Vy Nguyen on Founding Taskless - Episode 1

Introduction

[00:00:01] Jakob Heuser: Welcome to the Task at Hand. I thought it'd be a great idea to kick off with, uh, my co-founder, Vy Nguyen, who's the CEO and Co-founder at Taskless, and it started out as a telemetry startup and went through a pretty big change.

[00:00:16] I'm gonna let her talk about that a little bit. Introduce herself. I guess we'll start with that. Vy, if you could just kind of fill everybody in on what kind of got you to Taskless, what were you doing before? How'd you get to what you're doing now? And then we'll talk about what Taskless is.

[00:00:30] Vy: Sounds good. Uh, hi, I'm Vy Nguyen. Uh, as Jakob mentioned my background, I, I'm originally from operations. Uh, my education has been in supply chain logistics and information operations management, and my, majority of my work experience has been in large scale manufacturing. Um. I got into technology because the work I was doing was consistently having to help people transition that gap from what they were doing to adoption of a new technology.

[00:01:04] And I think we're standing at the precipice of a big leap for people in that space right now where we're talking about the adoption of AI tools. Where do those kinds of tools fit and how do they help people with what they're being tasked with working on? And the adoption, the people problems around adoption of those tools because it's not just deliberate and everybody is happy and everything's hunky dory.

[00:01:31] Jakob Heuser: That's kind of fascinating. So like Taskless, itself. Obviously I'm close to the products, but I'd love to hear from you about what problem it's specifically solving and who it's for then.

Building Taskless

[00:01:41] Vy: So we originally had been building a telemetry product and um. Things changed pretty rapidly with the introduction of AI code generation tools and coding assistant agents, where it seemed like there was a twofold problem. Engineers were being tasked with doing more product delivery work and less site reliability and quality control, there was this also this pervasive belief that we can just have the AI fix the quality issues later. If you are working in that space right now, I feel your pain. But what we ended up doing was we interviewed several hundred developers who were at various levels in their career. focused very specifically on senior developers because obviously this is the group of people that is being kind of tasked with being the shepherds. Bringing everybody else along on this transition to with the AI. And the one thing we kept hearing over and over again when I would interview these developers is drowning in code reviews. I feel like a janitor. feel like all I ever do is look at prs and nobody can seem to explain their code because 90% of the time, aren't the ones writing it anymore. And the issues went beyond. Things that you could do with code review tools. Um, we can get into that later if you'd like.

[00:03:15] Jakob Heuser: So talk to me about that decision. Obviously, I, I was in the room. But I was in the room wearing the technical hat. And I'd love to hear your take on the business side about what it was like when you realized what we were building. That telemetry product wasn't going to be the thing. Was it a single moment?

[00:03:33] Was it a slow realization? And how did you, how did you reach, how did you reach through your technical co-founder work through that decision? What'd that conversation look like?

[00:03:44] Vy: I think technical folks have a tendency to focus on a specific problem set, it's that. Core value of being, uh, tenacious when it comes to solving a problem. I really admire that about developers. Uh, the big message I had to get across to you, my technical who vendor Hi Kitty, um, was that people were not willing to pay for this product. They thought it was interesting. They thought it was cool. Everybody we talked to was like, that's really cool. I think that's awesome. And like?

[00:04:20] when I asked, how much would you pay for this? And said, eh, I wouldn't pay for it. And other question was, how much pain would you go through if you didn't have this product?

[00:04:31] And they, the realization was people were more willing to put up with a lot of pain in this area, whereas the immediate pain was shipping product. And ability to ti in a timely, timely matter, get all of these code reviews completed so that the product could ship. Um, consider the aspect of you're the senior engineer, you're a team lead. you're a principal developer at a small shop, or even a big shop, your job is some cohesion in the code quality and the base, the, the code base that's being written continuously. These things doesn't stop. Um, your feedback from your junior developers will be, stop holding me up. Why are you slowing me down? And their metrics are to ship as much product as possible.

[00:05:24] Jakob Heuser: Yeah, that's true. Nobody ever got promoted for code quality. Everyone gets promoted for shipping.

[00:05:29] Vy: Absolutely. But quality reflects itself in a num, a myriad of different ways. It's customer satisfaction, it's retained revenue, a lack of churn, customer, um, expansion and just the basic costs of r And d are well understood when it comes to having a high level of quality that can be maintained across different products you ship.

Too Much Code to Review

[00:05:58] Jakob Heuser: And so to paint a big picture, then what I hear is there's all of this code. Now people are with these AI generating tools, able to create more and more and more code, but senior engineers are still being asked to be the reviewer of all of that code,

[00:06:15] Vy: Mm-hmm.

[00:06:16] Jakob Heuser: and it sounds like it's hit a tipping point. And that kind of lands on that core thesis of Taskless.

[00:06:23] I looked at the marketing page before, just to make sure I was saying it right, this idea of telling AI once this idea of deterministic rules over just adding more AI to the AI pile. And I'm very curious how you got there and kind of what was ruled out along the way to get to that idea of a deterministic rule.

[00:06:43] Vy: I had one senior, very senior developer large shop kind of. Tell me the lay of the land for him. Um, he had been a high level manager of a team and had transitioned back to an individual contributor. And his main problem was, I don't write any code anymore. literally spend all of my time, 40, 60 hours a week reading code, something else wrote. I can't even go back to the engineers who submitted the PR because. In a lot of the cases, an AI wrote it and querying the AI about what it wrote and why is kind of a futile task. So when I pushed him on this question, well, why aren't you using, you know, copilot to review the prs? Why don't you use Bug Bot from Cursor?

[00:07:40] Why aren't you using Code Rabbit or any myriad number of other tools for specifically. Code review. He said, it won't solve the biggest chunk of my problems, which is the code is technically correct, it will function for now. And when I continued to do the interviews, I saw this kind of pattern emerge where people were saying the same things in different ways. It was. I had this problem and it only propagated once a week because of the way That, our container seemed to load this one file only on Tuesdays at 3:00 AM

[00:08:19] Jakob Heuser: that that sounds horrible. And it sounds like, uh, if you had just another markdown file that wouldn't actually fix it, you'd be like, please don't call this file on Tuesdays. And then there's no guarantees that. There's no guarantees your Claude MD would ever get enforced.

[00:08:35] Vy: There's no guarantee with any sort of generative, non-deterministic AI because it's a probabilistic model. That's just how it works. Did it follow the instructions in your skills.md? We don't know. Maybe it did, and you're just gonna make the assumption

[00:08:52] Jakob Heuser: Okay. And that kind of means that determinism becomes that big play. Like determinism is, this is what's gonna make it different. This is what makes Taskless distinct and you've kind of had a front row seat then as this is un as is unfurled. We started building AI when AI coding agents really started to take off.

[00:09:14] Vy: And that pattern, that that trend is not gonna slow down.

[00:09:19] think that one really large repo on open source was saying that 90% the prs they receive, and this is in the numbering, thousands of submissions a week were AI generated and the account was coming from what was known as a, as a known bot. And this person or this account that had been created for a bot bypassed their. Whitelist.

The State of Tools for AIs

[00:09:47] Jakob Heuser: Of segues into Then my next question here, like what does the landscape look like right now? I mean, other than thousands of AI generated poll requests.

[00:09:55] Vy: Um, I think the landscape is that it's close to impossible for open source maintainers on bigger projects to stay ahead And what that relate, uh, what that means is ultimately. We have slower release cycles around products that rely, rely on community submissions, and I don't think that's a problem unique to the open source. Um, obviously when you have 2000 engineers all committing code on a daily, hourly basis, that's the same thing. They're using the same tools. You certainly have a shop style for every individual business that is not ever gonna be captured in a static analysis code review tool cause

[00:10:41] Jakob Heuser: And when you say, when you say, when you say, when you say shop style, what exactly do you mean by that? Is it just the aesthetics of the code or is it just also lessons learned? Is it more than that? What is shop style?

[00:10:53] Vy: Shop style captures the engineering decision making that you went through along the way of. Making the decisions that you have represented in your code base, it might have been lessons learned for from who knows? Years ago. And I'll tell you a really dumb story of that. The fact that when you lose that institutional knowledge, it doesn't come back. I worked at a company that was founded in World War II and I had to fix a software problem that, uh. was built in FORTRAN. You can guess that I wasn't yet when this product was built, but the person who built it was dead. They were long dead, like nobody could understand how this product worked, and we had to make some significant updates to the base data that it was being, that was being used when it displayed the query. I was the only person that they felt capable of untangling this mess, and I don't know. FORTRAN, I had to learn.

[00:12:03] Jakob Heuser: And you had to learn without the assistance of something like Claude Code 'cause that that clearly wasn't. Around at the time

[00:12:10] Vy: There's no Claude code skills file for FORTRAN based mainframe ops.

[00:12:18] Jakob Heuser: there might be now, which I guess kind of comes to this idea of like,

Is it a Bubble?

[00:12:22] Jakob Heuser: there's all this technology, there's all this money pouring in.

[00:12:26] Is it, is it a bubble or is there a bigger platform shift or is it kind of both at the same time like. What do you see happening around this?

[00:12:34] Like you can now code FORTRAN I don't know why you would, but you could. Um,

[00:12:40] Vy: that use this still.

[00:12:41] Jakob Heuser: yeah, so like, there's clearly some nuggets of usefulness in this technology, but it also seems like there's a lot of hype and a lot of like extra funds being poured in to pump up that hype machine.

[00:12:53] Vy: I can talk at length about the economic of the capital bubble around AI and the, the major debtors in the industry, being the ones that have to benefit the most from it. That's a different topic.

[00:13:09] Jakob Heuser: Well, let, let me ask you a more pointed question then. Um, if all the hype disappeared tomorrow. What's the embedded things or the valuable pieces that are left behind? Like what's the good stuff when all the hype is gone?

[00:13:22] Vy: I mean, this quote unquote AI revolution at its core is basically just. Natural language processing for an a machine learning model, it's good at all the things that machine learning is good at pattern recognition. Um, maybe to some extent when you extrapolate pattern recognition down to what it is, you can use prediction engines and something like the highlights game.

[00:13:51] Tell me what the difference is. And we see that example in real life where if you were to ask copilot or Claude code, tell me what's in this diff and don't just tell me how many lines and all that stuff. I already got that from my, from my GitHub. Tell me what changed and why. It's the why part that it struggles with, because that's intent. We learned, I don't know what, from 2018, that intent is probably the hardest thing for machines to parse from human speech.

Taskless' Future

[00:14:23] Jakob Heuser: So when you see all this playing out. Do you feel more or less confident in what Taskless is building?

[00:14:29] Vy: I think as we move in a world where people. Take in the hype and try to expand, use cases into things that machine learning models are legitimately just not good at, but can, based on a reward structure of the algorithm being what it is. Um, it will be, it will be expectations of correctness that don't exist. And in the space between existing and correct Taskless that void, whether the key use case be proton, uh, folding in which, you know, we are, we're talking about amino acid folding and that that process of it, in and of itself is non-deterministic. If we don't know when it will happen, we don't know how it will happen, and it can't be repeated every time in the same ways. But on the other end of this, doing silly things like geospatial analysis. Look at a map. Show me what changed. Highlight for me the areas that are different now. Same satellite, Passover.

[00:15:42] Jakob Heuser: I love this 'cause your fascination with technology. I feel like your path to where you're at, at Taskless is a direct result of coming from not the typical founder path. It wasn't big tech into product manager, into yc, into Surprise as a company. You had a background in operations, which meant you gotta play with a bunch of different systems.

[00:16:06] You got to see a bunch of different business use cases in a way most founders never get to see. Uh, my question around that is really just in that role and through your role with USC Marshall and a bunch of other groups you work with, you talk to companies that aren't AI companies and they're trying to figure out what AI means for them.

[00:16:26] And I'd love your read on what you're noticing when you talk with these companies.

[00:16:30] Vy: When I talk to companies that are, they're either under pressure or they have adopted AI in some sense, but they're trying to make, they're trying to capture the value around those acquisitions. There's a lot of expectation, and then there's the, the hard piece, which is the business transformation and culture change that has to happen alongside. You can't just deliver software and expect people to automatically make your metrics jump the way the sales person said it would. Right. I, um, I think the biggest challenge will be in ag. Where there is a strong cultural resistance to change and technology, and even when we are talking about industrial food systems like and Cargill and, know, all of these companies.

[00:17:24] that produce massive amounts of crops and it's, it's requiring all kinds of very disparate and, you know, non. Um, what is the word I'm trying to say? Non-structured data. It's weather patterns. It's climate change analysis. It's, it's something as simple as policy decision impacting whether your water supply will be delivered when it is. And then there's other stuff like, you know, The genetics of your product. All of these things. Are within the highly capable silos that can help people process. But I think the challenge is that when you, when you oversell something and that promise is not realized immediately, then you have lost the trust

[00:18:18] Jakob Heuser: The trust is kind of a big thing too, isn't it? Like when you talk about the trust.

[00:18:23] Vy: Would, that happened directly in the ag industry.

[00:18:27] Uh, so John Deere mainstay of ag, they build tractors and heavy equipment and combines and the big machines that are needed to do the kinds of work we ask of our farmers. Now, a horse and plowing gonna cut it. So John Deere introduced this whole subscription model Tractor as a service, if you will. the trust.

[00:18:49] was annihilated almost from the instant it was introduced because people could not fix their equipment. Guess what they did? went on the af the second market And bought all the old equipment that didn't have a service contract to type, uh, attached to it, and it had no software. So we're back and I don't know the nineties when it comes to Ag Tech.

[00:19:10] Jakob Heuser: And so to tie that back to the AI stuff then it sounds like if an AI, if a company wants to adopt AI in. One of these more legacy spaces, um, trust is gonna be paramount if they don't manage.

[00:19:23] Vy: regardless of your industry. So we talk about developers, you all are a very skeptical bunch, and some of you are actually very cynical when it comes to new things. Which is funny because you're the ones building it, but if you told developers that I only have to tell the AI once, and now they're going back and having to see the same problems again.

[00:19:46] If I don't have a PR rejection rate go to zero immediately, don't believe you anymore. So what we are working very hard on is making sure we deliver that in as little of those interaction loops as possible. You make a rule the first time and it works the first time and it works every time thereafter.

[00:20:06] You never see that problem again. I'm sorry to interrupt your question. I think you had, uh, maybe a different track in mind.

[00:20:12] Jakob Heuser: Uh, it's fine. I was thinking specifically just around how these legacy companies with the trust we bring in AI to these legacy companies.

Trust in AI Systems

[00:20:24] Jakob Heuser: And it seems like just based on your John Deere story, that they get one shot, they get one chance to really keep that trust and those legacy systems aren't gonna be tolerant like tech companies to, oh well it works 80% of the time, a hundred percent of the time, and that's not gonna be acceptable in most industries.

[00:20:46] Even if you get a that up to like 95% of the time, that's still not good enough for a lot of places.

[00:20:53] Vy: If you went to a doctor who said they were effective 95% of the time, would you continue to go to that doctor?

[00:21:00] Jakob Heuser: No, that's, that's totally valid. Like my doctors write a hundred percent of the time, 95% of the time. And you know what, I'd actually really, I wouldn't be okay with that. I wouldn't be okay with that at all.

[00:21:15] Vy: the human cost of that is very obvious to you as the person who is being worked on.

[00:21:20] Jakob Heuser: Yeah. As a person receiving medical care, I'd like, I'd like at least a couple nines in my, in my guarantees here.

[00:21:27] Vy: I, I'm going with seven nines at minimum.

[00:21:30] Jakob Heuser: I think that hits on a really core note though. Like as an engineer, if I'm, if I'm tasked with, I need to ship code that works. My AI keeps generating code that is semantically wrong. It calls methods that we don't call anymore. It calls services that we don't use anymore. I'm writing a UI and it has decided to use CSS modules, tailwind, and inline styles all at the same time.

[00:22:00] All of those are technically correct and are going to literally blow up in my face in a matter of weeks. I don't think more markdown files is the answer, and I think you'll agree because of obviously with Taskless what they're building, but it feels like adding more markdown isn't gonna solve this problem.

Why Markdown Wasn't Good Enough

[00:22:20] Vy: Consider markdown your skills, dot MD files to literally be your call center triage. Flow chart, customer calls says, I have a problem. what's your problem? And then you kind of go down this decision tree and you're supposed to arrive at a solution and then you hang up the phone and they give you a five or whatever. Um, the more complex the systems and the longer those decision trees become, the less likely you are to have a favorable or even desirable outcome. Like you have a 20,000 line md, how are you certain that it followed every single one of them

[00:23:01] Jakob Heuser: I can actually say it from firsthand experience that it doesn't. Uh, you

[00:23:06] Vy: you added to it because the last time it screwed up, you

[00:23:09] were like, well, don't do that. Let me fix this.

[00:23:11] Jakob Heuser: Oh, yeah. I put it, I put it in asterisk and everything. I was like, asterisk, asterisk. Please never do this. Asterisk. Asterisk. And, uh, yeah, it doesn't, it doesn't work consistently.

[00:23:23] Vy: So now you've got a dot MD.

[00:23:25] File that's 20,013 lines long and it still doesn't work, right?

[00:23:30] Jakob Heuser: Yeah. And it's very frustrating.

[00:23:31] Vy: So you propagate this across multiple systems, multiple languages. Tailwind in and of itself, you know, it has very Different ways You can do specific things like call color.

[00:23:47] Jakob Heuser: You can call it dynamically. You can use their bracket notation to put arbitrary values in, and there's no guardrails like the AI is going to generate what it thinks is best.

[00:23:57] Vy: Suppose you use a contractor and now the contractor's AI is using their MD file, not yours. And even if you gave them yours. It is still working off of the presumption of how it worked before. And now

[00:24:10] Jakob Heuser: Even the, or even the model changes. I, I can't guarantee that sonnet, I can't guarantee sonnet works like GPT and is going to read my markdown file and behave the same.

[00:24:21] Vy: So I had talking some with a, a developer who was working on a personal passion project on their own. It's a game. game. And every single time there's a new model released, they go through this exercise of running the same code exercise through the different model and seeing what changed. Is that a good use of your time if you're being paid to build a game? I mean, that's

[00:24:48] Jakob Heuser: Erases all the productivity you were hoping to get if now you're constantly rechecking stuff.

[00:24:53] Vy: Yes.

[00:24:55] Jakob Heuser: Yeah, that

[00:24:55] Vy: And there

[00:24:56] Jakob Heuser: sounds pretty terrible.

[00:24:57] Vy: there was apparently a guy who received a, a code change and they were not aware that the contractor had been writing code in traditional Chinese, so he couldn't read it. And so he asked Claude to translate it. So you can imagine how that went as a bad game of telephone,

[00:25:20] Jakob Heuser: Yeah, that, that feels like a lot of meaning gets lost almost instantly.

[00:25:24] That probably could, that probably could have been fixed by just having some rules up front that say, this is how we code here. And the agent being told this is wrong, if it generates it wrong.

[00:25:35] Vy: and Yeah, that can be you put, you stick in your skills MD file. Let's suppose you did, and it still came out in Chinese. Now what?

[00:25:47] Jakob Heuser: Yeah, you almost need some sort of like

[00:25:50] Vy: Well, you can't go back and do it again.

[00:25:52] Like you want a rule and you want the rule to be the reason. This code never goes to you

Taskless' Core Thesis

[00:26:01] Jakob Heuser: So the big thing that Taskless ultimately sells, and this kind of comes back to Taskless's core thesis, um, I was kind of just noodling on this before our chat today, and the core product is consistency. What Taskless sells is a guarantee of consistency. It happens to be in the form of rules. But there's this belief that consistency is actually, if you, if you go with speed and there's no consistency, it's just a liability.

[00:26:29] You're just generating a bunch of stuff with no guarantees.

[00:26:33] Vy: Garbage in, garbage out. you know what consistency translates to at a hyperscale

[00:26:38] Jakob Heuser: What's that?

[00:26:39] Vy: you cannot fucking scale if you can't execute the same way consistently.

[00:26:45] Jakob Heuser: We should probably just print that one out.

[00:26:47] Vy: Okay. You want a t-shirt?

[00:26:49] Jakob Heuser: Maybe I, I'd wear it. I'd wear it. I, I think that's a really important thing though, like speed without quality is this liability. It's a problem waiting to happen and I don't.

[00:27:02] Vy: before, like my background is in is is in operations, and when I think about. A well functioning operation. It is functioning within some boundary of expectations. There can be really good and there can be kind of bad, but it's always within.

[00:27:18] Jakob Heuser: So let me ask you something then. If, if Taskless thesis is correct, that consistency and determinism is how you solve inconsistent AI output, what does the world look like in a few years if, if AI's core belief is true?

[00:27:35] Vy: If Taskless's, core belief is true?

[00:27:37] And let's say in a perfect world everybody has a Taskless something like it, I can stop looking at a dashboard on cloud uptime that looks like a freaking stoplight.

[00:27:48] Jakob Heuser: Fair

[00:27:48] Vy: it red, yellow, and green every day, all the time?

[00:27:52] Jakob Heuser: because shipping software is hard.

[00:27:54] Vy: I know that, but we have been shipping software this whole time and it was never this bad. I think to my point, huge companies with giant budgets and like 3000 engineers to throw out these problems are screwing up like this every day. And it was so bad. The Financial Times wrote about Amazon's failures. And yet and you, we have shipped code that had an uptime of five nines with no, without even trying. What are they doing wrong? Or maybe what are we doing right? And it's stupid to think that it's something wrong with the company. I don't think it's the company. I think it's the culture around you do with the tools you have. The resources that are given to you.

Advice For Your New-Founder Self

[00:28:46] Jakob Heuser: Uh, one last question I guess before we wrap up, because Taskless has been a product through a pivot and you've been on this journey for a while. What's something you would go back and tell you and your co-founder based on what you know now? If you could go all the way back in time, use that time machine to go back.

[00:29:04] It, it's been a while. Uh, if you could go back, what would you, what would you tell them?

[00:29:09] Vy: I would go back to the basics of, um, software development is about your customer development and foremost. If you aren't solving a brand new problem, how are you solving it better?

[00:29:23] Jakob Heuser: That's good advice.

[00:29:24] Vy: big lesson we learned was don't build anything before you know what you're gonna do. If you don't understand the problem, don't build anything.

[00:29:32] Jakob Heuser: There's a lot of products out there right now that are building to solve a problem without, or they're building a solution without even asking whether or not it's a real problem yet.

[00:29:41] Vy: I don't know if you've seen these billboards when you're driving Long 1 0 1, but what I see every time I think of an Everything app is a lack of product vision. We don't know what this product does. We don't know what our customers value, so we're kind of hoping we throw a lot of money at the wall and something sticks, and then we have this sort of message that emerges out of this pattern, and I think that's a stupid way of using a hundred million dollars in capital.

[00:30:08] Jakob Heuser: No, that that's fair. Especially when the companies that are. Adopting this don't even know what their problems are. I think Taskless is in this unique position because engineers are specifically saying, code review suck. I have to say the same feedback every time. This is getting old and also putting it in like my CLO MD agents MD file isn't doing anything to make this better.

[00:30:30] Vy: The more technology you throw at a problem where you don't have business fundamentals, the less likely you are to hit any kind of target whatsoever.

Outro

[00:30:39] Jakob Heuser: Good advice. I appreciate you having this chat. I'm looking forward to when the hat's on the other side and you get to interview me. Um, everybody. This is, this is Vy Nguyen, CEO of Taskless.

[00:30:49] This is the task at hand.