Everything’s a Fad (Including This Podcast) — with Benn Stancil
This transcript and summary were AI-generated and may contain errors.
Summary
In this episode of The Test Set, Michael Chow and I sit down in Times Square with Benn Stancil, who spent a decade building Mode Analytics and now writes the Substack at benn.substack.com. Benn’s blog isn’t about any one thing; his stated goal is that readers leave feeling like “that was 10 minutes I enjoyed” — no takeaways required.
Benn traces his path from a DC think tank job (which he compares to a data analyst role, except Congress people don’t care what 22-year-olds think about the economy) through Yammer and into Mode, which he co-founded in 2012 and sold in 2023. The original Mode blog was 538-inspired pop-culture data analysis — the first post was about Miley Cyrus — and was a deliberate counterweight to “here are five tips for running a data team, tip five is buy modeanalytics.com.”
Much of the conversation is about a cultural shift Benn frames as going from Nate Silver to Rick Rubin. In the 2010s, being “data-driven” was a status signal — you showed off your big data team, your A/B tests, your charts in board meetings. Now the culturally ascendant companies (Cursor is his example) say they test their agent harnesses “mostly by vibes,” and taste has replaced quantification. The modern data stack, like big data before it, was an era — a useful bucket for a set of companies that launched on Product Hunt in 2017 — and that era is essentially over. He walks through the Jordan Chiles / Paris 2024 Olympics gymnastics scoring controversy to illustrate how over-quantification of judgment produces its own kind of failure: we spent a year in court fighting over how many seconds you have to protest a deduction rather than who did the best routine.
We spend a good chunk talking about AI and what it does to the economics of creating things. Benn’s framing is that software is becoming content. If you can vibe-code a working app in a weekend and someone else can in two days, the market can’t absorb 10–100× more software the way it can’t absorb 10–100× more TikToks. He points to Steve Yegge’s “Gas Town” post (layers of Claude Code instances orchestrating each other) as the current extreme — men in tar pits burning enormous amounts of money for a thousand attempts where only one or two can win. His counter-model is a “boy band”: small, committed groups making a profitable business without scaling into a traditional corporate structure. I talk about how AI has let me revisit a 20-year mental backlog of side-project ideas I never had time for, and the discomfort that comes with not wanting to close the laptop anymore because the agents could still be doing things.
On writing, Benn uses AI as “a thesaurus” and occasionally as a Codenames-style partner for finding unexpected associations between ideas — but never as the writer. He describes walking into a Lorde concert, noticing the album Pure Heroine, and retroactively finding the thread that tied together a half-written post on AI addictiveness. The secret, he says, is that you cheat: you start with the analogy and glue the ideas onto it, not the other way around. He cites Matt Levine (makes you care about things you shouldn’t) and a chemistry Nobel laureate who gave talks with 300 slides in 20 minutes as influences.
We close on VCs — Benn wants “the Simon Cowell of VCs” who would just tell you honestly that they hate it — and on what traditional BI looks like in an AI world. His guess: we don’t build chatbots that write SQL queries for us; we bypass SQL altogether, because LLMs do approximate “math on text,” and a lot of the reason we quantified things in the first place was that math was the only way to look at data at scale. Unstructured data — Starbucks counter conversations, audio clips of users yelling at your website for $5 — starts to become collectible because there’s finally something to do with it.
Key Quotes
“My ambition with it, and it’s a blog about data stuff, so God knows if it actually can do this, is I want people to read it and leave it thinking that was 10 minutes I enjoyed. Like, I had a good time. There’s no takeaways.” – Benn Stancil
“The modern data stack — my view on that is always it’s the same as big data. It’s an era. It’s a useful term for a thing that happened over some period of time. Data companies that launched on Product Hunt.” – Benn Stancil
“Status has become a lot less about how well do you understand the numbers and more of just like — Rick Rubin became this weird, like, that’s who we all want to be, not Nate Silver.” – Benn Stancil
“There’s a thousand companies in San Francisco right now, two of them are going to work out — because only two can. A whole bunch of people are like these agents: the point of their work is to be thrown away.” – Benn Stancil
“It feels more like software is just content at that point, as opposed to a tool for anything.” – Benn Stancil
“I basically use AI to write stuff in one way, which is as a thesaurus. The thing it writes, I do not want to ever say a single word that it ever writes.” – Benn Stancil
“How did you come up with that? The answer is: you cheat. You didn’t come up with that. You came up with one association with Lorde, and then you read some Lorde lyrics and you found four others that felt kind of attached, and then you hook other ideas onto them.” – Benn Stancil
“I would appreciate a VC where you walk into it and it’s a little bit of the Simon Cowell of VCs. They’re going to hate it. I’m not going to be offended when they hate it because I know they hate it. And if they like it — Simon Cowell liked my song — oh my God.” – Benn Stancil
“The point of numbers was to do math. We quantified stuff so we could do math on it. But if you can do math on stuff that’s not quantified, do you do that much math? Maybe not.” – Benn Stancil
“You can just do stuff, and then the way you’ll figure out what’s actually exciting is by doing that.” – Benn Stancil
“I have a whole mental backlog of projects I thought of building over the last 20 years that I just never had the time or inclination to work on. Now I can have a terminal tab that’s just working on that side project I thought of eight years ago. It’s nice sitting there watching Netflix at home and just nudging along my little agents working on my personal side projects.” – Wes McKinney
“Do you like being able to close the laptop now? No, and that’s becoming a bit of a problem. When I close the laptop, I feel guilty. I’m like — the computer could be doing things right now.” – Wes McKinney
Transcript
[Podcast intro]
Welcome to The Test Set. Here we talk with some of the brightest thinkers and tinkerers in statistical analysis, scientific computing, and machine learning, digging into what makes them tick, plus the insights, experiments, and OMG moments that shape the field. Today we’re sitting down with Benn Stancil. You may know him from his writing at benn.substack.com, where he’s just as likely to weave in Lorde lyrics or Olympic gymnastics scoring drama as he is to talk about SQL. Benn spent a decade building Mode, one of the early modern analytics and BI tools. We talk about how he got started, we talk about the layers of hype inside the tech ecosystem, and we talk about why he’s still optimistic about writing in an AI world where output is cheap. I really like how he uses AI as a source, an idea generator, but not necessarily as a writer. And I have to say, for someone who comes off roundly as a pessimist, Benn gives us a lot to be optimistic about. So without further ado, Benn Stancil.
Michael Chow: All right. So Benn Stancil, welcome to The Test Set, where we are really interested in talking to the people behind the data. We’re here in Times Square, and I’m joined by Wes McKinney, who’s a principal architect at Posit. And I’m Michael Chow, a software engineer at Posit. And I think we’re so excited to talk today about all of the takes you’ve had over the past decades on data. And I feel like, maybe just by way of introduction, one thing that really stuck out to me as I was looking at your writing is that you’re the type of person who seems to be writing lyrical essays on data that might involve a SQL poem, or a bait-and-switch where you give someone a quote about analytics and then reveal that it was actually a 1904 quote about accounting. I’m so curious about your writing and thoughts on data, but maybe to start, do you mind telling us a little bit about yourself?
Benn Stancil: Sure. So I started working in tech 12, 13 years ago. Initially worked in DC, did a think tank job. It’s very like a DC job — I did policy research. It’s actually pretty similar to a data analyst job. It’s just, instead of telling people, PMs or engineers, what people are doing on their product, you’re basically telling Congress people what you should do to the economy. And Congress people do not care what the 22-year-old think tank thinks you should do to the economy. But it’s the same kind of idea — you take a bunch of data, make some recommendations.
So I did that for a bit, then ended up joining a tech startup in 2012, was there for a little while. And then me and two other folks who worked on the data team at that company started a company called Mode, which built an analytics and BI tool. We did that for about 10 years and then have kind of done a couple of things since then. It got acquired in 2023, worked at the acquiring company for about a year. Spent a little bit of time doing some side projects, spent a little bit of time working on political campaigns, and periodically yelling at the internet about whatever data nonsense or tech nonsense or 1904 quotes about accounting, I guess.
Michael Chow: Right. And I saw you started on the Mode blog pretty early. You were sort of like the early posts on the Mode blog?
Benn Stancil: Yeah. So when we started Mode, there were three of us. There was a guy who was an engineer who could write code, and like chain him to his desk and be like, go build our thing. There was the guy who was our CEO, who was personable and could talk to investors and could make friends with people like Wes, who was also working on something similar at the time. And then there was me. I was neither of those things. I was not going to go out — like, please stay inside. But also you can’t write code. So what are you going to do?
And so I basically started writing this blog as I needed something to do. Why was I there to begin with? I don’t know, you’d have to ask them. But I started writing this blog that at the time was very 538-ish — do analysis on pop culture stuff. The first blog post was about Miley Cyrus. There was stuff about parking spots in San Francisco, and weed — prices of weed scraped from a price of weed website — and stuff like that. It was not related to tech at all. Or Mode, I guess.
Michael Chow: Or did it have Mode?
Benn Stancil: No, it did not. I mean, nothing existed at that point. The first blog post was three days after we started the company. And we were fairly intentionally trying not to do “at the end of this, here are five tips for running a data team. Tip five is buy modeanalytics.com.” So it ended up being sort of niche-popular around data people. The original inspiration for it was things like the OkCupid blog, which 2010 data people were obsessed with.
Michael Chow: Called Dataclysm, or?
Benn Stancil: No, no, that was the book. But it was that sort of stuff. Some people liked it for that. But eventually I started having other things to do. We had customers and we had support tickets to answer, and I had to respond to support tickets. And so I kind of stopped doing the blog. And then the Substack was a return to that — or an original attempt to return to that — and ended up taking kind of a different direction. It’s not pop culture data analysis stuff to the degree it was.
Michael Chow: Right. And you’ve been doing the Substack for like five or six years?
Benn Stancil: Five years, I think. I started doing it in the middle of the pandemic — I want to say during COVID.
Michael Chow: I think it’s so interesting. We’ll have a lot from the Substack to discuss — so many interesting topics on data and AI. But maybe one last bit of context: I’ve run on the Substack a couple of times. This Substack is not about anything in particular. You’ve at different times tried to set people’s expectations for what the Substack is about.
Benn Stancil: Like, I don’t know why you go to it. I’m not really trying to give anything in particular. Originally it was: okay, you spend a bunch of time building a data thing, you have a bunch of ideas, so “I’ll write something” — intended to be more like “here are tips for doing stuff.” I got kind of bored with that pretty quickly. And it became more of “these are the things that are interesting to me.”
My ambition with it, and it’s a blog about data stuff, so God knows if it can actually do this, is I want people to read it and leave it thinking that was 10 minutes I enjoyed. Like, I had a good time. There’s no takeaways. The intent is not to have someone come away and be like, “oh my God, I’m going to go implement this.” It’s “okay, I learned a thing. It was kind of fun. Whatever. I can leave and walk away and not have something I’m supposed to do with it.” I don’t know whether I deliver that or not. For a blog about SQL stuff, probably not. But we can dream.
Wes McKinney: Yeah, I mean, I read the Substack. In the early days I was like, wow, Benn’s publishing a lot of Substack posts, and it usually would start on Twitter — you know, before it was X — and it would be something like “it’s Friday, so let’s fight about this.” And it would be some topic in data engineering, analytics, startups, the tech industry, whatever was on Benn’s mind at the moment. The consistent theme is that it’s a blog that has a lot of technical content and observations and data about what’s going on presently in the industry, but mostly it seems like it’s written to be enjoyable. The reason I keep reading is because I enjoy reading it.
Benn and I met — I was also doing a startup in San Francisco back in 2013. I was working on something in the same kind of domain, visual analytics, business intelligence. My company didn’t go very far; Mode ended up surviving over a decade. But we got to know each other back then. I remember my first impression of Benn, especially reading the Mode blog posts, was: wow, this guy knows a lot about SQL.
Benn Stancil: Yeah. This person has really suffered writing some tremendously complicated Vertica queries, essentially, or whatever you’re using back at Yammer.
Wes McKinney: Yammer was Vertica?
Benn Stancil: Yammer was Vertica.
Wes McKinney: What I like about the writing — I read a lot of blogs. I prefer to just read people’s blogs, especially smart people who have spent a lot of time thinking about things that I haven’t thought as much about. I’m especially interested in people that have contrarian takes, or that are not accepting the standard dogma, because I feel like in the tech community, especially in the startup ecosystem, there are almost these mantras that get repeated — these shared delusions or things that you’re supposed to believe. In San Francisco in particular, there’s essentially a monoculture, a groupthink where if you don’t conform to the group think about what is good and what serves the overall operations of the tech ecosystem — in the last 10, 15 years, up until very recently, the zero interest rate period — money was cheap, startups were raising tons of money, and there’s this dogma that “it’s good, it’s great, everyone’s raising a ton of money.” But is it productive? Are the companies making money? Are we advancing society?
I don’t know that there’s any easy answers to that. And just being contrarian and poo-pooing on things because you don’t like them, or because you’re not part of companies raising a lot of money or being nominally successful — I think it’s worthwhile to take a step back. We’re building a lot of stuff, but is it helping? The modern data stack is very complex. The whole landscape diagram has thousands of companies on it now. You need a microscope to check out the ecosystem of different open source tools and companies. It’s a little overwhelming trying to make sense of it.
Benn Stancil: Well, we’ve cleaned that out some, the modern data stack. There are fewer now.
Wes McKinney: Yeah, there’s some consolidation happening.
Benn Stancil: Yeah, we’ve removed some from the map. But it is interesting — I know for the modern data stack stuff, I listened to you and Tristan talking I think last year about that post he released, “Is the modern data stack useful? No.”
Wes McKinney: That’s an interesting one. You got to ride along with him and almost discuss this concept. I know you’ve hit on the modern data stack as a kind of hyped-up thing.
Benn Stancil: My view on that is always it’s the same as big data. It’s an era. It’s a useful term for a thing that happened over some period of time. What is big data? I don’t know. It was the thing we all got excited about in 2010 or whatever, like we’re going to predict the future with just magic and data science and stuff. And nobody really talks about that anymore. That’s not a thing. But was it successful? Kinda. Was it what people said it would be? The Economist had all these articles — I remember 2014, it was like “oh my God, the world.” Did that happen? Not really. But it’s useful as a marker in time.
I think that’s basically what this was. The joke I kind of always make about it is: it’s data companies that launched on Product Hunt. That’s sort of when this happened. I don’t know if people use Product Hunt anymore, but they certainly did in 2017. There’s a grant-bottoms-up, consumery feel. It wasn’t “go sell to giant companies.” It was all product-led growth type of stuff. So it’s useful as a bucket to put these spiritual things in. Does it still exist? Not really. But that’s not because everything’s dead. It’s because that era is past. We’ve moved on to something else.
Michael Chow: That is interesting. It seems like it connects to some of the stuff you’ve talked about lately. In the talk you gave, “In the Long Term, Everything’s a Fad,” big data is kind of being at a point in time and serving a purpose for specific people. I was really curious about this talk because if I had to break it down, it’s like 15 minutes of you talking about the mechanics of gymnastics scoring, very complicated, and two minutes tying into data. But to your point about wanting people to be entertained for 10 minutes — for the gymnastics juicy goss alone, I feel like totally sold.
Benn Stancil: The stuff that we do is not that interesting. A giant set of lawsuits over who wins medals in women’s gymnastics is way more interesting than the ins and outs of SQL.
Michael Chow: Right. Do you mind really quick, just recapping the little tidbit of juicy goss, what went down in the Olympics?
Benn Stancil: I think I can do this. Okay. So there’s a women’s floor exercise final in 2024 — Paris. A bunch of gymnasts go. Simone Biles is supposed to win. She messes up a couple of times. She’s in second. There’s a Brazilian named Rebeca Andrade who’s in first. And then there are two Romanians tied for fourth.
The way women’s gymnastic scoring works: there’s a starting score value, which is computed from the elements you do; there’s an execution score, which is kind of subjective judging — was it artistic, that sort of stuff; and then there are deductions, which are the “you stepped out of bounds” things we all yell at people from the couch about — “oh my God, they took a step on their landing, a tenth.” After all but one person had gone, there were these two Romanians tied for fourth. One of them had the tiebreaker because she had a higher execution score.
And then the last gymnast to go was an American named — why am I suddenly blanking on her name? Jordan Chiles. Jordan Chiles gets a score that puts her in fifth behind these two Romanians. Her coach realizes that they calculated her difficulty score wrong. Basically, they added up the elements wrong, and had they calculated correctly, she would have come in third. So the coach protests, she ends up getting third. The judges say yes, we should have calculated differently.
After that happens, the woman who was in third and then became fourth said, “Actually, she protested too late. You have a minute to initiate the protest; she initiated after a minute and 24 seconds.” It goes to some sort of tribunal. The tribunal said it was after a minute and four seconds — so actually, Jordan Chiles does not get that change in score. She goes back to fifth. But then the other Romanian who was in fourth the whole time said that she got a deduction for stepping out of bounds, but she didn’t actually step out of bounds. There’s video. Her heel was above out of bounds, but she never touched. You have to touch to be out of bounds. She protested: “I should have gotten third because had I not gotten that deduction, I would have actually been above the other Romanian.” Even if Jordan Chiles had her score commuted correctly, she would have been above her.
So it became this huge thing, lawsuits. Ultimately I think it is still in court. There’s a Swiss court that does arbitration for this stuff. The initial ruling was basically that the first judging was correct because Jordan Chiles could not actually change her score — the review of a deduction is not reviewable. All these protests, a whole thing. The point was, we were fighting over all of these weird, tedious…
Michael Chow: It became a legal battle over what’s the letter of the law, not whose routine is better.
Benn Stancil: Yeah, at the end of the day, if people want to know who did best, we’ve entered some very detailed quantitative hell to figure out — maybe in an unsatisfying way. You see how we get there: there was a giant controversy in 2002 or 2006 where the judges messed up, and it was more of the perfect-10 thing. Judges subjectively perfect-10, a lot of it was just the “Russian judge” kind of stuff. And so they’re like, okay, we have this much more structured thing. There are these giant rule books with tons of elements.
But then it becomes this weird thing where the whole game is playing that rule book as opposed to doing the best routine. The best routine becomes gaming the math. There’s stuff in figure skating where you get bonuses the later you do elements. People will try to do harder elements later because there’s a special 20%. What used to be artistic sports became extremely quantified around weird rules. How many seconds do you have to protest? Became what decided the medal.
Michael Chow: Yeah. And I find it so interesting because I really saw myself in that tale. The point you made about — if the CEO hits me up and asks what should I do, and I break out all these formulas, when they could go to ChatGPT and get a very quick answer. I’m curious if you could recap that idea: we used to be very quantitative and model-oriented as a huge virtue, but now there’s this world where we can actually get answers.
Benn Stancil: It’s all vibes. In the gymnastics case, before, it was vibes — you’re an expert, tell me how I did. Okay, who does the best? “I don’t know, this person presumably knows more, and we can all dispute it, but it’s kind of vibes.”
To me, the data stuff was — and this is sort of the “everything’s a fad” — it’s better, kinda, but a lot of it is just that was the culture. The culture of Nate Silver became popular, and things like OkCupid, a bunch of people got into this stuff. Our generation basically got into the “well, actually, let me show you all the numbers and why you’re wrong,” and that became a signifier of how smart you are, how successful you are. There has been a lot of pushback to that. Things like this gymnastic stuff happen when it gets so quantified that it’s just like — my God, what are we doing? It’s been a year. We still don’t know who won this thing, and the routines are 90 seconds long.
There hasn’t really been an alternative, but now AI kind of is an alternative. You can just be like, “what’s the vibe?” You give it a thousand support tickets, and you say, “what’s the vibe?” And if I can get an answer in five seconds, and I can do that as much as I want, I’m not going to go ask an analyst to give me some giant report of a bunch of numbers that don’t align — “well, did you think about this way or this way? It depends on how you count what a user means.” What’s the vibe?
Michael Chow: I feel like we’re about to cut Wes loose on this, because I feel like Wes has a lot of opinions. But if I can just put a thread through it — it’s so interesting to hear you describe the Nate Silver / OkCupid blog era, which you mentioned were kind of like inspirations for the original Mode blog posts, and then the shift into today of more vibes, which is a lot of the topics we hear right now on your Substack about AI and people vibing out. Wes, you came up in a recent post that’s very vibe-code related.
Wes McKinney: Yeah, there’s definitely a lot of topics there. I think all three of us have spent a lot of our careers focused on building tools and environments to make it easier for people to do things with data. That might be writing SQL queries and computing things, looking at the answers, making plots. Or in the case of pandas, it was “I just want to be able to read CSV files, get a little bit of data out of the database, clean it up, wrangle it, and get to the answer.”
Surprisingly, in 2007 when I started working, I was shocked at how difficult it was to undertake those basic data questions. But now we’re in this weird era where the whole concept of usability of data tools and the environments in which we ask those questions is being completely turned on its head. I find myself, for probably the first time in my career, essentially not writing code — just looking at code and pressing return, more or less. The way that we work and the shape of our tools and how we ask questions, get answers, and make decisions certainly hasn’t completely changed yet, but it’s on its way to changing. We don’t have a clear picture yet of what that’s going to look like.
Benn Stancil: There’s a broader cultural thing in all of that, which is: in 2010, 2014, whenever, data was a status thing. Big tech startups had big data teams — not capital-B Big Data, but they had data teams, lots of people on them. One of the things they would show off is how smart they are about making decisions because they’re so data-driven. You go to a VC pitch or a board meeting — they want to see all the numbers. That was how you represented that you were with-it.
Now, to your point — not just in vibe coding stuff, but if you talk to companies like some of the leading AI companies — not necessarily the Anthropics of the world, but companies like Cursor, that cohort — they’re the big companies with not a ton of people. Data is not a big part of what they do. It’s very vibey. Cursor famously has said that the way they basically test their agent harness stuff is mostly vibes. They’re not really an evals-driven thing.
I think broadly what’s happening with taste in Silicon Valley — it’s a big thing. “Oh, we don’t need someone tweaking the numbers based on an A/B test that has this half-a-percent lift. What’s good taste?” Status has become a lot less about how well you understand the numbers and more of — Rick Rubin became this weird thing. That’s who we all want to be, not Nate Silver. Culturally, that becomes a goal. We don’t want to do the data stuff. We want to have somebody who’s the artist that can conjure the great ideas in their head. Maybe it works, maybe it doesn’t. I’m not saying it’s necessarily better, but that’s at least the ascendant cultural trend.
Michael Chow: It’s so funny you mention Rick Rubin. Maybe this came from one of your posts, but I was just thinking about that. I think he starts an interview with a brief minute of meditating with the interviewer, and I’m like — sure, that’s vibes.
Benn Stancil: He was like, “I just feel it out, the music.” And it’s hard to deny the impact. If you had done that in 2012, everyone would have been like, “oh, it’s a hippie thing. Your intuition’s totally wrong. You’ve got to watch where the users go and look at what they’re actually doing. Don’t tell me what you think they’re doing.” And now it’s like, if you’re Rick Rubin, you sit and have a minute of meditation and you say, “I just intuit it,” and people are like, “oh, that’s the guy.” Again, I don’t know — maybe it’s better, maybe it’s worse, but it’s at least who we want, I suppose.
Michael Chow: Yeah. I’m here for it.
Benn Stancil: I, as a data person, you shouldn’t be here for it. You should be like, what are we doing?
Michael Chow: Wes, do you have any burning AI questions? Because I feel like there’s a million things.
Wes McKinney: I have a lot. I don’t like to describe what I’m doing as vibe coding. Vibe coding has almost become a pejorative term — vibe coding means “I don’t care about the output. If it seems like it works, it works.” I still care about code quality, how fast my test suite runs, how fast the code runs, the long-term sustainability, the growth of the code base. I think a lot of vibe-coded software is going to reach some kind of breaking point if there’s not a reinvestment in architectural quality and design patterns. In a sense, design patterns have become more important than ever — right now it’s an army of agents building giant hairballs of code bases, riddled with bugs. How we reckon with that and manage that — time will tell.
It’s one of the most exciting times in my life in terms of technology. The closest analogy I can think of might be the beginning of the internet, like 1995. I was 10 years old in 1995, and I remember how exciting it was to get my hands on Netscape Navigator and start using the internet. I was too young to do anything meaningful in that era. I didn’t program. I wasn’t going to start a company. I was interested in video games — I did GoldenEye speed-running on the early internet in the late ’90s. But now, 30 years later, it’s this collective fixation on this thing. Everyone’s trying to figure out how to get the most out of it — how to start companies, how to become rich, how to be the first, the best, deliver the tools, interfaces, environments, solutions to make this fundamentally interesting and powerful technology work well for people.
It’s both exciting and fun, and I’m having a lot of fun. But there’s also this weird joyless grind happening where a big part of the tech industry is working harder and longer than they ever have before, and a lot of people are not having any fun. Maybe it’s because there’s so much money sloshing around in the ecosystem. There’s so much money to be made that there’s a lot of young 20-somethings who are like, “yeah, this is my opportunity to work 120 hours a week and have 15 Claude Code sessions going all the time, and then maybe I can become a billionaire if I keep that up for as long as possible.” I think there will be some kind of collective community burnout at some point.
It is very exciting. I’m on board with this new way of working. I don’t think there’s any going back, and I don’t want to go back. But how to actually make it work — how to build software, how to do analytics, how to do data science — who knows. I’m questioning everything, rethinking everything. I’m curious what’s been your experience and how you’ve gotten to know this new technology.
Benn Stancil: Quick aside: GoldenEye doesn’t hold up like you want it to.
Wes McKinney: Today, it really does not hold up. It doesn’t look good. It’s kind of clunky. It’s still charming. It has a special place in my heart.
Benn Stancil: I got one of those Analogue 3D — the FPGA clone of an N64 — and plugged in GoldenEye, played for a little bit, and was like, it had its time in its place. You want it to be “oh my god, this is going to be the best.” You’re like, “ah, this is not so great.”
Okay, so to answer your question — a couple of things on using these tools. The analogy I’ve used with this before is that it feels a little bit like when Google came out. When Google first came out — I was in middle school, high school, something like that — people were like, “oh my god, this is the new thing. You’ve got to learn how to use it. You’ve got to learn all the tricks, all the search booleans, the query languages for Google, how to exclude sites, include these terms, don’t use this.” Obviously, 10 years later, nobody knows how to do that. You just get a feel for Google. You get a feel for the thing. People who grow up with it understand how to search for something. They can find anything, and it’s not because they’re using fancy search parameters. You just kind of get a feel for the machine.
I think that’s what will happen with all of this. There’s a whole bunch of the “internet grifter” type of thing — “here’s how to get the most out of your five Claude Codes” — and it’s like, hey, this will all kind of come together, the best things will emerge, and you’ll just get a feel for it. I think there’s a lot of that.
As for what emerges from it — I don’t know. Somebody will figure out some crazy thing and release a product. Claude Code is an example of this — it kind of came from nowhere. Not really, but it went from being a side thing to “oh, this is the only thing anybody ever talks about.”
To your point about the Silicon Valley burnout stuff — this is a half-written post, we’ll see if I can make this analogy work — earlier this year there was a guy, an engineer, Steve Yegge, who wrote a post called “Gas Town” that was very viral when it came out. Basically he was like, how do I turn my Claude Codes into the most insane thing you could possibly imagine? There’s one mayor thing — Gas Town is a reference to Mad Max, I think? It’s a city in the Mad Max universe. There’s a king, and the king has all these other bots that do things, and it’s agents all the way down — four or five layers to it. He basically says this thing burns enormous amounts of money for a lot of noise. The point is a thousand things do something, and one gets picked up. You tried a thousand times and most of them are trash and a few are good ideas, and only one can actually get chosen because we’re only implementing one thing.
The imagery he uses for it — because it’s this Mad Max thing — is men in tar pits. Not the paradise of floating cars up there. Now we’re all in tar pits. And part of that feels like that’s actually a bit of what the world is right now. To your point, there’s a thousand companies in San Francisco right now; two of them are going to work out — only two can. You can’t have a thousand billion-dollar — it’s not going to happen. There’s going to be a whole bunch of people who are like these agents: the point of their work is to be thrown away. Everybody’s figuring out how to build stuff, and most of those things can’t work because they’re all building the same thing, and one’s going to win.
I think we haven’t really reckoned with that. It’s not like “okay, it’ll be fine, a bunch of people try a bunch of stuff, it won’t work out.” If you go to SF, there’s this sense of “oh my god, we’re all at the beginning of the internet and we’re all going to be enormously rich and famous.” Ultimately that’s a status game. There’s only so many people at the top of the pyramid. Status is zero-sum, and in some sense business success is zero-sum. There’s only so much money people are going to spend. It’s not like everybody can sell — you can’t sell a thousand $10M contracts to Coca-Cola, you can sell like two. One way to look at it is “oh my god, so much potential and creation and possibility.” Another way is “oh, this is sort of a salt mine that feels fun” — but ultimately, a lot of it is going to be industrial waste.
Wes McKinney: My optimistic take — and ultimately, even though there are a lot of questions about the financial private credit bubble, the infrastructure bubble and its effect on the economy, the over-concentration of the US economy in AI, essentially the S&P 500 and NASDAQ have become an AI play. If you’re in the US stock market, you’re essentially long AI whether you want to be or not. Your index funds, your 401k — all of it.
But I am also, on the technology side, an optimist. What I’ve been telling people is that initially there were fears of AI is going to replace software engineers, and we’re going to build the same amount of software; it’s just going to be built by AI. But actually no — we’re going to build 10 to 100 times as much software, and yes, it’s going to be built by AI, but there’s still going to be a lot of professional software engineers.
To your point, if we build 10 to 100 times as much software, who’s going to buy it? How much money is there going to be to buy software? One of the memes going around Silicon Valley lately is “software as a service is dead,” which I think is a little hyperbolic. Some SaaS products are hard to build; some turn out to not be. There are a lot of to-do list apps — I’d say to-do list apps are probably in trouble, especially because people can build their own. I’ve built my own to-do list app with Claude Code. Why not? But certain classes of software will become very difficult to monetize as there are free open-source versions or you can roll your own.
Right now you still need to be a pretty good software engineer to get good output out of Claude Code. But over time, maybe a year or two from now, you’ll be able to pull up your phone, voice interface, say “make me a new to-do list app that has these characteristics,” come back a day or two later, and the agents will have built your iOS or Android app and installed it on your phone. Nobody really knows what that compression of the software industry looks like — so much more software fitting into effectively the same amount of space. The population’s only growing so fast. Maybe it will increase GDP, world GDP — but not by 10 to 100 times, not even by a little bit.
Benn Stancil: Even if it does, ultimately, most people want to be rich because they want to be richer than their neighbor, not because they care about how much money is in their actual bank. A lot of this is just status. But to me, the easiest analogy for it is content creation. It used to be to make a movie, you went to Hollywood and paid a bunch of people to do whatever. Now anybody can do it. There are a lot more people making movies. There are a lot more movies. We only have so much time. We’ve given a lot more time to staring at TikTok than we did when we had to go to movie theaters.
But it’s not like everybody now is making — there’s a whole bunch of people on the cusp of sanity trying to become TikTok influencers. For some people it’s been great. But there are a lot of people on the fringe of that, and it seems kind of bad. It seems very difficult if you’re there. There’s a lot of stuff on that edge of “yeah, you’ve got enough to maybe make a thing, but not really.” That feels closer to what the software stuff will be.
Michael Chow: Could you say more about people on the edge? What type of people are you thinking of?
Benn Stancil: When we started Mode, starting a software company was a serious endeavor. Not serious in the sense that we knew what we were doing — but you had to go out and raise money, spend a bunch of time building a thing. From the day we started the company to the day we first put out the first product that anybody outside of it ever saw was close to a year. And that was seen as reasonably quick. We made a product that was pretty simple. For us to build anything that had any kind of meaningful feature set, it was going to take several years, millions of dollars spent on engineers. It was very hard to get something that was sellable at all to market without spending a bunch of money and a bunch of time. That didn’t mean everybody who did it was therefore a real professional, but there weren’t that many people doing it. If you did, you spent a lot of time thinking about it.
The same was true for content. You want to write a book? Great. It used to be: “I’ve got to sit down and write a book. It’s going to take a long time. You may be a bad book, you may be a bad author, but it takes a long time.” Even if you’re going to write some bad beach read, you’ve got to write every word, and that takes time.
Now you can write a book in 10 minutes. You can create shippable software in two days. That doesn’t mean “oh my god, the bar is that much higher.” But to your point, we can’t read every — if everybody writes a book in 10 minutes, we can’t read that many books. Those books aren’t going to get read. It’s not possible to consume that much content or read that many books or use that much software. And because you can create a thing that you could have created 10 years ago in two days, whereas it would have taken you two years before, the market isn’t growing in the same way. It may be growing a lot, may be doubling — I don’t know. But when you can create 10 to 100 times as much? What’s the market for video? It’s way more than it was 10 years ago, but there’s probably a thousand times more video people are creating now than they did 10 years ago.
The people on the edge are a whole lot of people who try to make that a career, but they’re making stuff that seems pretty good, that looks like it should be quality professional, that is good enough — but when there’s that much, a lot of it just doesn’t catch. The market is extremely saturated. I think software will be like that.
Michael Chow: It’s a funny issue. I feel like authors — yeah, if a person spends a year of their life on a book, we’re all going to die one day. For a person to spend a year of their life on a book makes me think, oh, that’s kind of remarkable that a person would dedicate a big chunk of time to this thing. I feel like it makes me want to talk to the person, understand what pulled them into it. But I could see if you spend like an hour vibe-coding a book — there’s something much less compelling about that. Maybe it’s great, but —
Benn Stancil: Presumably the other part of this is: if you spent an hour making a book, somebody else also spent two years. If you’re picking up a book to read, which one are you going to read? They have the same tools you do. If I’m going to build software in two days, you’re going to build the same thing in two years — yours is probably a lot better. It’d be weird if it wasn’t. If we get to the point where it’s not — where I can make something in two days that I can’t make a better version of in two years — then God knows what. That’s a weird world, hard to see around that corner. But if you spend two years on it, it’s going to get way better than any version somebody built in two days.
Wes McKinney: Yeah. I think now, if you do hardcore agentic engineering and spend a year working on one project, certainly it’s going to be way better than somebody who tries to vibe-code a replacement in a weekend, or the same year that you spent pre-AI. You would make a great deal more progress. It might have taken five or 10 years to do the same amount of work in the past. But what’s interesting right now is this feeling of — in the past, you had to be very deliberative about where as a software engineer you would choose to spend your time. What projects do you build?
For me, somebody who really likes building software — up until now, I really enjoyed writing code, but now I’m actively enjoying not writing code. Which is weird. At the same time, I realized at some point that I had a whole mental backlog of projects I thought of building over the last 20 years that I just never had time for. I would always have the same conversation in my head: “wouldn’t it be nice if I had the time to do that, but I can’t, because my time is valuable, I need to get paid, I need to pay my rent. I’ve got a plan for the future, I want to be able to retire.” So I would choose to spend time on software projects that serve my professional goals, working toward building something I could release out into the open-source ecosystem — something that would have value and impact for other people.
Now all of a sudden, I’ve got that mental backlog of all these orphaned project ideas that I just never had the time or inclination to work on because I have a life outside of work. I like being able to close the laptop and cook, or go on a vacation.
Benn Stancil: Do you like being able to close the laptop now?
Wes McKinney: No. And that’s becoming a bit of a problem. Now when I close the laptop, I feel guilty. I’m like, the computer could be doing things right now. That definitely is happening. But it is an exhilarating feeling — the ability now to revisit all of those old ideas. Especially because I can have a terminal tab that’s just working on that side project I thought of eight years ago. I can build the thing I always wanted, and it doesn’t require me to nudge it along a little bit every day. Every day I have a few more ideas, I nudge along the agent, and I keep grinding on it. I’ll continue to release things — some things are only for me, some things are all open source. But it’s nice, sitting there watching Netflix at home and just nudging along my little agents working on my personal side projects, clearing out the backlog of stuff that I dreamed about and never would have gotten built without AI. I’m having fun doing that.
Benn Stancil: That makes me think of a kind of interesting thing: how long does that feeling last? Or how far can it go? You saw this happen with ChatGPT when it first came out, where people would ask it dumb questions — “rewrite this email in the style of whatever.” It had this fun little gimmick. At some point it was like, “okay, I know what it’s going to do. It’s going to produce the thing. The artifact isn’t that fun. It was sort of the process of doing it that was fun, not the artifact itself.”
To your point, if you have all these things churning away in your backlog — your project backlog, your idea backlog — it’ll be cleared out eventually. It’s partly that, but it’s also — if you just said, “let me give you a two-paragraph spec of the thing I want, push one button, and it’s done” — do you think that’s fun? Is it fun if it can really do the whole thing autonomously? If RoboRev was done in one shot — if you could one-shot RoboRev — would you have fun building it? I don’t think so.
Wes McKinney: Yeah. Part of the fun is the process — seeing the thing take shape. I’m not an artist, but I imagine like what it might have seemed to Leonardo da Vinci to be chiseling David out of the block of marble. Maybe that’s a bit of a grandiose analogy, but agentic coding feels like that. The feedback I’m giving the agent, I can see the project taking shape. The agent’s behavior out of the box — it gets most things wrong. You have to go through and find all the things it got wrong, get them right, refine and refine and refine. It takes hundreds or thousands of iterations with the agent to make a thing that meets my standards.
I have high standards for tools. Maybe for other people it would be different. Maybe their one-shot tool would be totally fine for them — if it’s a piece of throwaway code that serves a one-time use. But if it’s a tool that you’re going to be using for years, the way I’m approaching it is: I’m building things that I want to be using five years from now. So I might as well put in the effort to do it the right way.
I don’t know — it’s fun. The feeling won’t last forever. Maybe a year from now, six months from now, it’ll just feel like work again, and at six or seven o’clock I’ll be happy to close the laptop and say, “the day is done.” The agents will be there tomorrow morning.
Michael Chow: It’s such an interesting point — the question of “do you like closing your laptop? Are you having fun? Would you have fun if you could snap it into existence?” It feels like questions of “what do engineers enjoy, versus a person who might contract an engineer to build something they want — who enjoys being in the process and maybe seeing the process sped up or automated, and who just wants it done.” Tools like RoboRev that you’re talking about — I guess I should explain what that is — a tool for reviewing commits, that augments the development process with AI. It is kind of like an engineer’s tool — something they enjoy having, something that speeds up the process. But there’s this whole other group of people who probably want to snap things into existence.
Benn Stancil: Even something like that — I mean, you built it, you can tell us — but even if you could snap your fingers and have the problem that tool solves solved, I’m not sure that gives you the same satisfaction. There is something about building a thing and using the thing and it being my thing. You see the spinning wheels, see it come back with stuff — there’s enjoyment in using it. It’s not all just “how do I maximize some utility of this thing that does my particular job?” If you have a button that fills out your tax forms, okay, that may be a different thing. But there’s a very different set of things: “I just do not want to ever see this again” versus “I built this partly because it’s a thing I like to use.” A lot of people do woodworking. They’re not like “snap my fingers and have the thing exist” — they want to do the thing. I think there’s still some of that here.
Wes McKinney: Yeah. For people listening, the thing I started building — RoboRev — as part of the agentic development side projects: essentially, like many people, I started running multiple Claude Code sessions, and I found that most of the work Claude Code was doing was full of bugs. So I started running additional Claude Code or Codex-from-OpenAI sessions whose entire purpose was to review the work of the other Claude Code sessions. I would have three Claude Code sessions and then three reviewers, and all they were doing was “review this commit, review this commit, review this commit.”
Claude Code would return control to me, commit its work, and then I would say, “here’s the commit hash,” and I would give it to the reviewer session. This went on for a while, and I was like, I’m losing time because I have to manually take the commit hash and give it to the reviewer — that’s wasted. The whole thing just didn’t seem efficient. So I said to myself, this whole process can be automated. RoboRev basically automatically triggers a code review whenever the agents return control to you — whenever they commit and return control. I found that really liberating. I didn’t have to run those extra Codex or Claude Code sessions anymore.
If I have parallel multiple code sessions, whenever I return to the session from another side project or another git worktree in the same project, not only has the agent finished what it was doing, but also the associated code review has been done asynchronously while I was working on something else. If I forget what I was doing, or I’m not sure what to do next, it’s just easy to say, “hey Claude, fix these bugs that were discovered in the work you just did.”
Part of the fun of it was making a tool, seeing yourself build the tool, seeing it work. A side corollary of that now is that a lot of people are going to prefer to use tools of their own making. I find that whenever I show new AI tools — or new tools I’ve created — to other people, usually the response is something like, “I’m exhausted with how many tools I’m being shown, how many new tools I’m seeing on the internet. The only tool I can manage right now is a single agent session, or maybe a couple of agent sessions. Any other tools beyond that is almost exhausting.”
I’ve also felt that exhaustion — the head-spinning feeling of 10 new AI tools and systems and solutions coming out seemingly every day. For the longest time, I was tuning out a lot of what was going on in LLM world. The only thing that really pulled me in were the CLI coding agents. Prior to that, I was, to be honest, not impressed.
Benn Stancil: That reaction — being shown 10 new tools — is the content thing to me. It’s like if iPhones and iPhone video editing had just come out: “look at my video.” “Oh my god, this video is very good. It’s as good as any video I’ve seen 10 years ago. But I can’t look at this much video. I’ve got to tune it out.” There’s this feverish pace to the whole thing where all these people make things that seem really good, but I can’t keep up.
Michael Chow: Benn, I know we’ve been down the AI rabbit hole for a while. One thing I’m very curious about — when I asked you, before doing this, what you’re most excited about, you said “joining a boy band.” I wonder if you could tell us more about your boy band dreams. It does seem like a fun gig.
Benn Stancil: Basically — it’s a model of collaboration. So is your Gas Town — probably not a bad boy band name, to be entirely honest. If you ever interview an exec, especially an exec from Salesforce — this is apparently famously a thing that Salesforce did — you ask them what they’re good at, and they’d say stuff like, “they’re incredibly collaborative, just the most collaborative, I’m just so good at working with them.” Great.
One of the things I also like is — I’m not like “oh, I’m collaborative.” I don’t really want to work with that many people. It’s a lot. You want to work with some people, but not an army of people. I don’t want to work with the whole Gas Town army. I think it’s more fun to work with a small group of people who are all committed to a project and want to do the thing.
It’s basically a model of collaboration, like a boy band. There’s a handful of us. We all have our fun little personalities. We do our own things. You could do your own thing if you wanted to, but you kind of like being a part of this thing. Sometimes you do a little bit of your own thing, but ultimately we all have to do the same song and dance, and you can maintain some independence. Each person in the boy band isn’t overseeing enormous departments. Their job isn’t managing a ton of people. The job is: “I make the things.”
Historically, one of the downsides of startups is in some ways you get punished if you’re successful, because you have to go hire a bunch of people, and you become the “oh, I have to manage tons of people.” Sometimes there are things about that that are fun, but a lot of it is administrative — to sort of downplay it too much — but it’s organizational work. How do you make sure everything is aligned? How do you make sure this department sends to this department? What happens when they disagree? How do you keep everything on the rails? I don’t particularly enjoy that work, but that’s what you have to do. The goal of building a company is to get to the point where you have to do that work. Otherwise, what have you done?
That may not be the case now, that you can do a lot more with fewer people. You can keep it relatively small. It’s just a handful of people working on an exciting thing. You can make — maybe not an enormous business, maybe not the next trillion-dollar thing — but a functional, profitable, successful business by all sorts of reasonable definitions with a relatively small group of people. At least, that is my hypothesis.
Michael Chow: And not to stretch the analogy too far, but in your dream data work or boy band scenario, what role are you? Are you the bad boy? The father figure? Is there a father figure of a boy band?
Benn Stancil: Not in the modern boy band. I feel like in the OG boy band there was always a responsible party — the one driving them home.
Michael Chow: The responsible one.
Benn Stancil: I think I would be more of the moody guy in the back. You don’t know that much about him. “What’s his backstory?” It turns out it’s nothing — there’s nothing interesting there, but he’s doing a lot of skulking around. Every once in a while he emerges and has his verse, and people are like, “oh my god, he never told me that.” I would be more that. Not trying to be the guy in the front.
Michael Chow: That’s fair. It is interesting — when a boy band succeeds, people aren’t saying, “we need to add 200 people to this stage.” Though there are some K-pop bands.
Benn Stancil: Yeah, right.
Michael Chow: But it’s not like you’ve got to scale the whole thing up. I think that seems more fun. Are you forming any boy bands right now? What do you have cooking? Or is this kind of a metaphorical boy band?
Benn Stancil: A metaphorical boy band. There are some ideas of the metaphorical boy bands. If I could snap my fingers, like Matrix-y, “I know Kung Fu” kind of one skill, it would definitely be boy band skills. Being a pop star seems like a really good time.
Michael Chow: If you get the chance, I think you should take it, 100%.
Benn Stancil: Oh, I would. But don’t hold your breath.
Michael Chow: What about other projects — are you working on projects right now? And have you succeeded in keeping the essence of a boy band in them?
Benn Stancil: It’s just me. It’s not really a boy band. It’s kind of a bad solo act. I like playing around with some stuff — similar to Wes, except I would say with less success. You have a bunch of ideas, you want to play with them. There are things that you think are interesting to experiment with. A lot of them are less like “here is a very precise problem I have,” and more like “this is a rough way of working or doing a thing that I feel like I could, I feel like there’s got to be a way to sort of solve this.”
So I periodically yell at the internet. There are parts of that process I really like. There are parts that are kind of a grind. I’m not sure how much of that you can AI away. There’s a delicacy there too — I don’t want it to be “push button, get blog posts.” That is not a thing I want to do. However, there are a lot of parts of that drafting process that you’re just like, “this is painful.” There are things you can do to help unstick you, where it’s like — the things that I do, I want to try to find ways around. So we’ll see.
Michael Chow: Just to get a little more concrete — you mentioned Wes, are you saying using like Claude Code for writing, getting assistance? What kind of stuff are you cooking on?
Benn Stancil: I basically use AI to write stuff in one way, which is as a thesaurus. It’s a pretty good thesaurus because you can be like, “here’s the word, but make it more X.” You can kind of guide it a little bit. Obviously that’s not a whole lot done, it’s useful sometimes, but that’s basically the only way I’ve found it useful. The thing it writes — I do not want to ever say a single word that it ever writes. Certainly you try periodically, but you try this and you’re like, “oh my god, I cannot.”
But there are processes in writing. There are points in the process of writing these posts where there is something akin to a thesaurus, for a broader set of problems. You have a bunch of loose ideas in a doc, and you’re like, “I don’t really see how to weave these things together. I don’t really know what the narrative here is. I kind of like this sentence that I wrote here, and I like this idea, and I think it needs to work here, but my god, it’s a mess.” That is the stressful point of writing these things, because you’re not just refining paragraphs. You have to be like, “does this fit together at all? Do I have puzzle pieces that can make a thing that looks like anything?”
There are ways you can attempt to mimic the creative process that helps solve that for me. That is again very much not “write that paragraph,” because I think it does a terrible job of that. People have written a lot of stuff about this. I cannot beat certain patterns out of it. But I do think it can help be like, “here are some ideas for ways to arrange things.” Most of them aren’t any good, but there is some inspiration you could be like, “oh, that’s kind of an interesting thought. Let’s play with that.” That is basically the equivalent of having a conversation with someone about it.
A lot of these posts end up getting written because you have these loose ideas, and you sit down and talk to somebody, and you’re like, “oh, I didn’t think about connecting in that.” They don’t tell you to do it. It’s just that they say something that triggers a thing in your head that connects two ideas in ways you hadn’t seen.
Can you use it in that way? The analogy I have — a friend has used this with me — is: have you ever played the game Codenames?
Michael Chow: Yes, I have.
Benn Stancil: You get random words that you have to find associations between. That process is where interesting things come up, because it’s not “here’s a word, find three related words that are very similar to it.” It’s “here are two very disjoint ideas, forcing yourself to be like, how might these be connected?” Something kind of interesting comes out of that.
That is a process that right now has to sort of happen serendipitously, because you have to walk around the world and do that. There are ways you could use AI as a conversation partner of sorts that isn’t just ChatGPT — it needs to be more guided than that. Basically, can you use it to find those serendipitous moments faster?
Michael Chow: It’s so interesting — hearing you talk about it, it sounds like you’re browsing the space of possible connections just to see what’s out there, and is there anything useful, but it’s pretty broad. Like storyboarding.
Benn Stancil: This is a bit of a weird example, but there was a thing I wrote a while back — I don’t know, a month or so ago — that was about the addictiveness of AI stuff. It wasn’t just “oh my god, ChatGPT psychosis” or whatever. It was kind of about how this feels like a certain kind of thing where we are creating our own realities in certain ways. You see that extending in other places. I don’t remember the exact news hook.
A couple of nights before, I went to a Lorde concert. There are a couple of songs where you’re sitting there listening to a Lorde concert and you’re like, “oh, that’s kind of an interesting thing.” You’re watching all these teenagers totally dissociate from the concert so they can take pictures for Instagram and not pay attention. And then Lorde has an album called Pure Heroine — “heroine” spelled like “hero,” female hero — but it’s also like AI is kind of pure heroin in that way.
Seeing that connection actually ends up driving a lot of the narrative. It’s not like, “oh, I have this exact article I write, oh, I can tack on this analogy.” It’s “I have all these loose ideas, and then you see that and you’re like, oh, there’s actually an interesting thread. This is the thread that can tie it together.” But you need to be sitting at that concert to see the thing happen. It requires these random collisions. Then one of the random collisions is like, “oh, that’s where the interesting connections are.”
For me, anyway, that’s where a lot of attempted creativity comes from. Can you speed up that process? Can you force it a little bit more? Code names is doing exactly that — you become very creative playing that game because what else are you going to do? There may be ways to say, “help me run more threads through this to see what happens.” It’s not going to write anything, but it will help you see more interesting things.
Michael Chow: It sounds like it’s not a rhyming dictionary. It’s not a thesaurus. It’s a little more granular than that, but it’s still not writing the thing. It’s just a different way of exploring the associations or paths.
Benn Stancil: Part of it, to me, is: people have asked about these blog posts before, and a lot of them have things like — why is Lorde the hook for all this? How did you come up with that? The answer is: you cheat. The answer is: you didn’t come up with that. You came up with one association with Lorde, and then you read some Lorde lyrics and you found four others that felt kind of attached, and then you hook other ideas onto them. So it looks like, “oh my god, you had all these ideas and you found that Lorde is the thing that perfectly weaved through them.” No — I started with a Lorde song and then figured out ways to tack on things that feel associated. It creates a much tighter sense of connection than you kind of should, because it seems like you found this analogy that maps so well, when in reality you’ve stuck ideas onto the analogy to make the analogy work. The analogy was the starting thing, and the ideas were glued onto that.
To me, the process for writing this stuff is not “what are the ideas, stick them on the Lorde analogy.” The Lorde analogy is actually the thing that drives it. Forcing yourself to describe everything through this analogy forces you to come up with more creative ways to say it. Whereas if you’re not doing that, you end up just writing a boring outline and then sticking on weird forced, half-connected jokes.
Michael Chow: It’s so interesting to hear the process. I’m curious — do you have literary influences? Are there people you love to read who have influenced your style?
Benn Stancil: Two things. Matt Levine’s stuff is great. He has a style — it’s very much a thing that you can’t emulate, because everybody sort of knows it, you can’t try to write like him anyway, he’s very good at it. The thing I really appreciate about him is that his thing is kind of about nothing. What’s the point of this? I don’t know. There are a bunch of interesting things in the world to look at, and I like to look at them.
I read his stuff because I find it enjoyable. No shade to Ben Thompson — Ben Thompson is obviously very popular. Ben Thompson writes about a subject that I should care a ton about, and I find his posts hard to get through. I read him because I think he has very good ideas, but it’s “this is fascinating, the structure of it is, and the ideas he has are” — but I don’t particularly enjoy reading it. Matt Levine makes me care about stuff I should not care about. I’m just interested in it because he tells a good story with it. I don’t care about any of the topics at all. The private credit bubble — how much do I care about that? Not really. But when he tells a fun story, it’s fun. I appreciate that style.
The other thing — actually this is where a lot of this originally came from — there was a guy, I don’t remember his name, he’s a chemist. He won a Nobel Prize in chemistry for Bucky balls. There’s some compound that is a giant hexagonal thing. He won a Nobel Prize for it, and it turns out to be completely useless. It was just a very stable ball. Clearly didn’t stick.
When I was in college, I went to some lunch thing that he gave a talk at. He gave the talk in this style where it was probably a 20-minute talk and he had 300 slides. You just can’t look away, because it’s slide, slide, slide, slide. Each slide has a sub-narrative to it. He’s giving a talk that you don’t really need the slides for — it’s not “and now we’ve got slide six and here are the bullets and here’s the diagram.” He’ll say something and it’ll be an image that sort of is related to it. It ends up creating two stories — there’s a story behind it that’s the entertainment, and the thing he’s saying. It keeps you really attached.
I started giving talks in that style. And the thing that creates is a certain pace to it — it requires you to weave a bunch of stuff in, because you can’t just be “AI, picture of AI. Computer, picture.” You have to come up with little vignettes that follow along throughout the course of it. I think it’s much more entertaining. A lot of the blog style actually was like, how does that look on paper? I’m not sure it delivers on that, but that’s where there are footnotes and stuff that are like — the whole thing is a little bit of “there’s a sub-narrative to it that’s me entertaining myself.” Because again, it’s about SQL, my god.
Michael Chow: It is interesting. It’s all the elements of stream of thought and collage and association, but it’s crazy that it all comes back to data. It’s not to butter you up too hard, but it’s interesting to read.
Benn Stancil: To some degree comes back — sometimes they don’t really tie together.
Michael Chow: Are you saying there’s no promise you might at some point drift off data and go full…?
Benn Stancil: Broadly, I would say, I at this point mostly have. There are still things there, but it’s not a lot about ETL tools these days.
Wes McKinney: On the subject of small teams and small companies — I feel like examples of companies that have remained small and yet been really successful, there just aren’t that many. Craigslist is always the canonical example of a company that still has the same very bland utilitarian website that it had 20 or 30 years ago. I don’t know when it started — over 20 years ago. And yet they have tons of revenue and a really small team.
Me as a software developer — I always found myself torn between the desire to build open-source software and make the world a better place through building tools and giving them away for free on the internet, with the downside that it’s hard to do that in a sustainable way. It’s hard to get people to pay you to do that for a long period of time. It’s hard to be able to hire people, because people need salaries, they need health insurance. On the other end, entrepreneurship and starting companies — invariably, with the way the world works, you get nudged into the venture capital and raising money, and that has a lot of strings attached. There are a lot of expectations in terms of how fast you’re going to grow and how many people you’re going to hire. That gets really complicated, as you know.
One optimistic hope I have is that there might be some new model of people that build software in the world and can create things that have impact that’s neither of those things, and yet can be fun and sustainable — kind of the best of both worlds. I don’t know what that looks like, but it’s something I’m really interested in.
Benn Stancil: My answer to that is: it looks like a business. It’s a small business. Only in software — and this made sense, kind of the economics of building software, there’s a ton of upfront costs — only in software would “we’re going to start a thing, and it’s going to be 10 people, and we’re going to try to make more money than we spent” sound weird. That’s the goal. You start a restaurant? Duh, yes, that’s the goal. But in software, it’s “that’s a ridiculous goal. How are you going to grow?” As famously parodied in Silicon Valley: “revenue? No, no, you want to be pre-revenue.”
There’s very much a model for a handful of people working on a fun thing forever that you can be sustainable. That’s a business. But startup — it’s not just the VC ecosystem; the economics of software has not been able to support that. You do have to spend $5 million before you can sell anything. You’re in a lot of debt, and that doesn’t quite work. But now you don’t have to be. You can probably be making money after a year of one person working on a thing. All right, that’s just a business. If building stuff is fun and a business you can run forever isn’t bad, then not a bad gig, maybe.
I don’t know. Maybe it doesn’t work, maybe software is too winner-take-all, so you have to get super huge. But I’m not that cynical about that. How many people do you really need to really love a thing to make it sustainable for a handful of people? Not that many.
Wes McKinney: It does seem like there’s a big space for a small, useful thing that people are willing to pay monthly for or something.
Benn Stancil: And it — this isn’t quite right because there are certainly a lot of things that are pure productivity tools — but it starts to feel a little bit more art-oriented, or content-oriented. It’s almost like Substack, but like Substack for somebody’s vibe-coded software.
One of the things I’ve noticed is that people seem to have replaced, in a lot of ways, decks with Vercel apps, because you can do it in the same amount of time. “Oh, you want to give a demo?” “I just made a thing.” You get random pitches from people — “check out this thing, check out this whatever.” I have gotten a lot of those that are links to Vercel apps now, because it used to be “here’s a PDF of some slides.” Now it’s “why would I do that? I can make a little clicky app thing.”
It feels more like software is just content at that point, as opposed to a tool for anything. I just have a much more flexible canvas to make a thing I want to make.
Wes McKinney: It’s interesting — the idea of a deck being replaced with a Vercel app, which is like a deck plus just a little bit of interactivity. The bar is just way lower.
Benn Stancil: It’s not that different from “what’s the simplest version for me to explain something?” A bunch of text. Okay, what’s the slightly more complicated version? A Word doc with text, some bullets, some nice formatting. What’s a very flexible version? A founder’s deck — well, it’s a deck. How do founders get rejected? Most of the time, it’s a quote not a squirt.
Like, every VC — you get rejected probably for emotional reasons, you get probably rejected because literally two minutes into the pitch they don’t take you seriously and they’re tuned out and off looking at something else. Maybe they reject you for real material reasons — they really evaluated and didn’t like it — but they’re not going to give you that memo of what they said. They want you to like them, so you come back next time if you are successful. They want to maintain optionality.
It’s very frustrating when you’re pitching, because you don’t get anything honest. You get a lot of soft “it’s just not for us yet, we’d like to see a little more meat on the bone.” And you’re like, okay. I am not a VC, nor do I give feedback that people should listen to, but I would appreciate a VC where you walk into it and it’s a little bit of the Simon Cowell of VCs. They’re going to hate it. I’m not going to be offended when they hate it, because I know they hate it. And if they like it — Simon Cowell liked my song — oh my god. I would pitch that VC, because the worst that happens is the thing I expect to happen, and there are a million others I’ll just go pitch for the actual money.
VCs have weird value systems. There are some weird incentives there. “What do you really think” is nice to hear in an uncomfortable way sometimes. Again, if Simon Cowell tells you you’re bad, you’re not going to be like, “I must suck.” You’re going to be like, “that’s his job. That’s what he does. That’s his shtick.”
Wes McKinney: You’ve pitched a lot of VCs. Maybe you have — no, it’s — I largely agree with what you’re saying. Especially in Silicon Valley, it’s rare to hear honest feedback. I’ve known VCs and have been in pitches where I’ve received honest feedback — “this is not good. You shouldn’t be doing this. You should be doing something else. The way you’re approaching this problem is wrong.” It is refreshing to get those takes. But usually, professional venture investors are looking at so many pitches and talking to so many founders that they’re just pattern-matching. They don’t have the time or energy, especially at the early stage, to think super hard about things.
It’s basically “have I seen any similar companies that have done this successfully? Does the way this product is being sold make sense? Are they selling to the right people? Does the tech, the team, seem credible?” Surprisingly often, you look at the backgrounds of the founders of a company and it’s “they can’t build this. If there’s going to be a successful product, it shouldn’t be this team.” That’s often the right answer, and it isn’t always the most — unfortunately, very frustratingly — it’s not the most competent and qualified founders who succeed in the space. Often the founders who were there at the right place at the right time came in second or fifth, and they’re very frustrated and bitter. But that’s how the cookie crumbles.
It’s a lot of pattern matching because ultimately, in investing, you want to build or maintain relationships, not burn bridges. You don’t want somebody to come away from an interaction being offended or being like, “that guy was a jerk. He was honest, but he was a jerk.”
Benn Stancil: You should never be a jerk, and to me — I feel like a lot of times VCs don’t give totally truthful feedback because they’re trying to be nice. A lot of times the answer is “you’re not the horse I want to bet on,” and I don’t know what use is telling somebody that. But as a result of that, you also just get very soft answers.
Wes McKinney: I don’t know if you want more of these, but you’re giving people candid feedback.
Benn Stancil: Yeah. I’m certainly not trying to maintain some sort of set of deal flow around that. But also, I don’t know — that’s what I would want. If you’re nice about it, it’s fine. I try to be. For me, that stuff’s interesting. It’s interesting to see what other people are going on.
Michael Chow: Anything you’re excited for this year? What are you most excited for?
Benn Stancil: This year? There’s a bunch of stuff. The way all this stuff changes things is exciting in a very — maybe we could just have a down year. Maybe we could just stop for a bit.
Wes McKinney: It’s all gas, no brakes.
Benn Stancil: We need a moment to catch up. Can we just all — I don’t know. I’m sure there are things that will come out that will be fun. Claude Code is a good example of that. You use it and you’re like, “this is pretty fun. I can do a lot of stuff. This is cool.” And yet there’s also half of you that’s like, “I am so behind. What do I need to be doing? Stuff all the time. Am I keeping up? Is everybody else doing the same thing? All of the ideas I had — are they all going to be built a hundred times over?” When things are moving that fast, it is both exciting but also terrifying.
There will inevitably be more moments like that, where something happens and it’s just — there is an opportunity, but there are also so many opportunities that you will miss because stuff is moving so fast. Exciting but somewhat existentially stressful.
Michael Chow: “Existential” has come up a couple times. Feelings of uncertainty, which is so interesting. It’s such a different response than “it’s a really exciting technology,” and there is this slight existential —
Wes McKinney: I spent a lot of the last — honestly, very honestly — I spent a lot of 2025 in a state of essentially existential dread about what it means to be a software engineer. Benn and I are a little different. I’m more of — writing C++ code, writing Python code, this is an engineer.
Benn Stancil: Say it.
Wes McKinney: The rapid pace of AI development was, for me, disruptive to my core identity as a person who has spent a lot of time becoming good at software engineering, getting really good at writing code. This feeling of essentially yanking the keyboard — the proverbial keyboard — out of my hands: “you don’t need to write code anymore.” There’s this feeling of “well, what am I for? What’s my purpose here? Where can I add value?”
I think at some point it clicked, and I regained my agency, and started having — as Peter Steingard puts it — “I can just do things.” So I’m a little bit more embracing this mentality of just doing things again and not worrying too much about whether other people are doing more, or other people are working smarter, and just focusing on: am I building things that are useful, things that people care about? Maybe they’re only things that I care about, and maybe that’s okay.
That existential dread has been present in a lot of the industry. I don’t know if there was any one thing that caused me to break out of it. I remember there were weeks where I would sit down and start writing code in Emacs or VS Code, and I’d feel so ineffective. I’ve gotten over that. Maybe that’s just a mourning process for the old world that feels like it’s gone. I don’t know if you felt that way.
Benn Stancil: Oh, yeah. And the “just do things” is a useful reminder sometimes. A lot of the stuff I do is this kind of blog stuff. The process of writing these things, or putting together presentations — it isn’t just “I’m going to methodically plod through something.” There is some discovery to it. You spend a lot of time wasting your time. There’s a lot of work that goes into it where you’re like, “that was bad, I’ve got to throw it away,” and that’s part of it. You can’t skip that.
The “just doing things” — it’s not to me “everything is happening so much, everybody’s going to create the new iPhone app that’s going to make a zillion dollars, I have to go create a new iPhone app and I’ll make a zillion dollars.” It’s that doing the thing is the process — how you figure out what to do, what the actual thing to do is. You can’t — you’re trying a bunch of stuff, maybe some of them will work, maybe they won’t, but the point is, whatever you find that does work, that’s the way to find something that does work: by doing what you’re doing.
I view it as — are any of these ideas that I play with any good? Probably not. But in two years, I’ll look back and be like “I may have come up with a bunch of terrible ideas, but I tried something. Oh well.” Or I’ll look back and be like, “yeah, it took seven failed things to find the eighth that was any good.” All you really can do is just do the first one and see what happens.
The part of that, though, that I think is particularly important as relates to startups, and is true with any kind of creative stuff — you have to be willing to move on. That’s one of the things about not having to raise a bunch of money that is helpful. You don’t have to go out and “I’ve spent six months researching a problem, I raised a bunch of money to solve this exact thing, we’re going to hire a team of people who want to work on this thing, now suddenly we have 18 months of time we’ve spent on a company to solve a problem, we’ve sort of half pivoted a couple times because this is the idea we have, and we built the tech around it, and you can’t get away from it.” You have this anchor that you’ve attached yourself to.
In a sense, you couldn’t just do one thing. You couldn’t do a bunch of stuff to figure out what’s interesting or what might work. Now you can. You can have five projects working at once, and maybe you realize one is a great utility you want to use, and one’s actually a thing that people want to buy, and one’s a thing that people want to buy but you hate it, so you don’t want to do it. Maybe you realize eight of them are all terrible throwaway things, but in the course of that, you found these other things you’re interested in. You can do a lot more of that.
My optimistic view of how you find your way through this is more of that. You can just do stuff, and then the way you’ll figure out what’s actually exciting is by doing that.
Michael Chow: Super interesting to hear. Maybe to tie it back to your way earlier point of “in the long term, everything’s a fad,” and this idea of more vibe-oriented feel — I hate to use the word “vibe” so much — it’s a really interesting point that people today, if someone’s worried about what to do now, it’s okay. You can do six or seven things really easily and explore what’s out there.
I never thought of this analogy before, but it’s like — I went to a liberal arts school. A lot of people show up at liberal arts schools not knowing what they’re going to major in, and the way you figure out is you take classes, you try a bunch of stuff, and you’re like, “that was really interesting, I want to keep doing that.”
Benn Stancil: It wasn’t an engineering school, or a lot of schools where you have to declare a major up front. If you go to one of those schools — I think European universities might be like this, they’re basically trade schools. “You want to be a nurse, you start out, you’re locked in, you go from high school to nursing college.” I don’t know how it works.
Michael Chow: “Uni.”
Benn Stancil: I don’t know what anybody is talking about, but you have to make that choice up front. Again, you can just do it — you can do whatever you want. But if you go to some liberal arts school, you can do whatever. Don’t worry about what you want to major in — try it. You’ll have a lot better sense in a year when you’ve taken 10 classes. I think it’s kind of like that. Just try it. How could you possibly know until you try it?
Michael Chow: Yeah, try a bunch. That’s probably super helpful advice today for people thinking about the future, the pace of technology and AI, wondering what to do — that they can just try things.
Wes McKinney: Do you reflect much on thinking about the analytics work that you used to do a lot more of in the past — the data analytics team at Yammer, the early kind of BI dashboarding world of the 2010s with Mode? If you were in a business and had to solve those problems, what do you think it looks like now? Is that space still something that interests you?
Benn Stancil: It’s kind of a big question mark — the fate of that whole segment of the industry. Even the whole concept of business intelligence is a little bit unclear. Everyone wants to make their own custom dashboards now. Why should you be forced to use somebody else’s dashboard builder interface? It’s funny to think about remaking these fields that used to be well-defined, with a set of tools that we’d use and an established set of practices, being — not totally thrown out the window, but maybe on their way to… I don’t know.
Do I want to work in that stuff again? Not really. There are still a couple of things — “would be kind of nice if this particular thing existed.” My guess, if I had to say where it goes: traditional BI looks very different, and we kind of get rid of it. Partly because it never really worked in the first place. People have their actuals, they have a bunch of dashboards, they’re mostly little apps for salespeople to look at their list of leads and how stuff is going. The idea of these things being “I go here to find — I drag-and-drop my way to insight” is not really a thing that happens. They become very focused on the presentation of it — “I need it to look like all the styles and whatever.”
I kind of think that does become: we figure out a way that that becomes — people get really accustomed to chatbot stuff. I don’t know that that necessarily means — tipping back to the gymnastics point earlier — I don’t know that that becomes a chatbot that writes SQL queries for you. In some ways I think it becomes more of “we just bypass SQL queries altogether.” If I want to understand what my customers are talking about, I could drag-and-drop my way to doing a bunch of analysis on top of some structured data set of support tickets, or I could just be like “tell me the vibe of the support tickets.” I’ll probably just do the other one and never bother with the analytical stuff first.
The other way to talk about this is: we have done data work because it’s the only way we’ve been able to look at things at scale. If you have a million support tickets, how do you look at that? You can’t. But with math — math works really well at any scale. You can average a billion numbers just as easy as you can average 10. How do you average a billion conversations? You don’t. There’s no way to consume that — have researchers read some of them and summarize them manually. Now we have a thing that can very approximately do math on text.
So the point of numbers was to do math. We quantified stuff so we could do math on it. But if you can do math on stuff that’s not quantified, do you do that much math? Maybe not. There’s a real essential thing there.
It does feel like there’s a place more for nice tools for people that are continuing to work with — when I open a CSV, how do I open a CSV in a really useful way? There’s a bunch of things that are sort of like Excel, or it’s pandas, or it’s Posit in a way. But Posit is designed for a much richer experience. There’s still people who want to look at data and manipulate it, and there still isn’t really a tool that does it. The ones that start in that space — Mode tried to be like a nice workbench for that sort of thing — get dragged into the BI world. There’s a lot of startups that started as these nice data-oriented utilities and then end up being a BI tool. There could actually be something that’s just “this is a really good way to view and manipulate CSVs.”
Wes McKinney: Yeah. In the last two years, I’ve worked on exactly that inside Positron, which is Posit’s VS Code fork. The use case was: I have CSV and Parquet files in my workspace. I’ve got Python code, I’ve got R code. I want to be able to just click on the CSV file and see what’s in it, see some summary statistics. I’ve got Parquet files — I want to be able to just click on them and look at what’s inside. I shouldn’t have to write a bunch of code to analyze it. Why not just open the file, look at what’s inside? That turns out to be not only useful for humans, it’s useful for agents too.
Every BI startup, I think, underestimated the value of “I can just send you a file.” There’s this Excel-files-are-an-anti-pattern thing — “don’t send us it, it’s in the cloud, it’s a web thing.” But for a whole lot of people, it’s really nice to be able to email you a file that is the data itself. It’s a static version of the data. It is the manipulation on top of it. It’s the chart, the whole thing in a box. Whereas “oh, it’s a BI tool, it’s attached to a database that runs queries and all this stuff” — yeah, I get why that’s useful, but so much data work is just files.
Michael Chow: It’s interesting to hear you talk about the vibe aspect — that there are these big sources of data we used to have to quantify because we couldn’t put eyes on them, that now could be fed into an LLM to get a useful gist out of. But there are also these small things — emailed files — that we’ve overlooked, or even just cracking a CSV open that we’ve overlooked at times. They’re just really simple glimpses at things that people want to do.
Benn Stancil: The other version of this to me is: we’ve spent — whenever we started all this data stuff — 20 years figuring out ways to collect it, because we had a thing to do with it. “Let’s collect web events on every single click you’ve ever had on the internet. Why? Because we have something useful to do with it.” There’s an enormous industry of data collection, largely structured, because we have something that can be stuffed into something, and was that thing useful? I don’t know, but theoretically. We didn’t actually bother to collect a lot of other stuff, because what are you going to do with it?
An easy example — this seems like a thing that could one day exist: you go on a website, a button like an Intercom type of thing pops up and is like, “would you like a $5 Starbucks gift card to yell at this website for 30 seconds?” I’m sure there are lots of reasons it doesn’t exist. But why would you ever pay $5 for a 30-second audio clip of feedback? What am I going to do with a thousand of those? But if today you have a thousand 30-second audio clips of people telling you feedback, would you pay $5,000 for that? Absolutely you would, because we have something to do with it. Before, it’d be “oh, it’s not going to cost $5,000, it’s going to cost me $25,000 because I’ve got to have someone read them and there’s going to be a user researcher who goes through them and parses them and makes a bunch of stuff. It’s a huge process, and I really trust it.” Now that data is useful.
There are a lot of places like that where we’ve never — we’ve started to do it with things like sucking in emails and Slack conversations and whatever, but that feels very early in “what data would we collect now that we actually can kind of make use of unstructured stuff?”
It’s a little bit Black Mirror-y where — other examples: there are a whole bunch of conversations that happen every day at a Starbucks counter between the barista and the customer. That’s really useful, if you were Starbucks, to figure out what you could do better. Is it collected? No, because what in the world are you going to do with that? Now, what are you going to do with it? A lot. Do you want people to collect that? I don’t know. We very much got comfortable with people tracking every single thing you do on the internet. We probably end up getting comfortable with that because someone figures out how to do something with it. We have a thing we can actually do with it. Those sorts of sources start to emerge.
Michael Chow: Fingers crossed that I get to yell at a website in the next two years.
Benn Stancil: It is scary — the Starbucks thing — where sometimes these things start to feed back into themselves. If the Starbucks worker and me are in a panopticon and we know it, our interactions change because we know. But it doesn’t — you don’t use websites differently because you know what you’re clicking on is. At first you may be like, “oh, I’m going to mess up, this is fun.” And then you do it two times and you’re like, “I am one of a million visitors to ESPN.com every day. I’m not going to — fine. I think they won’t click.” Oh yeah. For a lot.
Michael Chow: We’ve had a sweet long jam sesh. Benn, appreciate you coming on. I feel like the advice to people who are wondering what to do is so helpful — to explore things. We’ve heard so many interesting things, from “in the long term, everything’s a fad” to takes on AI to writing practice. There’s so much to chew on. Really appreciate you coming on and chatting with us.
Benn Stancil: Yeah, for sure. Glad y’all could have me. And yet, do things include “don’t take my advice.” A lot of that is: don’t listen to random people like me tell you anything to do. How’s that?
Michael Chow: Thanks for coming.
[Podcast outro]
The Test Set is a production of Posit PBC, an open-source and enterprise tooling data science software company. This episode was produced in collaboration with creative studio Aji. For more episodes, visit thetestset.co or find us on your favorite podcast platform.