[GHW GenAI] Fireside Chat with GitHub's Idan Gazit & Snowflake's Chanin Nantasenamat

本站添加: 2026-05-09

[00:01:08]Heat.

[00:01:24]Heat.

[00:01:51]How's it going folks? Rosundo from MLH here. Welcome to Global Hack Week Gen AI. We're really excited to have you.

[00:01:57]Obviously, things have been rolling along. you've just left the uh left the stream with Mary who uh gave you all the uh updates on global hack week. This stream is going to be focused on basically having a fireside chat with two of uh two of uh our guests from the industry uh Don Gazit and also uh Channon Cenamat from Snowflakes and uh from GitHub and Snowflake respectively.

[00:02:22]So if you have any questions feel free to drop them in chat. Obviously the questions we'll be fielding are regarding this stream. So, be sure uh if you have any questions about Global Hack Week in general, go to discord.mmlh.io to get set up there. Someone that's going to help you out. As far as getting checked in and registered for this workshop, head over to the link scrolling across the bottom of the screen there. You're going to get registered, checked in. We want you to get credit for being here. Obviously, just um really excited to be here. Let us know where you're from. Let us know where you're joining from. We always love to see that. If you're joining from MLH, if you're joining from dev, anywhere, any part of the world that you're from, we would love to hear from you. Uh we also have a really exciting exciting special guest that I hadn't mentioned before. So as my time here runs off just uh make sure that you do stay tuned in because we are going to have two sections of this fireside chat.

[00:03:09]It's going to be the first section with uh our friends from GitHub and then the second is going to be from with our friends from Snowflake and Streamlit.

[00:03:17]And it looks like we do have a couple of folks joining from India. Love to see it. Oh, I have such a nice Thank you. I appreciate it. Um, we got some folk we got folks from Pakistan, Sweden, very nice. Kenya, Nairobi, love to see it.

[00:03:31]All right, so without any further ado, I'm going to go ahead and bring up someone that you may or may not know, but is near and dear to the hearts of all MLERS across the world. Um, our very own co-founder, John Godfrey. Give him a warm round of applause, y'all.

[00:03:45]>> I wish I could hear the Oh, you have a sound effect. I was gonna say I wish I could hear the round of applause, but um, hey everyone. So excited to be here today. I'm John, one of the co-founders of MLH. Um, and honestly, I'm I'm personally uh really looking forward to this AMA that I'm about to start with Edon Gazite, who heads up GitHub Next.

[00:04:07]Um, for those of you who don't know, GitHub Next created Copilot among a lot of other really experimental and interesting ideas. Um, and Adidon has a really kind of cool background in tech, working on products like Heroku, open- source projects like Django, and a lot more in between. Uh, so I am so stoked to welcome him here to Global Hack Week.

[00:04:30]Uh, and looking forward to our conversation.

[00:04:33]Let's kick this off, Ro.

[00:04:36]Awesome. Welcome, Edan.

[00:04:38]>> Hello. Good morning to you all.

[00:04:41]>> Good morning. Thank you for joining us early in the morning. I feel like we have a hundred time zones from around the world here and you might be the earliest.

[00:04:48]>> Uh uh I hope so. It's I'm going to I need to be smart now like at 7 in the morning so it's all good. It >> it's a challenge but I believe in you.

[00:04:57]Um so I want to start off at at kind of like a really high level here. Um you know the theme of this global hack week is generative AI. And for me personally, like one of my favorite things about using these tools, especially coding uh generative AI tools, is the ability to quickly like experiment and play with ideas that I have in my mind and see what they look like in reality. Um, and I saw in another interview that you, you know, called GitHub Next this fool around and find out, which I think is an incredible descriptor. But I'm curious like for the folks in the audience here, you know, what does productive fooling around actually look like? Like how do you tell when something is a real idea versus just like a cool demo?

[00:05:46]>> Um I mean that's a that's a really good question actually. So it kind of it kind of depends. First of all, obviously at the end of the day uh we exist uh as part of a business. you know we have to deliver business value and I think um every incubation team from the side compared to like a regular engineering team um it seems like ah we just get to frolic in a field a green field of daisies and rainbows and we get to do whatever we want and that's true to some degree but like we have to prove that we're responsible stewards of that freedom you know like if if I just you know wander off and do stuff and it doesn't have any impact then um sooner or later somebody's going to be like why do we have this team around right so um that freedom needs to be bought by demonstrating that we're spending our time that we're placing our bets not every bet needs to succeed and in fact if it did that would also be a violation of our jobs because that means that we're being too conservative for a team that's supposed to be the scouting team you know if we're not out there scouting um and sometimes falling on our faces then we're not really out there at the edge Um, so, uh, we're looking for impact and that means finding new products that people want to use, you know, like we're the team that originally made C-pilot. Um, like that was not obvious. And hand on heart, like at the time I was also like, is this going to suck? Because every AI thing I've ever used before has sucked. Um, and then no, no, it didn't. It didn't actually suck at all. It was great. Um but but it wasn't obvious and it required quite a lot of sweat and and figuring it out. But um I guess to answer your original question initially what we're looking for inside the team is like does it make other people vote with their feet like on the team? Are they sitting up and saying that I want to work on that? That's that's that's the bees knees. Um so um so that's the signal that we're looking for. Now this is also a very unique team. It's a team of only seniors, right? Like there are no juniors on this team. Like just there's no juniors on seal seal team six, you know, kind of a thing. Um uh and so that experience that we hire for on this sort of weird team um is also that's the compass that we take with us out to see like you know the first signals that we're looking for is from them. But then we go out and we, you know, after we prototype the thing and it's important that we build it fast, we have to put it into people's hands and then see how they're using it. Like we're looking for product market fit. So take pretty much any book ever written about startups and apply it to like what do we do? Like you know try stuff, iterate fast, figure out if it's good.

[00:08:36]And not less importantly, it's also knowing like when to cut your losses and say this doesn't have impact. I'm not going to keep throwing good money after bad, good time after bad. Knowing how to say like, you know what, I don't got it.

[00:08:50]Let me move on to something else that has a shot of like escaping or uh orbit.

[00:08:55]>> So, um yeah, I guess I don't know. Messy answer.

[00:08:59]>> No, I mean it's a great answer. I I guess the question it presents to me though is like what made you double down on Copilot? cuz I remember what the early days of these agentic coding tools were like. And this is not like a knock on Copilot, but like there was a sort of battle between the um maturity of the models themselves and the maturity of the harness, right? Like C-pilot and they were kind of at odds for a while, you know, like people used to complain about like sloppy code and it doesn't feel that way anymore. But how did you know in those early days that you were on to something even though the output was in many ways like kind of uh you know like not where it is today right like maybe not even up to snuff you know >> I mean it's I like god I feel like a dinosaur but it's just like it wasn't even agentic in the beginning right remember it was just doing completions in the early days uh >> uh and those completions weren't good to begin with either right they were like they were good maybe 30% of the time uh at the very very early days and it it feels like yeah man this sounds this sounds like you know back in my day shy we walked uphill to school both ways um it was it was a like you know the original GPT3 which was then called codeex which then they used the name again like that was um uh because openai is amazing at naming um so like it was I think the context window was like 4,000 tokens so it's like a tiny context window compared to like you know today we'd be like what is this a school for ants um uh and uh and the the responses were not that amazing. It was particularly strong at Python and a little bit at JavaScript because that's what was over represented in the training set. It wasn't good at any language. Um, but at the time like seeing it actually complete the things that uh that we were about to type or to save me like you know I could write a comment and then um uh pro sort of prompt it that way because at the time there wasn't really prompting is just it was looking at the file that you were editing. Um, so like in the early days, as much as it sucked, we had to actually build the user experience around it to make it so that um, and this is a funny sentence to say, to say that it was ignorable. Like we needed it to be something where if the suggestion was bad, it didn't cost you time. Like we tried different interfaces like like sort of a Tinder style interface like swipe right to accept kind of a thing.

[00:11:33]But it turns out that that required too much commitment from me as a developer that I would need to like switch into like co-pilot mode. Um when what we wanted is that it it's like if you wait long enough it'll it'll show you something and if you don't like that suggestion you don't have to like cancel out of it. You just do what you were going to do anyway which is keep typing.

[00:11:53]So like success with this stuff and particularly when it was bad, when the models were not as good was also about the user experience that we wrapped it in. Later on as we moved into like this more agentic um uh style where I I give it a prompt and now I wait a minute, two minutes, five minutes for it to like do something. Now obviously people are like yeah I wait three hours and it implements a whole thing.

[00:12:18]>> Yeah, it's compilering.

[00:12:19]>> Yeah, exactly. Compiling just like the XKCD. Um but uh like so then the sort of the incentives changed and we got into this mode where we couldn't make it ignorable. Like you know in the original cop we had to make it tolerable when the models were bad. And then when we're waiting five minutes I don't ever want to wait five minutes and then be like nope that's trash that's terrible. So then we really had to focus on correctness. How do we like make the experience definitely hit the target or at least close to it? So like it's not so much like how did we know if it was good or how do we know how to make it good but instead it's just like at every moment in time based on the experience of the human driving these tools we had to focus on these different uh like things here we had to focus on like um making failures tolerable and here we had to focus on correctness even though they're both AI tools but like how you hold them matters and and sort of figuring that out is is just a part of the job. You know, it's an end to end like it's not enough to just have the best model. I don't think the models actually matter very much. Like they they they do but in a very temporary sense like but you know three minutes from now all the models are going to be great and then does it really matter which one? Like you know like my Ferrari and my Lamborghini both of them can drive 1,000 miles an hour but the speed limit is still you know 55 or whatever it is. So, does it matter how much better this one or that one is than the baseline? No, they're all they're all great. What matters is the experience.

[00:13:57]>> Um, I I kind of feel similarly that the incremental gains that you get from switching models every time a new state-of-the-art one comes out are are pretty minimal. But I guess like does that suggest to you that models are sort of um either becoming or already are like commoditized?

[00:14:21]>> Um yes. Um I mean I don't want to say commoditized but uh I will say funible >> in the sense that you know it's like mix and match. Sure. Opus 45 definitely a step change moment right um that one was like first past the post in terms of >> just an elevated level of capability of reasoning uh an ability to dispatch it on longer harder tasks >> but now like okay codeex 55 is also sort of like hitting that same level and you know the folks at Google are no dummies I'm sure that they'll like you know show up with like another banger and like Gemini whatever version is the next one.

[00:15:03]Um, uh, I think it trends towards everything being good enough. And the analogy that I use here is like it's like cell phones.

[00:15:11]>> I don't know like 15 years ago, you would buy based on who had good coverage where you lived and worked.

[00:15:17]>> And now everybody has good coverage everywhere. And so you buy based on other factors like will this one give me a new phone that I want and will this one have other like benefits. So it's like at some point like the coverage thing stops mattering and I think we're almost at that inflection point today.

[00:15:34]>> Yeah, absolutely. Um, so I I want to transition to another topic that is, I think, highly relevant to a lot of the people watching today, many of whom are, you know, relatively early in their careers. And there's this debate going on right now, especially in computer science classrooms, but I think in the industry at large too, about whether using agentic coding tools is um sort of like uh cannibalizing the skills development that you would get through traditional means and whether that represents something like cheating or if it's actually just the future of software development and needs to be part of the learning process. I'm curious like what your take on that is and what advice you would give to folks who are you know just coming up in the industry now.

[00:16:21]>> Um I have a nuanced answer to this. I mean is using Google cheating like >> or or Wikipedia when we were growing up, right?

[00:16:30]>> Yeah. Like you know it's just like hold on I'm going to go check out Stack Overflow. By the way, original copilot when we looked at it, one of the first things that we said is just like, ah, this is the Stack Overflow killer because like why would I go to Stack Overflow when it can just suggest the thing that is the amalgamated wisdom of the internet like right here and under my cursor. Um, I think no, I don't think that it's cheating any more than like like you know like ah like now we have electricity to do this thing that we had to do using like manual labor like is that cheating? No, it's just the advancement of technology. But it's more nuanced than that. It it's up to you and how you use it. Like, yeah, you could just stick your brain in autopilot mode, not apply your intelligence to a problem, and be like, AI, try it. No, try it again. No, like, you know, and just sit back and like, you know, hit the easy button. The problem with that is that then you're not actually developing the skills that you need in order to go from junior to senior because the difference between junior and senior is experience. It's wisdom.

[00:17:35]It's judgment. It's the ability to say like here are the different trade-offs and in this situation these are the trade-offs that make sense. It's not something you can be like ah like what's 2 plus two? It's four. Like you know no it's not like that. So, if you're using AI to rob yourself of like time hands-on like wrestling with like difficult subject matter and like thinking about it, then you're cheating yourself. Um, however, if you're using AI to like accelerate your own learning and accelerate your own trying of things, like I think back to like when I was in college and uh coming up in in the industry and like man, if I had had like a tireless robot that could answer all my questions, like, you know, even in the middle of the night when I'm sitting there like bashing my head against some exercise set that like, you know, my graphics programming teacher like set us and like the whole class is like on man this is going to date me like we're on like AIM or ICQ like you know instant messengers back in the day being like the hell are we supposed to do here are you getting this to render at all like you know I remember like talking with my classmates imagine if I could have had that back then how much that could have accelerated my ability to build and see how things are built like you know having just a tireless senior engineer on tap that's that never gets tired of my it like that could have been like amazing. So, I think it's in how you use it. It isn't like straight up good or bad. Like a knife. I could use it to murder somebody. I could use it to cook a fivestar meal. Like, you know, it's up to me how I use it. Um or maybe it's a chainsaw. Like, you know, careful how you use it. It's very powerful. Um like for good and bad, you know.

[00:19:28]>> Yeah. Um you just like unlocked a core memory for me. I still have my ICQ number burned into my brain. I can just recite it. I'm not going to cuz someone's going to look up something stupid I said when I was 13. But um >> yeah, that's no one in chat is going to have any idea what we're talking about.

[00:19:45]Um, so you know what you're saying like align with my experience of using these tools too, but like a lot of folks in academia are highly skeptical of this. And you see this thing going on right now where like people are doing coding tests on paper, right? because they want people to be able to articulate like pseudo code or how certain you know concepts work or things like that without consulting an outside resource. um you know I know you have a CS degree like we are of a similar age here like if you were going back and making this pitch to your most skeptical CS professor like what would you tell them because I think what you're describing here is like this very optimistic view of like how it can be used to improve skills and I agree with that but how do you convince a skeptic here who's in the weeds every day like teaching students who frankly you know I was pretty lazy when I was a student And so I would cut corners if I had the chance. Right. So like how do you kind of bridge those things?

[00:20:52]>> Yeah. I mean um when I think back to the best pedagogical experiences that I had in college um uh they were all open book exams. Um which I think is maybe a good proxy for the world in which we live today, right? like um the best professors were not interested in testing my memory like my ability to like recite facts on commands like ah what's the big old runtime of quicksort like I I don't know that's what Google is for back when I was in college it's not that there wasn't Google I'm not that old but like um but it wasn't like now where I could just like type in the question get the answer it was done like you know I'd have to go and do like a little bit of of homework so what's true now that wasn't true then is that my need to like have like a whole stack of trivia in my head around topics in computer science because like when I think about like CS 101, CS 102, what was I learning? Exactly that kind of stuff. Um uh does it matter for me to like have that at my command? Yes, you do need to acquire a certain baseline of knowledge in order to be able to like be fluid and fluent with that. But like after you're done with 101, 102 and you're starting to even like sort of in the upper classes of college, um you're like your teachers are trying to teach you the beginnings of judgment and tradeoffs and like okay like real problem solving and now it's no longer about just like being able to recite facts out of memory. So to those professors I would say like hey the world is about to change. All these facts are going to be like mine on command like from any device in my pocket. I'm gonna have I'm gonna have the equivalent of like today's supercomputers in everybody's pocket and they're going to have internet everywhere they walk around on the bus.

[00:22:53]They're even going to have internet on airplanes. And so my need to remember facts like memorizing the dictionary is going to go away. And all that's going to remain is the judgment part. And so like train people for that. That would be the the pitch that I'd make for them.

[00:23:09]And I'd say like here's the tools that you can use to help them with that that thing. So I still think that they'd say okay but like we still need people to have like you know you can't write literature without knowing how to spell.

[00:23:23]And that involves learning how to spell a lot of words. Um so sure you need to learn how to spell and I'm going to test your spelling quizzes and stuff like that. But very rapidly you're going to move on to classes in creative writing.

[00:23:36]How do you string together a sentence?

[00:23:37]How do you make a good argument? How do you write in a compelling fashion?

[00:23:41]There's more in common between that and software engineering than math today.

[00:23:45]You know, sure there are very mathy disciplines in software engineering. But like generally overall, how do you express yourself in logic has more to do with how do you express yourself in any other written medium. So >> yeah, I um it's funny like I have a history degree and was a self-taught programmer and there are some things right now with building software that feel very reminiscent of that in a way that they didn't like two years ago. Um it's a really weird like uh I I don't know. I I didn't I never expected that to be the case um to be honest. But um you know the thing that you're talking about here I I've commonly heard referred to as like taste right and like people in the AI world use this term taste to mean a lot of different things.

[00:24:32]It's kind of a catchall. Um you know how is your taste evolving and how do you think people can develop taste today coming into the industry?

[00:24:48]Um I don't know. I mean like the taste thing it's because it's such a like a broad term. Um trying to think of like sort of like which side to like pick at it. Um there's obviously there's taste in terms of like um what's going to be good? What is going to like what are I'm making something. Are people going to vibe with it?

[00:25:12]um u there's taste in the sort of trade-off sense of like I was saying like you know um we need to build a system and there's trade-offs in like cost and like which technologies am I going to use but then also how am I going to like compose the different Lego bricks of like the system together in order to make it robust in the face of failures and like the difference between junior and senior is like the senior has more wisdom accumulated experience that they're going to say like I've seen this work and I've seen this not work and here are the qualities of the things that tend to work and you can extend that into pretty much any domain. Um so that's what I think taste is and I don't think there's really a shortcut to acquiring that. Sure, some people maybe have good sort of they they come with good spidey senses from the factory. Um but um but for 99.9% of people that I've encountered, it's there's no substitute for just like hands-on time making stuff. And coming back to the AI thing, it's like on one hand it feels bad because it's like turning up the crank like once you get into industry and you're starting to like you're in a role, you're in a job and you're using AI. Part of what this moment in history feels bad about is this this idea that like we're all working in this factory and like somebody went and took the assembly line speed and they cranked it up to 11 like um and now everything's like flying fast like so fast and it it's it doesn't always feel good. Um, but on the flip side, it's like, uh, if I in my own career could have elected like here are the moments where I want to like turn it up from like five to seven and then I just gain experience that much faster because like things that previously projects I never would have done, side projects or whatever, just to like level myself up or or things that I would try at work, like the cost of trying things yesterday and the cost of trying things today wildly different. like the marginal cost of trying things has dropped to zero. And so I can use that to accelerate my own like accumulation of wisdom because now I can do things that previously is like man that's going to take weeks and like maybe I have better places to spend my time. But now it's just like man that's going to take hours and I can do three of them in parallel like on a weekend. So like why would I not be doing that to like level myself up? It's it's we live in the future and it feels like science fiction because it's like man my ability to like just try stuff is through the roof. So >> it's a little bit addictive.

[00:28:03]>> Yeah, it is. But like also burnout like now I'm also seeing this even among senior engineers. They're like I've parallelized myself into feeling bad.

[00:28:12]Like you know that's that's a problem.

[00:28:13]And I think like we're all sort of navigating that that feeling together as an industry right now.

[00:28:20]>> Yeah. Like I found myself um on a Saturday I like wake up and have my coffee and start in a side project and suddenly it's like 6 pm and I'm like what did I do all day?

[00:28:32]>> But like >> but like okay but like rewind a few years like before AI if that had happened you probably would have given yourself a f a high five. You could be like, "Oh man, I love it when like >> time falls away. I'm in this zone."

[00:28:48]Blink and like 12 hours go by on the clock. Like, you know, that's that was an awesome feeling.

[00:28:54]>> But now it's like almost kind of like tipped over into like because I'm not doing that in one thread. I'm doing that in like five threads or whatever. Like, you know, I've got like I've got like five clouds or five codeexes or whatever and they're all busy like humming away from me. And then like at the end of the day I'm like woof I feel like I had a work day on my Saturday like you know that doesn't feel good. Um so it's you know it's nuanced.

[00:29:18]>> Yeah it is. Um so when we're talking about kind of taste and the idea that folks who are more senior in many ways have like better pattern recognition or things that they can kind of reference back to.

[00:29:35]Um, you know, related to that, like one of the other things that I'm seeing is that the way people work and the way people work together in an organization are changing very rapidly. Um, and I'm curious like in in an organization as big as GitHub, like I know you're in this skunk works unit, but like what are some of the emergent behaviors you're seeing in the organization at large now that people can experiment and ship stuff so rapidly? um >> both in engineering and outside of engineering. I I suppose >> I that's a good question. I think it's to some degree there's there's some amount of um blurring the lines between those different functions of the business. Like what's the difference between a product manager and an engineer? Well, the product manager maybe understands more about how to talk with customers and how to um sort of um figure out and ddup requirements and and whatever. and the engineer knows more about code. But previously the product manager couldn't have tried things on their own and the engineer maybe couldn't have like done the sort of like ghetto product management themselves or it would have been like you know like more difficult and now it's just like we're all hybrids like where previously you know unicorns were out there people who are like designer developer like mixes and whatever um those hybrids um but now like everybody can be a hybrid everybody can can be a little bit of something else. And so I'm seeing now like, you know, even inside GitHub, like what are the profiles that we're looking to hire for? It's like, you know, it's not full stack makers. It's like we're talking about sort of like interdisciplinary hybrids. We don't want just like product managers who know how to write a product requirements document. We want them to be able to prototype. And the engineers, we want them to like feel comfortable slapping together something that like, okay, maybe it doesn't look the best. It looks a little generic because that's what the tools, the AI tools know how to bake, but like that same engineer isn't like blocked on having a couple of hours of a time from a designer to help them go from zero to one. they can do it themselves and through that like you know start to get their first bits of like like okay now I see how this works now I see how this feels I'm not just philosophizing I can put my hands on the thing touch it feel it like you know and get a sense for it like that's what everything is trending towards is like sort of like everybody being sort of um these crossf functional multi-capable kind of kind of people and that's what we're looking to hire here. Um, and so that's the guidance that I give to everybody is like, you know, once upon a time as a hybrid, you know, when I was coming out for roles, people be like, "Oh, you're a designer and a developer, like you know, are you the best developer?" And I'm like, "No." And then they're like, "Okay, so you must be like really like, you know, like a like like a top tier designer." I'm like, "Nope, not that either." And they're like, you know, because everybody knows this sentence like jack of all trades, but everybody forgets the back half of the idiom, which is and master of none. And so then they'd be like, okay, so you're not the best this and you're not the best this, why should we hire you? And uh what I learned over the course of my career is that the response was because I can glue those functions together for you, right? I can help your designers and your engineers like meet in the middle and and actually deliver good outcomes, business outcomes. Like we shipped the product, we shipped it on time. It was good. People vibed with it.

[00:33:18]Um uh and now that's coming for everybody. Like everybody can be that kind of hybrid. Everybody can be a jack of all trades and master of none. And where previously like over the course of my career that wasn't valued as much, people wanted to hire deep specialists.

[00:33:36]And there's still roles where like what you need is a deep specialist, but like nowadays the average role benefits more from a generalist that's able to sort of jump on everything. And so like I'm happy to see that happen. Uh you know, I wish it had happened earlier in my career, but like better better late than never, you know.

[00:33:56]>> Yeah, I I could not agree more. Um we're we're coming to the end of our time here. I have a couple of weird like tangential questions I want to end on here. Yeah, let's spicy. Spicy.

[00:34:06]>> Yeah, the spicy ones. So, the first one is what is the weirdest half-finish prototype that you have on your computer right now?

[00:34:14]>> Um I have a like a prototype. So, you know how like the context window in every like AI model is of a fixed size and I don't want to burn it on things. So, I find that I have conversations with AI that are like my conversations with other human. They tend to be very branchy.

[00:34:38]>> Like I'll like ask it something and then I'll be like, "What does that mean?" And then I'll end up like down a rabbit hole. And now I've just consumed like a whole bunch of my my context length going on this like conversational side quest and I don't even remember how to like branch up. Like how often in like normal human conversation do you like how did we get on this topic? I don't remember what we were talking about beforehand. Like so um I I have like like a prototype for like it's a different chat experience on top of like any one of the models >> um that like uses kind of like git branching semantics in the conversation and then like how do I construct because like the way these models work it looks to you like every time like you type in a chat message and you send it and like that but what's actually happening is that at every moment in time it's taking the entire chat history plus your new message and sending all of that as context to the model. Like the the models are stateless. They don't like retain things across terms of uh like sort of conversation. So like it represents the entire chat history is like like a like a graph and in order to like produce the state of the conversation at a given like tip of that graph like it could just walk that like graph all the way up to the root. But that way if I go on that side quest, that's its own sort of like token history as opposed to like my main thread of my conversation. Like I don't need to pollute it with that side quest of like what does that term mean? And then I'm like using up the thing because like when I hit the the context window maximum, it's like it now needs to compact or throw stuff away. So like I don't know that's something that I'm building for fun on the side for myself because I find that I want it and like I know it exists like when I use claude it's possible to like you know like branch conversation histories I'm pretty sure it exists in codeex as well. Um, but I just didn't like the interface for that and I wanted the sort of exercise of like how would I structure this like in the database in the product? How would I represent it from a UI perspective? So half-finish thing that I'm hacking on for fun.

[00:36:49]>> That's awesome. Um, so I think my final question here will be uh maybe a spicy one, maybe not, but like what's your most controversial AI hot take right now?

[00:37:00]>> Controversial AI hot take. I think it's it's the one that I said earlier is the models don't matter thing. Like right now, like, you know, you open up the internet and everybody's just like, you know, like loyalists to like this model or that one. It's just like I'm a Claude Stan. I'm a Codex stand. I'm a whatever stand. I'm a Gemini stand. Like, no, I like Quen. I like Deep Seek.

[00:37:21]>> Doesn't matter. Doesn't matter. Like, you know, it's it's like the models are electricity >> and like what I care about is my blender. Like my blender doesn't care what electricity it runs on. Like what I care about is like is the blender good?

[00:37:36]Does it have a nice interface? Does it look pretty on my kitchen counter? I don't know. Whatever. Whatever matters to me. Um uh but the underlying intelligence is a utility and it's relatively undifferiated. It's the experiences that matter on top and I believe in that very very strongly. You know, like in the early days of like electricity, you know, all the theaters were like, "Come see shows here. We have electric lights."

[00:38:05]>> Yep.

[00:38:05]>> And then three minutes later, everybody realized that like actually it's the show they were coming to see, not the lights. Like, so now we're finally crawling out of the stage of like everybody's just like, "Look at us. We have AI." And like, no, that's not the part that matters. It's the like, you know, the content that matters.

[00:38:23]>> Yeah. Excellent note to end on. Um, yeah, thank you Edon for the the super interesting conversation. I feel like we could go for another hour here. Um, I I hope everyone who's watching enjoyed.

[00:38:34]Um, definitely go check out githubnext.com.

[00:38:37]They are launching all sorts of crazy uh experimental projects. Um, and you know, I hope everyone enjoys the rest of Global Hack Week. We have another excellent AMA coming up after this with Chonin from Snowflake with Roseno, uh, our lovely host. But uh thank you all and uh happy hacking.

[00:38:56]>> Yeah, thank you. Thanks so much for having me.

[00:38:59]>> Thank you so much, John. Thank you so much, Edon. Let's give them all a warm round of applause, y'all. Love to see it. Remember to check out all Remember to check out all the good things that Don is working on at githubnext.com.

[00:39:11]Uh and for now, uh we'll say adu. Bye y'all.

[00:39:17]Okay, so we got to keep things rolling here. Thanks for thanks for sticking around. Thanks for joining once again.

[00:39:22]We've got a ton of great things happening at Global Hack Week. I do have the pleasure of announcing our next guest here. Uh Channon Natsenomat is a PhD and a lead developer advocate for Streamlit open source at Snowflake. Uh you may have heard of him through his uh other moniker or also known as the data professor. So if you've ever been um on YouTube learning about things um AI and data related, uh you may have uh seen Shannon there. Uh, so without any further ado, I'll go ahead and bring him up and let him talk a little bit more about his experience. Channon, welcome.

[00:39:57]Oh, looks like you're on mute there. All good.

[00:40:02]>> A a pleasure. A pleasure to be here.

[00:40:04]>> Yeah, absolutely. A pleasure to have you. Um, you know, we've got a lot of folks joining us from all over the globe today. We saw folks from India. We saw folks from Pakistan, Kenya, Sweden, the US as well. Um I'm sure a lot of the folks out in the audience are really curious about your personal uh personal experience and um you know your your journey in tech and your your career journey in general. Do you want to talk a little bit about that before we get started?

[00:40:31]>> Uh yeah sure. Um yeah so I'm currently working um as snowflake particularly on streamline open source as a developer advocate. So like creating content whether it's blog tutorials or um workshops and also YouTube as well. Uh also on the on the snowflake developers channel and also on the stream open source channel.

[00:40:53]>> Awesome stuff. So I'm going to go ahead and pop up your creds here and uh let's get started. You know I think uh a lot of folks uh in the audience would really be curious about your journey from academia to tech. So uh you spent 15 years in academia as a professor of bioinformatics. So what was the exact aha moment that made you decide to transition into tech and start creating content on YouTube?

[00:41:18]>> Right. Yeah. So um so while creating content on YouTube, video after video, I feel like the the field is moving fairly quick and there's still a lot of unknowns. So like the more videos I create on various topics in Python, data science, AI, I mean there there's still a big gap of knowledge that I feel is very interesting to pursue. Um like as a kid growing up, I like to tinker with, you know, new gadgets that that that are coming out. Um and I think the same kind of happened, you know, like playing with new Python libraries or new frameworks kind of brings joy. So one one aha moment was like I see a social media like on LinkedIn and ex Twitter um a lot of my friends on on YouTube or or also um content creators on on Twitter they kind of moved to developer advocacy and then I was also like tinkering with with that idea because we were at the time it was COVID like midcoavid and working from home and you know like It's kind of like the boundaries that once kind of separated um us from working at a different place. It's kind of blurred, you know, like typically we will go to a physical location, but then during co it's like everything is virtual. Um and and there there seem to be a lot of remote work happening at that time. So it's like the perfect mix, you know, like perfect storm kind of kind of scenario. Um so I saw that I already created a lot of content on Stormlet. So I I looked around and um see that there there's like like an opening initially applied for a uh a data row but then like after talking to uh some of the co-founders the CEO Adrian uh eventually kind of pivoted to more of a community role and and that's like the first time I heard of developer advocacy.

[00:43:14]>> Awesome. And uh you know as you transitioned from you know obviously academia and then going into developer ad advocacy and then creating your own data science data science content online what were some of the biggest challenges you faced during that during that time uh during that period. Yeah, I guess like the biggest challenge would be time, you know, like the time to create content on the side and also like for work and also kind of like the differentiation that you would have kind of like the boundaries at which um like if you're creating your own personal branding content, you know, like there there's a certain boundary that we should have and and also like a certain persona. Um, so yeah, I mean managing both of these um and also like finding the time necessary to to create the proper content for each of these um separate um personas.

[00:44:03]>> Gotcha. And you know, you mentioned this was during a time where you were, you know, everyone was sort of on lockdown there. There was a lot of like different things happening in the world where people probably had a lot of extra time on their hands. um how are you how are you like you know bridging that balance now like now that everything is um starting to transition back into uh uh a lot of inerson work and um you know I'm I'm not sure what your situation is at at snow at streamlet at snowflake but I would imagine that there's not as much free time or even not as much appetite for some of this uh digital content as there was previously uh could you talk a little bit about um what the differences between like you know when you started and now are >> right Yeah. So um so when I I started it was like kind of like middle of COVID and it kind of subsided and so what what what changed was that there are more like inerson events you know like attending pyon presenting there um and also like um there there's like the annual snowflake summit um spanning over several days. So yeah, there there's opportunity to to network with people and also to meet people that you already know like on on Twitter on on YouTube um but then kind of meet them in person, you know, like watching their videos or reading up on their social posts or their their Medium posts kind of and also having the opportunity to finally meet them and and even like um I think at Pyon last two years I met Eric Matthysse, the the the author of the Python crash course.

[00:45:35]>> Awesome. It's always great to meet these people in person. I have I've always dreamt of meeting the folks that I talk to at Global Hack Week in person at some point. So hopefully, you know, fingers crossed we all we all get to join up and grab some coffee together at some point.

[00:45:49]Um so pivoting back to um you know, your your journey from academia to tech. Um how did your how does your or how has your academic background really influenced the way that you teach some of the complex and technical con concepts to a broad audience?

[00:46:05]Right. Yeah. So, um I I like to think like um like my transition is not totally starting from from scratch. So, I think like there there are several transferable skills that I take from academia. Like first is obviously teaching. Um second is you know like dabbling, experimenting. It's kind of like research. Uh learning a new tool is very much like doing research. Um, and also like you know creating projects, you know, like having that the proposal similar to how you would do it in academia, you know, writing a research grant, but you don't actually have to write a research grant because the company already sponsors the project that you're working on. Um, which is less of a mental uh burden because in academia you have to get proper research funding to make things work. Um, but yeah, I mean in corporate there's a lot of um support there. Um, so yeah.

[00:46:59]>> So basically stay in school folks because it's just going to be a benefit for when you get when you get into the workforce. Um, you know, we kind of talked a little bit about this previously like you know how how you're kind of balancing things out um from your content creation and then like shifting over to a more corporate facing role. Um, so at one point you were managing a research lab, you were teaching and you were growing the data professor channel. Um, how did you, and we sort of spoke about this earlier, but this is a little bit like, you know, more honed in, how did you manage your time and what advice do you have for students trying to build their own personal brand while studying or working full-time, >> right? Yeah. So like when when I was when I started my so when I was in academia and there was an opportunity to to acquire a like a supercomputer in in our lab and that kind of required me to become an admin like the computer admin which at the time I had no prior experience in you know like Unix administration.

[00:48:02]Um so at the time I knew only R and a little bit of Python and so that kind of led us to the opportunity of having massive computer power you know like at your fingertips but then the thing is how do you actually use it? So yeah, I spent like several months or so, you know, like reading up on, you know, administering the the computer, installing, you know, all of the various um softwares that we use like for at at the time it was molecular docking and molecular dynamics, you know, like simulating how proteins move in um in time and space.

[00:48:38]And then um you know like using a lot of parallel computing uh running a lot lot bigger machine learning models, hyperparameter optimizations and all of the great stuff that typically before I would have you know run manually on several different computers and taking a USB drive to kind of combine it together. Um so yeah it kind of gave me the opportunity to learn uh more of that.

[00:49:05]>> Gotcha. I mean, I think I think this is a good time to sort of uh transit.

[00:49:10]You're talking about some of the tooling and some of the things that you're learning um at, you know, as part of your career journey. Um as and I know a lot of folks in the audience are probably really interested in um you know, some of the tools that you use now and some of the things that you're working on at Snowflake and at Streamlit. So, between Python, Streamlit and uh Snowflake Cortex, so the tech stack for building AI applications is always changing very quickly. Uh what do you think an early career developer should focus on mastering first if they want to build like you know like the next generation of uh generative AI applications?

[00:49:45]>> Right. Yeah. So um so I I think there there's several pieces here that we could unpack. Um first is are they coming from a computer science background? If so then they could you know go straight ahead to using the tools. If they're coming from other fields like biology for example from which I I also uh majored in uh for my undergrad then it's picking up the key concepts of coding. It doesn't have to go into super depth, you know, but but kind of like understanding about compi compilation, understanding the the con the vague concepts of variables and for loops and you know, all of the essential topics that are covered in in introductory Python books like like the one from Eric Matis. Um and so having that you know vague concepts or actually if you can actually implement it by hand using AI tools like cloud code or cortex code uh that snowflake also has for data ecosystem and also like chat GBT if if they're more of a generalist. Um so a any AI tool I I feel like you could start from any of these tools that you have access to whichever is most accessible to you will give you immense power. Um but then if you have more experience then definitely use you know like more like I think they call it the AI har the different AI harnesses like plateaued.

[00:51:11]Yeah.

[00:51:12]>> Gotcha. I mean, I feel like a lot of folks are sort of torn between using AI and uh while while they're still early on in career and still trying to learn things, using AI and becoming dependent upon it and then also like you know um rapidly prototyping and creating an MVP as quickly as possible.

[00:51:31]So you talk a lot about that like the power of building an MVP. So um >> do you think the idea of being a vibe coder who prototypes quickly to get feedback is uh more beneficial for folks at that stage especially for like a college student that's just building their first um data or AI application?

[00:51:48]Um do you think rapid prototyping is better than trying to build a perfect project from day one? I mean for for a beginner um rapid prototyping really helps them to to get that initial time to wow moment you know like if they're able to build something fairly quickly and it kind of makes them you know motivated to cons continue the path. As for having a perfect product, I I I feel that it's more suitable for um if you're if you're monetizing your products or if you're working at a larger company. Uh but then for solo builders who are just building out their portfolio or or even learning about the tools, rapid prototyping uh would be the way to go. And and actually speaking of which I I also had a course launched last year um in collaboration with Andrew Ang's um deep learning AI and and also it's already up live on on Corsera on rapid prototyping um with trimlets with trimlet apps.

[00:52:50]Yeah.

[00:52:52]>> Awesome. Um definitely check that out.

[00:52:54]Drop the link. I'll drop the link in Discord for y'all later. Um I do Just wait a few moments for Rosento to join us.

[00:53:09]>> So sorry about that. I clicked the wrong >> Awesome. Awesome. Yeah.

[00:53:14]>> Um yeah, so I think this is a good time to jump to a question >> from one of our audience members. It's it's sort of in line with the topic that we have here. So um now due to rapid development of coding agents, do we really need to know coding to write code?

[00:53:30]>> Yeah. So if if so I I like I actually have a talk upcoming talk at Pyon um next week and that's a great question.

[00:53:38]So as a generalist without knowing coding I feel that's vibe coding but then if you have the necessary CS concepts engineering concepts then it's more of a vibe engineering I would say or agentic coding that um Andrew Carpathy also just proposed recently. So yeah, I mean harnessing whatever you have already from engineering or CS and then mash it together with agentic tools makes you um allows you to do rapid development using um agent coding. But yeah, I mean by coding for if you're not actually checking your code. So that's more by coding.

[00:54:18]>> Yeah, totally. Uh I think uh the idea of being the human in the loop and being a subject matter expert about what it is that you're building is going to be really helpful for anyone that's out there using a lot of these AI tools. And um I think uh Idon had spoken about it in the previous uh segment of this as well like you know don't just sit there and like ship it over the wall. You've got to you've got to check check the code and make sure that it's actually doing what it's supposed to do. Um yeah.

[00:54:44]So as far as your background in tech, I know you know obviously um you come from a sort of like a biology background. It bridges like chemistry and computer science.

[00:54:56]>> So how do you see the intersection of AI and like and life sciences evolving over the next five years for new technologists?

[00:55:05]>> Right. Um so I mean in the past two years open fold uh from deep mind has pretty revolutionized the field of life sciences. The rule is protein structure prediction algorithm and so essentially protein folding is solved before we would need to either do experimental uh determination using X-ray crystalallography. Um but still that's like the golden still the golden approach. Um and also you know molecular docking because now with AI we could actually use it to solve you know like instead of doing like large experimentation to manually test each of the parameter to see whether it influences um the biological outcome because in biology there's like empirical testing where you control the experiment you control the factors that are influencing the outcome and let's say that you want to see whether water influences the growth of the plant. That's like one one variable, one factor. How about sunlight? How about the um acidity of soil and imagine you could do a hundred of these, right? But then what if you could, you know, simulate the plant growing? What if you could have each of the soil particle simulated in virtual reality? Um so these are some of the things that are now possible with AI that kind of helps you as well. And being a domain expertise uh or or subject matter expertise expert there allows you to to utilize AI to transfer that knowledge that you have on a specific domain and then create something new and you know you don't h you don't even have to know all of the deep principles but just providing your intent and then you're able to get that you know baseline prototype ready for you to proceed with experimentation.

[00:56:57]>> Yeah. So it just it seems like basically all we need is to um ensure that we're actually utilizing some of our skills and our knowledge that we have and like applying it generally to um whatever it is that we're trying to build. Um as far as and it sounds like someone's joining us.

[00:57:14]Nope. Uh we are good. So as far as um that is because I know we've talked a lot about your past and a lot about what you're currently working on and a lot of how you transitioned into this uh current current role that you're in.

[00:57:26]what's next for for you professionally and what's next for like you know the data professor and the content man uh the content development side of things for you?

[00:57:34]>> Yeah. Well, uh what's next is I'm ready to you know kind of level up in the AI engineering side. Um so as you might know I'm I'm a self-taught coder. Um so but then like before I actually had that aha or click moment of learning coding I used to struggle you know like learning about the key concepts for for years um as I was you know working in academia but then what clicked was the moment where I apply I finally apply what I'm learning to actually build something that solves actual research problem um instead of you know like solving the toy problems that are mentioned in the books but kind of look at the toy problems that they're solving and then try to figure out how that could be applied to real world problem. Solved that and then kind of rinse and repeat over and over and then it kind of clicked and then once I I learned about Python then I made a transition to learn a bit little bit about R got some depth there. Um and the opportunity to do that is from from some of my undergrad and grad students who are using R uh to build models and then kind of pivoted back to Python because you know you YouTube subscribers wanted more Python content. Um and so what's next is yeah leveling up in in AI engineering and trying to see how we could bring that back to bioinformatics.

[00:58:59]Actually for the data professor channel that I'm still managing for the past three weeks I've been doing live streams from Friday which is today and also Saturday sometimes Sunday. So over the weekend uh I would have a live stream on building bioinformatics project from scratch where I actually show you know like here's the data explaining the concept also if we encounter some errors I'll also try to debug that live and kind of you know talk about the thought process that's happening there something that's more more practical um versus a veil well well vetted video content that could maybe be finished in 15 minutes. And I do know that, you know, in this world, in this day and age, short form content is king. Um, and uh, the live stream might not be performing so well, but I think it might help others, even a few people, u, that would be cool, um, to, you know, kind of get started, whichever field they're coming from, whether they have a computer science or a biology background.

[01:00:05]>> Awesome. It looks like, uh, it looks like everything is like coming full circle. basically the content is being driven by the needs of the students and then also like you're kind of like um utilizing some of the some of the things you're you're you're learning about AI yourself uh to kind of espouse that knowledge. So, it's really cool to see that. Um I guess it brings us to our last point. I know we're coming in close on time here. Um if you were sitting in the audience as a college student today or just like as someone that's, you know, watching this stream and wondering, hey, what do I do next? How do I become a developer? How do I break into the industry? What's one skill or habit you would double down on for uh 2026?

[01:00:46]>> Wow, that's a great question. So, um I feel like the the best skill that I would double down on is not to overthink things and not and if possible, you know, kind of break free from the imposttor syndrome aspect of caring what other people think of you like and also kind of feeling that you might not have what it takes to, you know, proceed further.

[01:01:10]And I'm a I'm a testament to that because I'm I'm from a non-technical background, biology and you know working now in tech there's a lot of things that I don't know but the things that I do know I try to pull you know aspects from what I already learned from bioinformatics and work in that field and kind of pivot that to to what I'm doing now. And I think asdan mentioned in in the earlier uh session you know like having the skill at as a developer as a designer and kind of mashing it together. Same here. Skills from biology, from computer science, maybe a little bit of design. At 16, I created um some web websites, you know, from from scratch. Um but then with the help of a tool called Dreamweaver, >> uh pretty old already. I don't think it exists anymore. Um but then yeah, I mean it's like just start, just build and yeah, so that would be my my advice.

[01:02:03]Just build things and ship it and see how others perceive of it. provide you the the feedback necessary to repeat um the improvement process.

[01:02:12]>> Totally. Uh you know at MLH we're firm believers in learning, building and sharing. Uh you know get your hands dirty, get hands-on. Obviously every single experience you have whether it is a success or a failure is experience nonetheless. And um you know I think that's uh I think that's great advice from from your end. Um we are we are at time here. Uh I think uh you know everyone here has had a lovely time. I apologize for anyone whose questions did not get answered. Please reach out to us at Discord. We will try to get those questions answered for you there. Um I guess do you have any parting words of wisdom that you'd like to like to share with the audience before we go?

[01:02:51]Uh yeah, perhaps one one tip um is that you know like agent skills are become a very big thing nowadays and as a one person you know solarreneur managing YouTube on the sideline I would recommend embra embracing more on agent skills to kind of capitalize on whatever unique skill set that you have put that into writing convert that into a skill then you'll have like superpower to automate a lot of the task without you having to actually u implement them.

[01:03:19]Yeah, leverage skills and the best way to learn is to do. That's the uh the slogan I always end my videos with.

[01:03:27]>> Awesome. Well, uh you heard it here first, folks. Uh the the best way to learn is to do. Um Channon Atenat, PhD, lead developer advocate at for streamlit open source at Snowflake. Uh really great to have you here. Thanks so much for sharing your time and thank you all uh for joining in and watching us. We have a lot of new content that is coming up next. Uh we've got some folks from GitHub that are going to be teaching you how to use GitHub Copilot CLI. So make sure you keep it locked in for that. Uh Shannon, once again, thank you so much for joining us. And uh for everyone at home, keep it locked in. Keep stay tuned in because Global Hack Week is just starting. All right, y'all. Have a good one.