises into into machines this was a long time ago um and we needed to make sure that we repeated the process the same way so it was like how do we make this happen how do we automate we said when you're ready press this button and it we just go through the program setup and things like that and then and then I you know as as it comes from there like I I was often thrown into the deep end I I started with some Java in this this program language called Lotus script which is back in the day for Lotus Notes remember yeah yeah so um yeah and AI so so definitely like often I was thrown into the deep end and AI is very much The Deep End right like I I I'm sure that I'm sure that everybody all the the viewers have had this experience where your your boss is like we need to add AI to this and you're like I don't know how do I so so I had that I had that experience um we were I was at twilio uh and um we were looking to build hackathons to let people um see what would happen if they worked with data in Ai and I had never I mean I had done a little bit of chat GPT at the time um and so how do we let people build with this this tool and so um we came up with this idea of how these people are coming and we don't know what their companies are so we don't know we don't have any data for them so I was like what if we generated a company so like I was like could it do that and you know like that's that's what I love about AI is like you're like can it do that you're it can and so it's like you know like I I gave it a prompt and I said you you are a creative entrepreneur um in you always start companies and I want you to start a company in this industry and it did it and I was like in the sunglasses industry and it said I said what's the name of your company and it said shazzle and I was like oh that's a great name and then I said what's the mission statement of shazzle and then it gave me a mission statement I was like whoa what is what are some of the problems that you're having and it was just you know it was hallucinating but in a nice way right so I took that and I saved it and then I you can replay messages right if you pass these messages and so then when we brought people in they played inside of this generated world where it was and then they were able to ask questions and it actually worked really well for a hackathon um educationally I think that's a really fun Tool uh to use before inter interview you mentioned that you working on a team that's called emerging Technologies right yeah and that's sounds pretty cool because it's something that uh you know it's like a state-ofthe-art things or like emerging things yeah yeah so so so um my team here uh at Cloud flare we're called emerging Tech and incubation uh and we run like a startup we get it we we see like hey how's this thing working like what can we get this out there can we move quick uh and we move very very quick I'm very impressed by by that even before AI uh I was following Cloud for a long time and Cloud seems as a company like the velocity is amazing like you're adding things very fast and even if it's not polished it's it doesn't really matter because it's like the this constant one after another feedback loop and Improvement so do you know what the secret is and I'll I'll let the secret out have you you've heard the dog fooding example right so for those of you who don't know dog fooding is uh there was a CEO that worked at a dog food company and he said if I don't eat this dog food how do we know the dogs like it something like that so basically like eat your product um some people say we got to sip our own champagne people like that more than the but then people say well I don't like the alcohol we got to come up with a new one because I don't think either one of those are great but uh so we use our own tools so that really drives things forward um and especially if we're in this this you know in this ETI and this emerging Tech and incubation we use the tools to build the tools and then when that happens you're like hey we need to fix this we should do this oh this can it do this it can oh we should make we should make a tutorial about the fact that you can do that you know that sort of thing and speaking about velocity uh you've just released like pre tools related to AI in C right right yeah workers AI uh AI Gateway yeah could just you are a huge fan yeah yeah we are could you speak briefly about each of them yeah yeah absolutely so um so workers AI there's there's this we we believe in the open models right so um I'm being careful when I when I say I don't say when I say open source because people will say well some people will get you know pedantically upset about that as they should because open source is a movement these are these open models these foundational models are there and there's fine tunes and and tunings that you can do um and they're available on hugging face but when you have to do that yourself you have to figure out where to go and where put them things like that so what we've done is we have um I'm about ready I was about ready to jump into our data centers but I haven't even talked about that yet for people who might not know about Cloud flare so we have um we have data centers all over the world uh because we are on a mission to help build a better internet so we know that the internet's not great everywhere and so what we've done is we've we've put data centers everywhere so the latency when you wherever you are in the world like Poland if you happen to be in Poland and you go and you hit hit that it will be as quick as possible uh because the data centers are living there that's over 300 locations over 300 locations right and so part of building a better internet helping to build a better internet what we what we saw was people are building these AI apps and they're going to have that same problem right so these new so so we said what if we sprinkle gpus around and we let people have this this quicker inference on you know it's popular for people to say like the edge that sort of thing we've been saying recently region Earth so so you don't think about it we you take that away um you hit these models that have been deployed for you wherever you need them right when you go when you go to use them so that's that's what we've been that's what think so that's that's what the workers's AI is so we've put a we've put a a rest API in front of it it's a single rest call to it um we have uh been pushing a lot of models to your your point you were saying that we've been moving fast we've been like trying to find out what exactly are all the use cases that are going to help everybody build these tools so we've been pushing up uh models that are in different use cases there uh and tweaking them to make them uh as fast as they Poss I was amazed by the number of of module of models that they are it's like I 20 maybe even more 50 53 53 53 and 57 today we'll get back to that but it's impressive uh I wanted to maybe step back a little bit uh because worker what is worker work is just a function right it's aun and what's really nice about Cloud workers is that they support web API right so they take request as a input and return response which is like a newer say way of thinking then let let's say the other approach is like no JS or something like that yeah right and Cloud workers say I not like improve on that right so then you can from the worker you can call the model directly which is is very convenient I would say and and this way you can like interact with like a bunch of uh interesting things and and build on top of that so I think the Simplicity is a good point right yeah yeah absolutely yeah so so and and with workers AI we want to make sure that you didn't need to be inside of a worker but if you were it should feel seamless and it does it's a really great feeling so so workers are that that it's an isolate is what we call them it's this V8 JavaScript Engine with the with the standard web apis and there is no cold it's just boom it's on and it's it's you know I know you're fans but it's fast yeah so then Cloud provides like a lot of building blocks I would say right for example we have workers AI we have workers workers are you can use AI with that we also have vectorize so we can use Vector databases and then you have all those models that can for example uh change let's say text to embeddings and then you can store them in something like vectorized could you maybe uh describe a little bit uh how it works how it works and maybe potential use cases yeah yeah sure so so Vector databases have been around for a while right so like we we I think if you think about like I'm looking for a product that's similar to this one you you encode what is happening there and then you kind of like compare them you get you get oh there's these three things are pretty close to each other so um the vector database is great for that what we've been finding is that for like um a lot of times what's what happens right these these models are trained and they don't have your data so what you need to do is you need to some way get your data in there and you typically your data is Big it's a lot it's a lot so so um we have models that will do some embedding so you might have seen that in open AI has an embedding model too so so you've you've seen these embedding models and there's there's some open ones so we have some open ones that are there and we uh encode them we make a basically uh turn them into to vectors which is like a think of like an array of floating Point numbers and we we can push them into this Vector database which we have which is on the edge as well so again deployed to region Earth you don't think about it it will be close to where your inference is uh we have a thing called smart placement where we'll get it working the mo the way that makes the most sense for your application because you might have outside applications that you're hitting as well uh but if you're using vectorized and workers a they're right next to each other so it will be extremely fast um and that database that that lives there you say I I want to get information I want you to find me the related records and then you get these records back and then you can go to whatever data store that you had or you could store met metadata in there and what happens is you take that information and you can inject it into your AI prompt so you're you're you're because you don't want to put all your data in there I there there are the the uh the interesting thing is that like there was a time where like you could never put all your data in there but now we're getting to a point where these models are getting so fast that it might be and I know that there's like a lot of uh companies that were like uhoh there goes our there goes our whole tool um but what we're finding is that you still need this you still need this Vector this Vector database because you need the the speed and the and the Precision you don't you don't want to push all that like you you could push all that data now some of these context windows are gigantic but that's expensive right because you especially for ones where you pay by the token absolutely and the the pricing model for incl is pretty like uh I would say flexible and like very transparent and and it seems very cheap at the same time sometimes it's even free pages right host the whole application it's not related to AI but it's kind of impressive that you can you can do a lot for free yeah you so that right like remember our mission right our mission is to help build a better internet and we want people to be able to build and and so if you're unable to build because you can't afford it that's not great right so we want to be able to provide that and what happens what we found happens we've done this from the start right so like when we've we've always had this like super generous free tier and the reason why we have that or actually one of the benefits of having that is we have people all over the world so we can see attacks the internet is constantly being hammered and attacked and there's different things and we can see that from around the world of what's coming in so so because we have all of these it's actually everybody wins by that so so you know we we like to say that like we want you to pay when you're ready to pay so so that that's what the free tier is about we want you to come and we want you to see and feel how it is and when you're ready you'll be ready and and until until you're not you're building and we want you that's what that's what we want you to do to be able to do so so we talked about the workers the eye and then we talked about Vector eyes and it's all super fast and it's on the like Global region yeah but then there is uh AI gateway gateway so what's the use case for yeah yeah great question if it's already so fast so yeah yeah yeah so not everybody's ready to switch over to us yet right so so that's that we we we know that we still want you to have a good experience right so the AI Gateway one of the things that it does that lets you run everything through right so you run everything through it so you can see and it's not just limited to our models right so any sort of if you're using a CLA or using a a GPT you can run it through through this AI Gateway and we you can see there if how um how many things are going up you see pricing and things like that and we're constantly adding to it so we're adding a nice firewall there where you could like make sure that all the things that were going out we safe I know there's like a there's a huge safety concern right uh you might have seen today that uh released lard 2 um which is it's so cool so so there's there's there's two concerns that happen I think in in uh and building AI apps people are concerned like uh oh no someone's going to say something bad to this and the model's going to say something bad and so what you can do is you can put this llama guard up uh in front of it we we have the ability uh to we give you the ability to do that currently to build that but there's no reason why that shouldn't be a tool that we also have that you just say turn on the guard and I don't want it I don't want it to anything unsafe to happen uh that I've defined because you can Define uh the categories that make it safe so so the G the Gateway is like a nice place to watch everything that's going through and um also you can switch right so if if you are running through the Gateway you can switch your back end whenever you want to so in situations where you know like it happens sometimes a CEO gets replaced for a weekend and you might get concerned you you could switch it you can switch uh if you're running through the through the Gateway you can you can do that so it seems on the on the one hand is very conv to use cloud on the other hand it's very flexible because you have all those options if you you not fully immersed in this ecosystem you can still have some like a uh let's say Middle Ground you sound like a fan no no it sounds like a fan it sounds like you're paid and just to be clear we're not we're not yeah no no just so maybe maybe to challenge you a little bit Craig because there is a lot of vector database out there still uh and I think uh cloud has this like a position that could win uh because you already have other things that are very convenient to use and as you said it's it's better to have back database that's close to you or reduced latency or what not and and how do you see this uh maybe competition playing out and the other question is do you think people because Cloud for seems ideal for building startups on top of and do you think that's that's a good way of thinking about that I mean yeah I mean like so like the again back to the back to the mission I'm going to keep my hit hammer on the mission um we want startups to succeed right we we want people to build and we want that cre creativity out there and we want to push things forward of how however it needs to go safely and securely and as quick as possible and so yes startups are that it's excellent for startups and we love startups I mean like at our core we we love startups right so um I I think that that's a great way to go but we also we also support a lot of the 4tr 500 companies right and so so they know us on that side right they they know us on that side and we want to let them also build and feel safe in the same way that they feel safe using our other tools so uh when you before the interview you said that you were kind of busy doing something important done sorry for being a little bit late uh we're launching meta The Meta 3 uh wait let me make sure I get it right meta llama 3 I need to say that correctly that I was told that that that's why we say metal Lama 3 is is launching today and it's like it's the next level of the state ofart state-of-the-art model uh is coming out you know that people build on top of these models as they go so I think we're going to see a really nice quick boom to it I'm I was getting the documentation out and I have a couple of uh really quick tutorials I have an idea of uh unboxing as a service so it's a little python notebook that you just kind of go and you get a play with it right away right CU our models are out there and our models are that are beta which this one is coming out in beta are free so like all of the inference is free so like I I I I try to yell that from the rooftops here in San Francisco like y'all you could get this this for free just go it's an API call so yeah go play because I think it's super important right I think it's super important and I I know that like we were also talking before about San Francisco we're very lucky to have so many people who are experienced with it but there's a lot of people out there who are like I don't know I I I don't have the time to do this I don't have a credit card to do this this is free yeah that that was my next question because there's a lot of uh uncertainty like it's notot we don't know about the Future How It will play out right right we are just playing with something new we don't really know how it works uh some things and how how you see the future do you think it's like more on the are you Optimist or are you pessimist oh I I am like I I think the best thing for literally everybody to do right now anybody who's even getting started coding or even thinking about coding or never wants to code is to touch AI because it does stuff that you don't think is possible and it will change what you think you can build right like I I I like I said I've been in I've been in this game for a long time we were talking Y2K times and it's stuff that like it does stuff that I was told growing up I'm never going to be able to do that I am never going to be able to look at an image and say this looks like that like I can literally do that I could literally build an an app at a hackathon that I never thought was possible and it's here and it's accessible and um English right like like it's in English half of the stuff that half of the amazing stuff that we've been building is literally in English uh it's not even a coding language anymore so uh I I really would love to encourage everybody to touch it and play with it because it will change what you think you can build and and I we're seeing that I mean like you're seeing that here in San Francisco the amount of startups that are coming here that are just doing things that like oh it can do that play with different ideas and like yeah yeah yeah yeah just just having fun yeah just having fun exactly and with Cloud I think it's it's convenient at the same time you're have this like a huge infrastructure so if ever your idea becomes something like more than just an idea you still can scale that's what we scale yeah that's what that's what we want we want we want you to have that and we want that success right so that's that that mirrors both sides of of of what we were talking about there with like if if we could like maybe evolve a little bit on this point uh what recommend in general to people starting maybe today they want to learn programming and should they learn programming or should they start with AI should should they maybe just dive into Cloud as dogs yeah yeah yeah so I'm I uh I've done a lot uh in my past of like teaching beginners had of code and I think uh that if you are thinking about learning to code you're probably going to end up AI using AI to cheat and when you do that know that that is part of AI that is part of what you're building right you're going to be like so like you you also do uh education as well right so so one of my I I would teach beginners python right and so one of the things that I would do in there is it would end with uh we're going to play a number guessing game because you get variables and you get Flow State and you get uh you talk about loops and you talk yeah all all that stuff so so at the end here's all the stuff that you did we're going to build a little game and run it I can go and I can say can you make me a game that does this and it does MH now what's impressive about that it will generate it'll generate the python I can also take an English and say in the system message I you are a game show host right so when you talk to these assistants you talk to it and you say you talk to the system message you say you are a game show host you are going to play a game called guess a number the number has to be between one and 100 and all this and the user is going to try to guess and I want you to make up like fun game show host things and it will go and will actually run it will run a chat Cod you there's no code it's all in English uh and and when you take that and when you when you see that and you see that it it is capable of because because right AI has context and AI has a loop and we're pushing that the stuff through and it's more creative than the code I would have written myself you can see that magic you can see that power there I actually have a talk I have a talk where I where I do that and I take it to the next level where I I build an entire karaoke assistant so like like talking about like what you want to motivate and it's sending you links to YouTube things all in English it's it's uh I like what is this world it's it's Bonkers it's so Bonkers it's such a Bonkers St it's huge and in a way it replaces uh this old model of like a m teacher and and and a p pupil right like a person like a student who's because you have this feedback lo you can generate on your own right you can like have this because what you described could also be used for you to brain ter the idea you're working on on your own we do this every day right yeah yeah and in a more efficient way or like faster or without like all the it's it's incredible it's absolutely incredible I I recently pushed out an app that I didn't write the code for I I I I didn't even make the design for a designer said can we make this and I said well let's see and I dropped it into into a chat GPT GPT 4 and I said M can you help me make this and I was like oh actually I don't know can you make that Tailwind cuz I want to learn tail and I went and I saidit what does that linean and it literally taught me like exact what you're say so I do think that like that is that is how you should learn to code now um I actually was going to make a video a video series where where I replaced myself in it you know like where you're like I'm I'm an educator here and now uh you don't need me I'm out of here no longer need it it's a it's a it's a very interesting change of approach because uh I used to have some students uh learning how to code and they were ashamed of Googling things they didn't know because they f they they have to memorize everything and now with AI it's it doesn't even make sense to memorize like almost anything right right partner yeah there's like and there's there's there's Concepts and I think it's like about it's really about getting that that system message right so like if you could learn to write that system message to be like hey I literally know nothing about this I need you to explain line by line when you write code I want you to put a comment on every single line it works it will do that and it will like that back to that number guessing example it will write a line out for every single thing and it's they're pretty good comments they're better comments than I did when I taught the course to give to somebody I agree and and the code got the code that's generated by AI got much better than it used to be a year ago for example and it is it is constantly getting better it's acceptable I would say right now yeah and to your point uh because we are talking about starting with something but also if you're like faced with a large code base right right something which is very complicated you can use AI you can use those models to help you understand it more like efficient in a more efficient way so it's or a complex database complex databases like relations all that yeah yeah and you can put an English level in front of it too so like you know the level of like uh we need a report that does this and this oh actually now can you make it do this and you've gone and you've I don't know you've created some sort of jur ticket that's gone through like two weeks of stuff when you could have just put English in front of it and said ask the question and had it generate the stuff and I've seen amazing stuff with that like that and that's what I think is really powerful and like something that I I want to make sure that everybody knows is that literally every industry is getting because in the beginning we already talked about this your boss said I need you to put AI on this literally everybody has done it a little bit and where it's leaked out you're seeing if you could talk in English to a thing that's because somebody put AI in front of it right in in this in this text generation AI world right so basically after assembly then see then maybe I don't know typescript python probably English is the next programming language that we'll all use yeah AB absolutely absolutely I I bet I bet you know this this metal 3 I bet that when uh you play with this it's going to be better right because it's because it's progressing it's it's progressing and and what even my it understands better it understands it is better at uh when you when you talk to it mistakes that you make the same way that like we would make mistakes and it would get the squiggly line it knows now and it can I it's Al it's also faster right it's f F it's faster yeah there's just like infinite patience yeah absolutely absolutely I I'm a little bit concerned about when it learns about itself right like like you know how many blog posts like Well it can't do this it's like well it can but it's red that it can't so maybe I will ask you about this uh because we were like discussing it what risks do you see with AI what's like maybe yeah if we are talking about risks basically or something that could go wrong space I I I think it's moving so fast right and I think that um I have seen bureaucracy and companies right that like is there for a reason right like when you get to a certain level you need to make sure you go through all the legal things and all the concerns and we need to make sure that those people's voices are still heard um and also need to move fast so we're I I think there's a there's a concern in there and there's a lot of there's a lot of great minds out there that are thinking about how to how to make this safe and and uh and copyright and legal and all that stuff but we're still ripping fast right like I'm still I saw Mona Lisa rapping yesterday and I was like is that allowed like that's not that song and that is like literally the Mona Lisa are you allowed to do that with the monisa yeah I was I was impressed by those like models like sunno AI right it's the the songs generated are mind absolutely wild absolutely wild yeah and I think that like you know people will say like well that won't ever be something that I listen to and it's like you know how many times I've sent those links versus how many times I've sent a song to somebody lately like yeah it's so I I and there's there they do a great talk that that that crew does a great talk about like they're just changing what it means to share music um because like you're sharing this more personalized right it will be something that you like and it will C to your like a maybe around you also short funny story he sent me a a song that was I generated and I didn't know I I didn't realize yeah it was so good yeah it got that good yeah wow did you put in your playlist in your rotation it was like a funny song you know like a little bit trolling but I I thought just some YouTuber recorded it you know or or something I don't know that's awesome and Craig I wanted to ask you um about because we talked about uh those new things like related to music do you have other favorite projects that you are seeing outside let's say cloud FL what you're working on or or inside or maybe inside but things that you are maybe interested in or you want to start getting more into yeah I I think like in our in our video education world right like there we you you were at the hackathon this week but there's there are companies that are doing really cool stuff with AI and video and I would love I this has been my dream forever and I was told it's not possible and so now I'm like aha wait a second maybe it is possible so uh I would love for you to be watch you you watch a video you watch an educational video and you turn your camera on and I read your face and when you go H like oh you look like you might have cuz that's what I do in the classroom right so I I I taught in the classroom interesting and I miss that I missed the ability to be like o that didn't land but if I have if and and then if so so you see their face and you kind of control uh what's happening there and and Fork right like you you have recorded videos already but you Fork off into this other place and you you say oh did are you having a problem with that they're like or they're yes right like what whatever then you you continue going down this path things it's close and it's close like you know you like I mean it's not I I wouldn't I wouldn't subject anybody to that yet but the tech is close and it's good and it and that unlocks so much I agree because what's missing right now is like all this non verbal communication right but we we got pretty close with audio communication real time you can interrupt the model yeah but now when I see you Craig you're are smiling or maybe are sad on the question I ask something I can see that and this is something which is missing uh still in those things so that's that's pretty cool idea I I I uh it works for like Industries too like if you're like uh standup comedians need to do these these horrible open mics like you need to go to an open mic and like just bomb and see if the joke works or not right and what if you just did that you had you had an open mic where you had a recording everything and you just crowdsourced oh they laughed at that you know that was a good one like that's a it's a thing like and we're it seems wild but we're here I've saw some people playing with smell too okay so like they're trying to figure out how do we synthesize smell and I was told that's never possible and again like that's pretty cool idea actually that would totally change right and like you know it works well actually I don't know if you works at Burger Kings here they pipe a smell of like but it's not coming from the place I don't know how they do it or what they're pushing but like when you drive by like oh that smells good oh it's Burger King but it's not it's not there I thought it was the burgers I'm sorry to real for you yeah so uh Craig thank you very much for this interview for your time it was amazing to talk about all those things uh and looking forward what cloud first will announce next a and uh yeah thank you very much you so much thanks for us here yeah awesome good to seeing you guys yeah let's do thank you let's let's smile and W to the camera

Cloudflare let you run machine learning models on their serverless GPUs global network. Their built-in vector database, Vectorize, is not only convenient and easy to set up, but also affordable. Plus, you can enhance and protect your AI apps with their AI gateway, which offers rate limiting and smart caching to keep costs under control. It's not just about the range of tools—they're also incredibly quick at delivering new features. Cloudflare ships updates at a rapid pace and is constantly adding new capabilities to their AI offerings. Recently, Michał Miszczyszyn and I had the chance to sit down with Craig Dennis, a developer educator in AI at Cloudflare, to talk about their exciting AI tools and about AI in general.