# Leading Data Teams In The Age Of AI

## Метаданные

- **Канал:** Seattle Data Guy
- **YouTube:** https://www.youtube.com/watch?v=UVHOl0vE7dc
- **Дата:** 30.04.2026
- **Длительность:** 58:36
- **Просмотры:** 912
- **Источник:** https://ekstraktznaniy.ru/video/49824

## Описание

I've been having dozens of conversations with data leaders asking them about whats changing in the data world.

And of course, what is staying the same.

This week I'll be talking to Heqing Huang.

He's been leading data teams now for several years and I wanted to know what he is seeing.

How is he using LLMs in his daily workflows?

What still hasn't changed?

Come join me this week and bring questions.

## Транскрипт

### Segment 1 (00:00 - 05:00) []

All right, and we are live. Hey there everyone, welcome to another Seattle Data Guy live. Uh today we're going to be talking about so many different things, data teams, AI, kind of what's going on. Uh with that, with me today I have Huching. How are you doing there? Doing great. How are you guys? Uh I'm doing great. Uh for the folks who don't know who you are, uh would love to get a quick version of who you are, what do you do at Scale AI, and how do you kind of end up there? Cool. Hi everybody. I'm Huching. I am director of analytics at Scale AI. We're a uh sort of data engine that power everything that's happening AI from autonomous vehicle to all the chat app that you use today. Uh been here around 5 years uh doing everything analytics across all the departments across um you know, the full life cycle of analytics. Yeah, excited to be here. Yeah. Well, would love to know kind of how you got there. Like what was your journey or your path into data? Um I studied industrial engineering uh in college. All my internships are in the Midwest in the factories trying to collect data, figuring out how to improve the manufacturing uh productivities. Uh end of the day I realized the thing that I care about the most is the people and the data part. Uh so after my college I went to analytic consulting in the Midwest. And then later on got my master in data from Berkeley and then worked for a company called Lenovo like top company for a bit in the pandemic. And then um through a connection uh ended up in Scale AI about 5 years ago and been here since. Yeah, yeah. Since you brought up like connection, is that how you feel like a lot of people need to get jobs these days? More just a side note, but Yeah, um I think back then um it was hard. Right now it's definitely harder for for new college students. One thing um I usually say is you just have to show people what you can do, but not uh just purely based on resume and school. That usually means uh a portfolio on you know, Tableau of the world. It could mean your GitHub commit history for a project you're passionate about. Uh you know, like you just have to build out the reputation so that you know how to let people read the resume, they can just see how you work so that you know, you can get a job easier. Yeah, yeah. I do not envy uh people who are starting out. It feels so much it's just so different and more challenging. So Well, you've you know, you've kind of grown uh at Scale AI you know, up through the ranks um you know, being working more uh with the data, managing, now directing. What what have you kind of learned in terms of like ratios of like teams in terms of how many analytics engineers or data engineers to analysts and data scientists do you find are good at Scale's size? Um that's a great question cuz the answer changes a lot this year as well. Uh previously I would say um it depends on how many business, how many product you're running and then ideally for each function or product you have some analyst supporting it. Um so there are two aspects. One is uh depends on the function you obviously want some dedicated analyst uh to help with specific area. Um it really just depends on how much of your data or company's data driven. The second piece is I think more and more everybody is becoming more full stack analyst in a sense, right? People write models, people talk to business, people do dashboard, people do inside reports, and people set up experiments if they have the infrastructure to set up experiments. I think the more people can who can do that, the more leverage a data team that can have. I remember you have a blog about, you know, my uh like the importance of data team uh in the eyes of a business. I think the more you can be full stack, you really understand what it take to make the business successful. Um that is also what makes a data person successful. Um so the ratio is all about, you know, how much leverage you can drive in the business. Yeah. Yeah, that makes That also makes me think of uh I was reading I think it was just a tweet, so you know, take that with a grain of salt. Someone referenced the fact that you're like, you know, 5 years ago a product uh product manager or product person would be managing one product and now they're managing three. Is that something you're feeling the same thing with analysts? Either they're having to cover three different products or maybe at the very least like you said like having to go more horizontal and just like you know, be the full thing and do it all. Yeah, I really truly believe um you had to match what the business can offer plus personal interest. I mean, our in Scale's case everybody always own more than one area, right? You can have somebody who own finance, who understand all the financial metrics, but also own the metrics on go-to-market, who understand pipeline AR of the world. I think there are a lot of synergies by just by owning the two area, right? For example, the customer entity might be the same entity across the two parts, right? You can just own that. Uh the finance people are not talking to sales people or their pipeline, they're not using same definition. You can the data person go in and bridge the gap and say

### Segment 2 (05:00 - 10:00) [5:00]

we should kind of think the same way. Uh even better, we start to see more of a blend of an analyst and system person. Imagine you're doing finance analytics, uh go-to-market analytics, but you own some part of the salesforce as a CRM in the first place. You actually can have more control over, you know, this is what data should be in the first place um as a interface to enter data. I think with all the tools going on the world, everybody depend, you know, on the individual appetite, you actually can own a lot more in the full life cycle of things versus, you know, I build a pipeline, I build a model, I build a dashboard, I run the weekly meeting. It can always be the same person doing everything all together. And it's actually better that way. Yeah, um all the this is just something aside from our questions. I was thinking about this the other day where, you know, every time we get through some sort of technical revolution, the goal in theory should be to reduce the amount of work we have, but I feel like we've always done the opposite. Like, oh, we have electricity now. Great, now we can work more because we can, you know, night does not define when we're done working. Then eventually obviously laptops, no longer location doesn't define where you're working. And now it's like expertise doesn't define where you're working anymore either. I was like, look, you could get you could become an expert I mean, not expert. You could have AI help you kind of be more or better or fill in some of those gaps that you might have. Uh and so it's like I'm like, I think we're just going to work more. I don't think it's going to be less. Yeah. I really think good way though is like everything is in your control, right? You can figure out anything um as long as I have the right role in the product I'm using. Um I think it's a good thing for a lot of our analysts where we can see everybody's getting more and more curious about, you know, my stakeholder used to do this. I'm just bridging the gap now. Uh you know, doing more execution side, which is pretty awesome to see. Yeah. [snorts] Yeah, yeah. I think that's interesting. Cuz I remember seeing like some of the first of this was um probably when I first came in the industry. Uh like I remember working with a financial analyst or I had a friend that was a financial analyst at Amazon. um at like a more traditional company. And the friend at Amazon was like having to learn SQL, having to like really unblock themselves constantly. And the friend that worked at the more traditional company was like, oh, I just use spreadsheets and wait for the data person to get me my data and then I do the thing. So it's kind of maybe that next step of like what expectations are especially at certain companies. Yeah, yeah. I think that's definitely the truth. We see that business person with all the new tool they start to write more tech ask more technical question do more higher level self-serve. Before that was like dashboard, right? Right now is people are doing much more sophisticated stuff with data now. So everybody's doing a little bit more. Yeah, somehow we have more work. Hopefully that translates to more GDP, you know. I think that's that is the goal of a lot of us. Or maybe not us, but maybe the people above us. We're like, we don't get to see all of that necessarily. Maybe some of us, but uh you know, definitely the tax base or the larger organizations. Um in terms of like what people take on, how do you decide that is in terms of like what your team takes on versus what they push back on? Um like what do you have an intake process? Is it kind of individual people that pick things? Um I think one of the really good thing about Scale and that we're lucky that way is from the beginning uh our leadership set up a horizontal structure where everybody have the uh the flexibility to do everything. Uh in a lot of organizations, you know, finance have a data team, marketing etc. We have one data team that can do everything. So as a result, it's super flexible if one area suddenly become really urgent uh versus or some people have more interest in some other areas. Uh so how we work is really like a very flexible matching program with no process, right? You somebody say, I read this book on finance, that's super interesting. Can I do more finance analytics? And we look at what does our finance team require uh so that make their job to the next level. Then we help with them to do that work. Um often time we have more work than we can absorb cuz you know, since we're horizontal, it's sometimes just about sitting down with everybody, hey, what do you have in mind that you need help on? Educate everybody why this is important. Usually it result in itself. We never had to be the person to say no per se. They will say, oh, obviously this thing is more important. Please help them. I think Scale has been operating really good that way. Uh so we're able to match what people want to do um and the business domain. Another thing that's really helpful is we don't have a way certain way of how to help people. There are people who are There are some analysts are really good at, you know, building models. They just want to build models, do some special automation, they're done. It's fine, right? The models are sophisticated, etc. Some people really like building dashboard, right? Building dashboard, etc. But the really important thing is does it add value to the business? That is individual's taste to decide

### Segment 3 (10:00 - 15:00) [10:00]

what actually needs to be built to serve as data function. There's no There is no guideline on you should do ABCD to do the job. There's like decide what's important for you to execute on and do that. So it's everybody have a lot of room to be the data person they want to be. Yeah. Yeah, I think you know one of the questions I get a lot from people is on that like impact and you reference taste. Like how have you seen people find that and are there maybe different forms of impact that you can categorize? I know that's not one of the questions but I just Yeah. Something I'm thinking suddenly. As much as we can, everything come down to the profitability of the company, the pipe the future revenue pipeline efficiency of the company. All right. I think is a good localize. For example, let's AB test this button location on the page for CTA. I think it's a good thing to measure. But for startup, startup like us, we're usually able to tie some outcomes to the whole length of things. The second thing is I think reputation as a data person is super important. All right, like how people feel about working with you as a partner in the solving problem when you know our space is you know AI data moving so fast, how much leverage you can give people to make it work is super important. So outcome plus reputation or feedback is how we measure the success. And ROI is I think should be the same as ROI of you know how the company is performing. If the company is not performing well, doesn't matter how great your model is, doesn't matter. All right. I think that's what it is. And then it takes a lot of personal conviction and skin in the game to have that drive to make an impact. I think it's which I find in the line of work with motivation. Yeah. Yeah. No, I think it makes a lot of sense. Try and find the things that energize people and not just you know assign them something and be like figure it out. Yeah. I think you you've been posting I think on LinkedIn. So some of the things that you've talked about recently one you referenced the fact that your team recently kind of handled I think it was like close to like over a thousandish questions kind of on the system that you built. I think you've got Snowflake DBT Tableau or something and so you've kind of implemented I think a combination of like AI and agents to help out. Walk me through kind of like what that looks like and like how that's maybe does that impact your head count? Are you able to just do more and that's the benefit? Like what does that all look like and change and how does that change the day-to-day? Yeah. Um Our stack is DBT Snowflake. We have Tableau but the usage has been way down to where is wasting money at this point. And then we have you know Fivetran the world to make it work. For agent, we use Text SQL. We have some internal tool. We have some iPad tool like Workato for automation stuff. At the end of the day, these are just tools and people again have the flexibility of whatever they want to use. Some people still prefer to use Tableau because you know they have really good row level security based on the user email to surface departmental data, right? There are use cases like that. But what it looks like right now is oh one thing to mention is Scale has always had a really good foundation on our code base. Every single business logic is on DBT. All the everything is pretty well documented. The schema is very clean and we do a good job of constant our data engineering team do a great job of constantly archiving all models, make sure they're up to date. As a result, like when you deploy AI model you will have to do almost no work of describing the model because the logic is in code themself. Like you don't have to write down the definition of revenue if AI can go into look at your code, look at lineage of the model, decide you know this is the application revenue from the base model seven layer down to the final reporting model. You need to use the right zero English about it. So what it means is once we deploy like Text SQL always is the our preferred agent, business can just ask questions with very limited context and perform super well because nothing is it doesn't explore a database. We don't have to because the code is the description of all the of the business logics. Yeah. That's how it works now. Kind of nuts. — [snorts] — Yeah, I know I think I found something similar with with DBT especially if it's well documented. It like when people talk about semantics layer semantic layers often it's like this extra layer on top of everything you've already built and I'm like but it should be there, right? If you've got your metrics if you've built everything in theory somewhere in there you you've already done this. So yeah, that's something that I think I'm always kind of poking at and wondering kind of how different people are doing it. Yeah. All I will say is like startups were probably a little bit easier than to handle this. We're only been in existence for 10 years. They're massive company with multiple Snowflake data brick instances. They definitely need some of that layer like ontology of the world. So we're just in the lucky position we don't have to go through the pain cuz

### Segment 4 (15:00 - 20:00) [15:00]

everything is relatively already clean. Oh. Oh yeah. I mean as soon as you go from you know startup or company that was tech focused or data focused and go into enterprise, it just gets messy very quickly. Everyone has every tool. You know it's not like you know you go into I don't I don't know well enough each company but you know pick a Fortune 500 company. It's like oh do you use Snowflake? It's like or it's more like do you use Snowflake? Do you use Databricks? Do you use BigQuery? The answer is yes. You use all of them. Every team has a different tool and you know the chaos that ensues from that. Everyone's political or everything's political and you it's hard to get anyone to decide which path to take. Yeah. I agree. Yeah. Another thing you kind of referenced in some of your posts is like you said like the era of like analysts being ticket tickers is kind of over. So what do you feel like is kind of the job description of analysts today? And maybe you kind of covered this in terms of like them taking bigger bets. Is it kind of this more entrepreneurial like I think there's like some value here and I want to dig into that data and help maybe drive the business case for it or what do you see? Um my observation is let's say two years ago I feel like at least half the call you have as business is talking about oh this is how the reporting changed. This is the new chart. How's here's how to read the charts the definition. I feel like we spend probably 5% of the meeting time talking about that now. Like we just kind of take it as a for granted. Everything you want will be done in 20 30 seconds and we talk about what needs to happen, right? If we're talking about user strategy, here's how we retain user better. Here's my strategy and then back test it with my hypothesis cuz I can't do it instantly now. I feel like all the time are spent doing that and more strategic things. Or what I'm interesting in is people will spend more time not looking at the number but looking at the product themself. You know oh if you're a user, go through this journey to understand what you use actually go through and not describe it by data so that you have more realistic experience on what the data tells you versus to purely listen to data get a lot of insights. So in the sense you get higher level and very ground level at the same time to be more what say Gen Z say touch grass of the business and describe it with your data. I I think that's an interesting example there because it's like yeah like as a data person that's always kind of the challenge. I when I talk to a lot of like leaders or and business leaders or data team leaders, the common especially like the data engineering side is like they're so removed from the business, right? Like they like because you're like focusing on the minutia of like your deployments or like oh how my docker container set up and like it has nothing to do with the actual business in the sense of like well how do we drive dollars? It's like how do I optimize a query which has some impact for the business but it it's not necessarily as directly tangible. So yeah, touching grass for the business or whatever the equivalent would be there. Hopefully I use that term right. I'm not a Gen Z but I heard people say that a lot. I just want to be cool. Yeah. Every so often I have a few friends that I run to that are in that Gen Z age and unironically they'll use a heavy amount of Gen Z slang and I'm like oh I I've got to go get a dictionary. — Cuz I always thought it was more of a meme that this is how you know I obviously millennials have our own way of speaking as well but I feel like it's not as heavy. There's not as much internet slang in it. So yeah, it's always interesting to hear. But I think touch grass is accurate. Okay. Good. Um overall. And I was I was I'm trying to find I had another question in my head and now it's sadly left. So I will have to find it again at some point but I was really hoping to ask it. I guess on maybe the semantic layer and I think you kind of referenced this with like the code and like you you've said that data teams kind of own the semantic layer now. Like one, have they not always kind of owned it to a degree? Like I or at least they're the ones who has have to implement it. Or what do you mean by that? Um I think there are different like if you think about what is semantic layer here to do, it is here to show all ends you know the context between after behind each columns. A lot of context already exists in the code. If you're talking about let's say finances have a different way of reporting extrapolating revenues, everything ideally is already in the code. Like you don't have to explain the way I calculate ARR is take some month revenue extrapolate to 12 month. Like you don't have to in theory you don't have to say any of that if the definition is written down in your TBT or code base. But there are a lot of cases where things where um

### Segment 5 (20:00 - 25:00) [20:00]

are a lot of things where um you really don't have, for example, this specific contract 555 is a unique case because, right, we have a special handshake deal with contract 554 that they need to share the consumption, right? I think that thing nobody will ever know because they don't exist physically in the world and it only exists in neurons of your humans. Uh in that case, I think we usually work is uh we have to make it easy for business stakeholder to input that data and let our AI remember it. Um there are different ways, right? Like people use Notion, people use Google Doc, people write it down especially anything that has a can retain a memory it is helpful. Um Um but is but if you if the data team doesn't set it up correctly, you know, I can write it down in my Slack, I can also Google Drive, I can write it down in my Confluence. Then the knowledge become everywhere, it's become useless. You know what I mean? So, I think data team now owns the layer of managing it so that it how it kind of herd of cats per se, put all the things together, make sure it compounds, make sure there's not self-inflicting um to work. Yeah. Yes, which just reminded me of my next question or at least maybe more on like what you theorize will happen in the future. I think one of the things that I've noticed um over time is that as we unlock some of this data like make other parts of like getting data easier that just increases the type of data or like how much we want. Like we want that data you just referenced. in there but we don't like we want a way that we can pull it in um and actually bring it and answer these problems more accurately and cover some of these like edge cases um that's often in some sort of unstructured data. It's in a contract somewhere but nowhere else. It's in you know, where wherever you know, the the person who the sales person who did it didn't put it down in Salesforce, right? Like Yeah, you you've got to somehow capture it. Is that what you kind of see is like maybe the reason I think we'll still be very busy as data people is because of that. Like we want even more now. It's like great, you have free time. Get us all the rest of this data that we've never been able to get. Is that kind of what you imagine? Um it is what we imagine. It is already what we're doing. Uh two examples of what we're doing. One is um everybody have their favorite way of recording a meeting. Uh like Cronulla, uh Notion and one other they're like thousand to Clearly my favorite. I don't use them. If you don't know the social — Cluey? Clearly. Roy is that a say Do you do you actually use the product or is it just for the memes? — I don't use the product. I just — Okay, so just for the memes. — I love their marketing. Um but you know, Well, okay, but you can't love their marketing. You got to like their product, you know? You got to I I think that's always the challenge. Anyways, you going. I'm sorry. — That is true. I can probably use it on my local machine um on my personal machine. Um but imagine all the customer call uh or all the things you want to Salesforce to remember already exist somewhere in some other form. And what is the best way to transform that into structured data? Usually not numbers. Usually they're text blob of important information passed on to Salesforce. We're building that exact pipeline with human input. After a call, you know, transcribe it, parse it down to specific format with human in the loop review. After review, you put it down put it back to Salesforce so that memory retains. And also you can do more data science stuff where based on the memory, based on the sentiment of the deal, you have more better prediction about, you know, pipeline prediction of your quarter, right? So, unlock a bunch of stuff. Another case is you previously mentioned like oh, we have a lot of complicated PDF not in a standard format because, you know, AI moves so fast. What is the best way to parse out all the data with also human in the loop to validate that right back to our production database. I have a whole team also working that. I think a lot of the the gap between the data person and system person become less and less. It's becoming more into one now, right? Like we used only reporting. Now we can own the data, the generation of the data. Now we own how we collect the data, right? The whole thing become wider and wider. Um which is kind of cool thing that's happening. Yeah, yeah. No, I the more we can like I think as I think I read it I think it was like I think a few people have talked about it but I think I keep seeing it on I think it's Buko Capital or something on Twitter where they he references uh the collaboration tax kind of getting compressed. Where it's like yeah, if you don't have to have all this like let me go, you know, I'm the analyst, let me talk to the Salesforce person, let me go talk to the business person on the other side, right? Like that is really where a lot of stuff one gets lost and um just work gets slowed down because it's like oh, I got to wait for, you know, Jamie to finish something before I can do my part anyways, so, you know, but if you don't need that, if you can just get instant context somehow or already you are the context, uh you know, it really simplifies that whole process. Um Cool. Uh so, I think some your team has or is managing somewhere

### Segment 6 (25:00 - 30:00) [25:00]

in the I think trillions of rows, right? Like and having to manage all these queries how do you one how do you handle all the cases where this is wrong and then how often does it happen? And do you have some sort of QA loop or something to try to measure trust? Um Yeah, my trick I don't have a I don't have I'm still looking for a deterministic way um but the non-deterministic way is whatever we ask AI to produce a number um on a regular basis, we always have a checkpoint. Um usually a historical data point. For example, uh if I know my my my uh my January revenue which is booked, fixed, never going to change, I ask always ask AI reconcile this number first um before you tell me the current revenue for this week. And then usually we do one line differ, tell me about the revenue for that department on this condition, right? You have more complication so that it always know uh the number is correct. So, often time we find out we'll find out is even if you give them the exact query uh they can still copy and paste wrong per se, right? It's just the prediction model is like moving the embedding number from one place to the other. It can still make mistakes. But so, the only really good way is check the number. The second way is for like a really high impact work, we try to aggregate the model um to a point where it's just a fix per se so that agent never need to write a single join statement. Uh I think that's a goal, right? If you absolutely oh, never need to write a join, that's and I have everything I need, then you kind of know the end number going to be accurate cuz it really is just a select with very simple aggregations with almost no filters, right? Yeah. So, I think that's the two way. And then trust is interesting because I don't know how you feel. I feel like 6 months ago everybody be saying um oh, this model the never answer is wrong, the model is bad. All right? Now everybody saying oh, the number is wrong, and my prompt is wrong. Right? Nobody blame the model anymore. Everybody start blaming themself. So, there's a happening also like it's becoming easier easier. People know oh, if I give them the correct instruction, it will do it. I've seen it many times before. Maybe I need to work on my instruction. You don't need to show that anymore. People just know. Um It's a collectively changed. Yeah. I I feel like um that kind of aligns with when Claude became popular, right? Uh I think once Claude became popular, people suddenly were like oh, this is actually pretty cuz they focus so heavily I think on code and things like that. It's like this is actually pretty good. Um and so, yeah, I think people started to figure that out more. Which also just speaking of I think the way someone described this recently was like um how you're on autopilot like when you're driving from point A to point B and you don't remember how you got here. That's how it feels like with Claude where I'm like I remember there us talking about opening AI and Mistral. And then something happened about a year ago. Yeah. And now we're not talking about Mistral. I mean, some people still are but less. And now it's like OpenAI, you know, Google and and Claude and obviously, you know, XAI and a few others. But like suddenly I'm like how Yeah, when did this happen? Like did someone went back in time and like shoved that tool in there and we just all didn't realize. It just was all over on like reels. They did They must have done a really good job on like — Yeah. So much influencer marketing. So, yeah. Yeah, did you see that the whole Anthropic has a one person run growth marketing? — Yes, yeah, yeah. Also another crazy part. Just think about how much traditionally how many people does it take? One person with the right tool, you know, all the roles are being compressed. Yes, yeah. I think a lot of lessons from like Cursor from the beginning, right? They can you can pick your own model so the answer is wrong like oh, you know, I picked the wrong model. Yeah, yeah. I feel like a lot of people are same thing, right? Like oh, yeah, it's not a it's not my it's not the model's fault. It's, you know, probably my fault. I didn't pick the right model. Yeah, I can do better. I can make it better. Yeah, it's interesting. Speaking of which, you know, all this around hype or all this around models and it's hard not to obviously feel like there's a line somewhere where hype is maybe not everything. Um uh So, I guess in in your experience like you you've said that, you know, reading about AI all day um it's one of is one of the slower ways to keep up. Uh the you know, probably the best way is, you know, just doing it, building something, practicing it. Um what do you people kind of or what do you recommend uh people build in terms of like some sort of minimal viable thing uh that gets them started? Um anything. Um I think that the intent of the post is I realize people always talk about oh, this new feature. Then I was like have you tried it? Oh, no, I haven't tried it. Like what does it means? Oh, I don't know. It's just it's a cool news, right? But you try that they're like they suddenly have this

### Segment 7 (30:00 - 35:00) [30:00]

understanding about what the tool does even just being things. Uh, for example, my team we play a map guessing game every day. You basically it's like a Geoguessr game where you know, you point a location based on distance you have a score uh, kind of thing. It's like a pretty cool online application that my team member um, uh, he's trying to basically learn Claude code and in the CLI. He's never done it before. He just write code as a whole thing. And he suddenly understand how it works. Oh, here's how you know, CLI works. Here's how hosting works. I use Netlify. Oh, I should use Superbase to lock the results. He suddenly understand what it mean to write code an app that published to a website, deploy on Vercel on whatnot. So, people understand it. So, not because he tried this AI tool, but he understand how the whole thing works now. as application. What does it mean? Uh, you know, and how to design a UI etc. That's what I mean by you know, you can read this all day, but doesn't you don't grok it as much as uh, you use the tool themself. Uh, so if you want to build anything, it's really just anything you want to do. You do play a video game, let's try to build that video game, right? Do you have a little helper, build a little helper. I think he started it when he used Obsidian a lot. Um, he just want a tool a shortcut on the keyboard to write a note directly Obsidian. So, he write code that on a local machine. He's like oh, that's a great application. He understand like how Mac OS works and everything else that come with it. Um, so that's right the intense of why you should people just should try for 5 minutes and suddenly is worth probably reading 5 hours of content that you not never going to get. Um, That's — [clears throat] — good. Sorry. What did you say? It's weird. It's just got to try it. I don't know how to explain it. You just have to try it so that it just your mind get opened up in a way. No, I think I think that's a great way of putting it. I think that's that it's always been the case, right? Like case in the days of programming, right? Like it's like look, you can watch a million YouTube tutorials, right? Like I think I saw someone recently put this out. They were like oh, something I I've recently discovered this concept of like AI uh, AI tutorial hell or something like that. I'm like it's not a new concept. It's been there forever. Like yes, welcome to tutorial hell. It's not AI specific. It's anything. Like yes, it's very easy to get stuck in the let me start the basics 101. Then you get somewhere, you get stuck. You go do something else. You're like okay, I'm going to try it again. Then you start the basics 101 and then you get stuck and then you go do something else. And or you just keep doing every tutorial and you never go beyond it because it's like well, I need someone to tell me exactly what the step one, step two, step three is. Yeah. Yeah. Yeah, nobody we don't have to Google that anymore. The Claude code will tell you here's my plan. Can I do it now? So, yeah, please do it. Explain to Claude to Yeah, yeah, yeah. Uh, what do you mean Claude it's going to take 5 hours to do? You're going to do it in the next 30 seconds when I tell you to. So, don't tell me it's going to take me 5 hours. I had a this experiment in March. Um, it's a city thing I saw on X where like how do how do you measure your ability to write code is measured by uh, how long your code can run without your input. I was trying to maximize that. So, I had a session where to a point where a GitHub banned my account because I think I'm a spam. I thought the email I stick out like a whole week to unblock my own GitHub account cuz I have too many comments like PR just non-stop around the clock. Um, So funny. So, don't don't do it. But uh, it's doable if you really want to. I'm glad you got it unbanned. Um, I someone asked this question in the comments below. I think this is very similar. So, you're you're clearly I think you're clearly bullish on AI. But where do you think kind of the narrative is overhyped and where do you think it's kind of maybe underhyped or where it fits, you know, what where is that all kind of for you? Um, the thing is I'm very afraid to say anything that's overhyped because like the exponential growth of all the intelligence capabilities is beyond my brain can process. So, I feel like even if I say so, this will happen in 3 years it's going to be like 3 months. So, it doesn't matter what you say, the exponential growth is not comprehensible. Um, I don't think anything is overhyped. Maybe um, talking about AI is overhyped. Doing AI is underhyped. Yeah, yeah, yeah. Um, and also it's just like I think what also underhyped it is um, like your own judgment is probably undervalued. Um, right? Cuz like the Claude is already so smart, but if you try any real project, you realize there are five different way of achieving what you want to achieve. And then often time is your own taste of what to do. Um, that really matters. So, every time Claude tells me oh, that's a good idea. Then I it's like a reminder to me oh, why didn't you think about it in the first place, right? It's like oh, yeah, maybe I should have put it in the first direction. Like I maybe I should have thought it through uh, in the very beginning so that you know, I can save another 2 hour of write coding and just go directly to that route. Um

### Segment 8 (35:00 - 40:00) [35:00]

Yeah. How do you recommend people kind of build taste, right? Cuz I feel like taste and like your sense for what is good and what is bad is often built through experience, right? Like you when I first started learning how to code, um, right? You learn at school or maybe you learned on your own uh, and you think you're good. Or maybe you don't think you're good, but you're like well, I'm not terrible. Uh, then you come across your first really good project and you're like I'm Like I am not good. Like I it's funny cuz like because you've done bad and then you look at someone else's you're like no, that's good. What I did was bad. — And now I've kind of elevated my taste because I am now experiencing something better. Um, so how do people kind of I guess make that distinction now if they haven't had to go through that phase where they've even thought about where they made bad and now they see good. Yeah. Um, I think it's a very difficult question. Um, I think just being generally curious uh, as a long-term thing that is think of reading books uh, specifically like actual books or Kindle. I think it's really helpful. Um, I feel like a lot of for for work for business wise, uh, I do find that it's helpful to read like Harvard Business Review the case that people read in MBA classes. They should be like very concise case uh, where you can understand the topics. Uh, and but most importantly if you're write coding like a software project, the only matter thing matter is if you if would you use it yourself constantly or something you try to experiment, right? If you build something it take you a week of time that you don't want to use it. I'd rather use ABCD for it. Like obviously it's not good. It's more of like uh, refinement of not don't lie to yourself kind of framework. It's like oh, I really don't want to use it then the taste when I started project of designing is not as good as other product. Now, let me learn about like why do I use that product a little bit better. So, you kind of develop a muscle memory over time. Yeah. Yeah, I feel like I'd rather use Codex or Claude or Gemini. Like maybe it wasn't good. Maybe it was bad. Maybe whatever if you're trying to build like your own AI analyst or your own coding assistant on top of some of these models. Um, I've run into that with a few tools as well where I'm like this tool supposed to have AI. It's supposed to help me. I I get better results from just the raw model than uh, whatever else they've tried to put into this and then in service me and then put a margin on top of their tokens. Yeah, that's totally true. Um, oh, this guy want to say hi. Um, yeah, I think taste is developed by — dog? This is Mando. — Aw. Wow. It's too cute. Um, yeah, taste is come from keep doing things. Uh, would I use it? I definitely have started more projects that I would admit that I just totally gave up midway cuz I feel like oh, this is a terrible idea. But you kind of really learn a lot from these experiences. Um, it's also why I feel like it's really too early for people to ask what is the ROI of all these things. Yeah. People don't know what it what is this yet. Um, I think it we should spend responsibly uh, but not too anchor down like ROI or what have you achieved? How many how much money have your Claude bot make? I was like oh, this is that's another point. The point is the lobster is alive and then I learn a lot from being you know, making it alive or keeping it alive, you know? Yeah, yeah, yeah. How much did uh, Gary pay you for that? Uh, — No. Okay, we tank. Oh, his G stack is pretty good. Uh, I like it. How much did he pay you for that? — Oh, have you used it? It's uh, it's pretty small. No, I have not. No. Oh, if you don't know, G stack is this Claude skill. The funny thing is um, some point when you're using the skill, it literally on your Claude code will say uh, something like oh, you know, building product is valuable blah blah. If you're passionate being product, you should consider apply to YC. So, what I've what I've basically heard so far and I haven't used it so I can only say what I've heard and you're sort of confirming it is that it's somewhat helpful, but it's also some marketing. Like it's really decent amount of marketing. Like like you know, we've all done like as someone who does you know, has have has done social media content. I can't say I've never done something uh, to the equivalency of what you see exist in the past. But it's again, you're like you're trying to get people to notice you and that's you know, one way to do it. So, I mean clever clever, but I think that's I don't know if that's more of the goal. Maybe it's great. Uh, but that's definitely part of the goal for what I've seen. To quantify uh, I will say 90% value, 10% marketing. It's still majority value. Um, also I haven't done it today. Uh, apparently today Google announced a 70 word or line instruction on like it's just like free instruction for any coding agents like how Google do things. I'm going to check it out tonight. I'm very excited about like actual software development house, actual like very extremely smart software engineering and their principles and instill that to my own little workflow.

### Segment 9 (40:00 - 45:00) [40:00]

And then it's kind of nice. Everybody's been sharing like the best knowledge they come to this over 30 years of software programming. So it sounds like you need to release your kind of cuz you're kind of aggregating all of this. When are you going to release either an article or your own you know whatever you're building and learning. I know you're posting some of it on LinkedIn but like I feel like you need a longer piece of content. I have a little blog quchenhong. com but it's more like a personal journey and nothing technical. I also like posting silly things on LinkedIn because I like a timestamp to prove it myself wrong in like a month. It happens every single time I post it always going to be wrong in like very short time. I feel like that's a goal. It's like hey sober up you know don't make any conviction right now just keep trying. Whatever conviction you have right now it's going to be wrong. And it's the best thing you can do right? If you don't have conviction you just stand still doing nothing. First you have a conviction you spend time you learn from it you get better taste over time. Yeah yeah. What's been maybe one of the more interesting things that you've been wrong about? I've been there's so many things I've been wrong about. Um I always thought the market will be captured by um one every vertical one or two winners. Like analytic agent one or two winners legal or you know cloud right? But I start to appreciate human a little bit more in the sense that you know um it's not just the product it's the distribution it's how you work with people. So I feel like there are it's a pretty big pie for everybody in the market. So as long as you know you bet something right and you find your distribution channel you find your company you work with you're going to have a pretty healthy business to make it work. Um Yeah. I think that's my current take. Could be wrong. I think distribution and trust have always been kind of the big hard area right? Like that speaking again I think of Garry Tan. I think he said something like that or quotes attributed to him. It's like first time founders focus on products and then non-founders focus on like I think distribution is the word he does use. I always like I always add on and it's like and products cuz like obviously you need a good product but you know it's like if your product is mid but has amazing distribution see uh you know Microsoft you know you'll win or at least you'll win more money versus the product that's amazing but you know people don't know about or it's harder to get distribution on so. Yeah one thing I want to I don't know I just want to make a timestamp is like today is it easier or the worst time or the best time to buy SaaS stock? Like Salesforce. Okay well yeah okay that's like it's either really amazing guess. — I really don't know it's all the things we're allowed to But which side do you sit on cuz that's not a that is it's clear that it's one of those things right? Like it's either going to be an amazing buy and you're going to be rich in the future or it's going to be a terrible buy. Or it'll be flat I mean it could still be flat but you know yeah yeah I think I've definitely put a little into some of that cuz I'm like well if it does go up I won't be upset that I didn't buy but if it goes down I won't be upset that I didn't lose or I yeah yeah. So yeah it's either going to be amazing. I think that answer depends on like your trust about the leadership team and their ability to pivot and like transform themselves. Um I don't know but then it's like it takes personal judgment about another person who running a massive company which is a pretty interesting bet. Yeah. I know it's why they say you got to bet on yourself. yourself cuz you betting on someone else's is hard. Um Um that is also true. You bet on your convictions and hopefully that pays out you know in some ways. Yeah yeah yeah. Well one I really do hope you start putting out some longer content. I think people would benefit from it cuz you're testing it. Not everyone does. Obviously as someone who does content you'll get a lot of people who probably will either tell you you're wrong or will you know it's actually a mix. You'll get some people who tell you you're wrong. will thank you for doing it. ask you to try their product. And then you'll be like I don't have time. But I don't know I think I think it's a great way to share and let other people know and they can kind of see what you're doing cuz I think you're doing a lot of cool stuff. What are the things um you learned about analytics that you wouldn't learn without this mad this massive or YouTube

### Segment 10 (45:00 - 50:00) [45:00]

channel or this massive media audience? I think I learned more about maybe not marketing but kind of trends than anything else. I think analytics-wise I'm trying to think of uh I usually if I ever have to go deep into a subject which I'm always really bad at building series. So for anyone who's ever seen my part one part two missing part three series I'm sorry. I think I've got like two or three of those things. Um Yeah I think you often learn how little you know of a topic like you're like oh let me do a series on Airflow and then you're like okay I've used Airflow in certain context and so I understand like a piece of it but now that I want to go deeper and explain it to someone it's like oh I got to like really dig into it. So you'll often be forced to be like oh I'm going to put this out like is it going to be bad? good? Like what am I missing? What am I forgetting? Like do I need to take a different angle? Should I go deeper? Also I think it's I know that's not just analytics but like also learning to try like how to try to communicate ideas like how do I draw a picture that people will connect with? What memes do people like connect with and be like oh that's you know that's kind of what connects with people or what you know diagram. So I think those little things are all you get very immediate feedback which is nice. Yeah that's definitely a taste that you wouldn't get because is your own your yourself is the product. So if you really have a question of that how do you analyze your media marketing performance? What is your stack? What is your stack for your own hire? Well I think because I'm also busy consulting I poor what is it? Something about the cobbler's kids have poor shoes or something. You know every so often I think I pull in data like my Substack data into Snowflake and then like I'm going to analyze this for sure this time and then I'm you know after cuz I had like a few hours free or an hour free I'm like I'll do it. I'll do it later and then I don't. So you know I think there's definitely a lot of opportunity there if I were to lean more into the media side of my company but I use it more to get consulting work than I use it to like try to be a media company. So that's why I'm like well you know I get as long as I get consulting work and it keeps growing then I'm doing the right thing. Then the output is good. But there are probably opportunities I could also you know do better and put out better content and Yeah yeah yeah. I definitely I'm definitely more in vibes for my own stuff my own content. And then I also like I like being able to write what I want. I don't want to like if I if someone if I suddenly realize like oh I need to write more about this one topic I'd probably I don't know. Then you'd just be writing for algorithms and I think that's when you start to seem too much like everyone else. So I just write like when I make decisions about like content I just write what problems I'm seeing in the moment more than anything else. I'm like so you know everyone ought to be talking about AI but I'm like I'm still seeing SFTP problems in places. So I'll write about that because like great. The world is moving so fast but we're still in using technology from whenever I don't even know when SFTP was created now I think about it but we're still using that and it's still a problem pain and we're still just you know trying to make that all work. So yeah. Totally. One thing I was really curious about maybe it's already out there is how Mr. Beast run his analytics team. I know he has a pretty sizable data science team. I'm so curious about — so I know that um do you know uh I feel like I always say his name long but wrong the last name wrong but Luke Barousse? Mhm. Uh he's another YouTuber. I wonder if he like talked about Mr. Beast. Uh he did do it looked like he talked there. I think I recall hearing that he did some analytics work for them at some point and but it doesn't I was like maybe he's got some content about it but it doesn't look like he does. Um probably doesn't talk about it. Would have been cool if he did. Um Yeah yeah. I mean it'd be cool to know like I think what I do understand is like they do a ton of AB testing on um like thumbnails and I think I'm curious if they do like Disney does or like some of the other movie companies do where like they literally have people watch the movie and they'll like be filming you like track your emotions and track how you feel throughout the movie and really try to get like a sense for like when you're feeling happy, sad, you know, so I'm curious if if Mr. Beast goes to that level of like let me track exactly how people respond to every second of my videos. So His video had more views than a movie in the theater at that point. Yeah, I was always wondering

### Segment 11 (50:00 - 55:00) [50:00]

like if you measure analytics by the business success right? Like his marketing analytics got to be the most successful analytics team in the whole world for that reason. And then we're almost curious about what they're doing. Um And it's very hard to track cuz you it's like YouTube, right? Like you don't get the first party data. You get whatever YouTube gave you per se. Yes, which is not I'm trying to think. I mean it's nice that they you get to obviously see when people leave. So they give you like how when people like they literally give you second by second like when do people leave, when do people stay longer, when do people essentially go backwards to watch more of something or rewatch something. You'll often see if you do an ad break, you'll see a drop like a lot of people automatically skip ad breaks. Um So lots of little things like that. But yeah, you don't get as much information as you'd like for sure for sure. I'm sure there's a lot of extra information, but he's I think he's also just figured out to some degree also the vibes of like what people want. He gets He understands spectacle. He understands how to keep you engaged, you know, and he's willing to invest in it. I think that's the other going back to the conviction. It really is about like what are you willing to invest in? make a bet on? Cuz I don't think everyone wants to do it. Everyone's like, maybe it's right and then they don't do anything about it. So Yeah, shifting speed a little bit. I'm taking over the question now. Um Do it. For you got this question a million times a day. I I'm curious about what you say. Like for college new grads who don't have a taste of conviction yet because they're new to the job market and they studied data science for past four years or maybe a master degree and know exactly how to write random forest in Python, right? And then they go to job place, everybody be looking for a sequel person or they play that sequel person ready with the agent. What advice would you give them if they're graduating in June this year? Yeah. Yeah, I think I love that you called out the fact that like someone's got a master's degree in data science and then knows nothing about sequel. Just cuz I've run into that person a million times. And it's like And it's funny. I feel bad cuz it's like everyone wants them to do sequel and they're like, but I did not Maybe now they do. Maybe they've changed their programs, but I remember doing interviews at Facebook and when we were getting like master students in cuz we would be doing like I think like the internships for summer it would be it'd be rough. I'd be doing a basic sequel problem and or basic Python problem and even in like that. I'm like, you just need to go through a for loop and like check if it's an IP address or something. And they'd be like, can I use like Pandas? I'm like, no. Like can we just do a for loop? Just can you do it with a for loop? Cuz like you don't need that and you shouldn't have to like that shouldn't be the only tool in your toolbox, but I get why they get there and then I feel bad. I think you should still learn sequel. I think there's like a level of like data intuition and AI to intuition that you need to build. So you learn sequel to get an idea of like what bad sequel is, what good sequel is. Um You can now write more complex queries faster. So that's lucky you. You don't have to go through these 10,000 line queries. I mean you probably still will have to do that, you're not going to write them as frequently at least fully as most of the rest of us have or at least be copy pasted from one file to another and then built the rest of the query around it. But yeah, it's that's a weird thing to think about like not having to start some of these queries from scratch and be like there was a time that I had to write thousands of lines or at least go through and I still have to go through thousands of lines, but like just very different world. I think it's still worth it. You do want like a data intuition cuz I think that helps you get a sense for the business, right? Like you start to be like, oh I think I know where I could either find the data or I think I know like if we see something like revenue go up or suddenly or whatever it might be. I think I know where I could find out why that happened. Especially at a larger company where you've got so many tables and so many things like where are you even going to point the AI if you don't have at least a sense for what exists, what tables exist, how all of it relates? Cuz that kind of replicates the business to a degree and I think that's still worth knowing. Similarly with like data modeling. I think it's still worth knowing cuz you're going to apply it and if you don't, the AI will just apply whatever it wants. And then you might end up with a very, you know just crazy model that or data model that you're going to some not anyone maybe no one's going to maintain it. Just going to stick the AI over it and then it's going to get worse and worse. You're going to keep getting this mutating model that gets worse over time because no one actually fixes it cuz no one knows it's a problem and then eventually it finally collapses. So Yeah. Yeah, but Ben playing a role of somebody who are graduating I need a job, you know, like it's hard for me to find a job. What do I do? Yeah. Yeah, I think like you said you got to show. Sadly we're in a world where like I think a lot of people do need to see what you can do. Um Again, it kind of it's kind of gotten worse and worse over time. Like I used to be like, oh yeah, just like put out a few posts on LinkedIn. Now it's like, okay, maybe

### Segment 12 (55:00 - 58:00) [55:00]

like you have to do like a full project and like show people. But I do you think people like I have seen so many people get jobs that way? Like Eight years ago it was like one blog post and I've seen people turn that into a job. Now it might be a little more complex of a project, but you can do more because you have, you know, you you've got AI and I think people still want to hire people that know those skills. Like I don't think business people have suddenly picked up all those technical skills. And so if you can show them, hey, I can take these tools and make something useful. You know, people are going to get excited about it. And then maybe you're going to go back and realize that they were excited about it, but then when you look under the hood, it's a lot of like old ERPs and 10,000 line queries that you've got to fix and that's less exciting, but you know, you've got your first job and now you can kind of go from there. So yeah. I feel the same way. I feel like if I were, you know, graduating in one to two to three years, it is I got to do it now. Like start building projects now. If you're graduating this year, like I think it's okay probably. If you're graduating next year it's going to be I don't know 50% harder, but next year it'll be 200% harder. It's going to go scale from there at least that fast. Yeah. Yeah. Like I said, I think people I'm hoping we get to a place where people are willing and I think some companies are still willing to hire juniors, but like are willing to hire juniors. And I think the only other balance to that is, you know, you got to show people that you can start day one doing things. And I think that's the big thing is when I talk to companies and I this could be because I'm a consultant and people are often looking for someone more senior when looking for a consultant. Yeah, they want some they just want the problem solved, right? Like it's like, look, we don't want to bring someone on that we have to train. We want problem solved. So I think the more you can show I can be the solve, you know, the more likely it'll be that they'll be like, oh cool, you start tomorrow, you know. So Yeah, but it's I'm not saying it's easy. I'm not going to pretend, but you know, I think if you just sit and are saying, you know, oh tomorrow, right? Like we don't need to learn these skills cuz AI will do it tomorrow. I don't think that's a position you want to be in either, right? You Like you said, the best way to learn the skills is not to read about it, is to do it. You can start by reading, you know, a quick one hour intro, but like eventually you got to shift into building and doing. I agree. Yeah, I totally agree. What a time to be alive. — [snorts] — You know, I mean I have seen this meme of like it's like Ghibli-esque art and it's like remember the world before this one tweet and it's the tweet that Sam Altman releases ChatGPT and I'm like, yeah. I do feel like it was slower. I do think I do feel like we're going to work more not less somehow. Like I'm just like I think I mean I think it'll be an option. I think some people will probably work the same, but I think for some people who are competitive, there's like an opportunity to work more, but at the same time an opportunity to win more. So we'll see. We'll see what happens in the next three years, two years even. But yeah. Yeah, what a time. Exciting time. Awesome. Well, thank you. I appreciate your time, Hutching. Other than this, I don't know. Anything else you want to say before we sign off? No, this is fun. Um Awesome. If you if somebody in the school designing program are listening, you know, don't teach all the kids Python. Let them do more sequel projects. That's all I'm here to say. That's so funny. Awesome. Well, thank you so much again. And thank you everyone who's here. Bye. Thank you.