đ§đ The AI Startup Killing the $140B Survey Industry | Alfred Wahlforss, Listen Labs
How to talk to your customers with AI, advice for pursuing enterprise customers early, how to leverage investors for customer intros, and hiring for obsession
Alfred Wahlforss and the team at Listen Labs built the worldâs best customer research tool, enabling anyone to talk to their customers at scale, powered by AI.
Listen landed Microsoft as an early customer, and we talk about why more startups should pursue enterprise customers early, how AI is changing the $140B customer research industry, why 85% of survey answers are random clicks, and why interviewing customers at scale with AI gets the best customer feedback.
Listen recently announced raising $100 million from funds like Sequoia and Ribbit, and Alfred shares how they leveraged VCâs for customer intros before and during the fundraise, how Listen used billboards to stand out when recruiting, and how to hire for obsession.
Support this Episodeâs Sponsors
Numeral: The end-to-end platform for sales tax and compliance.
Flex: The all-in-one bank for business owners. Apply here.
Amplitude: AI analytics. All you have to do is ask.
Inquire about sponsoring future episodes here.
đ Stream on Spotify and Apple
Timestamps to jump in:
0:14 Listen: AI customer research tool
7:30 Fraud is a big problem in customer research
9:06 The $140B customer survey industry
12:08 Why running customer surveys is so hard
16:03 AGI will never replace humans
18:25 Surveys vs interviews
21:13 Importance of emotion in data collection
22:54 Using AI interviews to get product feedback
26:15 Building digital twins creates better data
32:22 Outperforming generic AI tools
34:17 Sweetgreenâs Max Protein Bowl
36:09 Jevonâs Paradox in customer research
40:37 Quantitative vs qualitative
42:38 Landing Microsoft as an early customer
44:50 Targeting enterprise customers from day 1
48:05 Building a VC customer intro leaderboard
51:53 Recruiting with billboard games
57:20 Hiring for obsession
1:02:07 Alfredâs favorite movies
1:03:53 Listenâs custom agent harness
1:06:24 Velocity Fellowship for Swedes moving to SF
1:08:34 Growing up with entrepreneurial older brother
1:09:46 No shoes in the office
Referenced:
Try Listen
Careers at Listen
Sweetgreen case study
Toni Erdmann on IMDB
Prior episode with Erik Bernhardsson @ Modal
Find Alfred on X / Twitter and LinkedIn
đ Stream on YouTube, Spotify, and Apple
Transcript
Find transcripts of all prior episodes here.
Turner Novak:
Alfred, howâs it going? Welcome to the show.
Alfred Wahlforss:
Yeah, thank you for having me. This will be fun.
Turner Novak:
Really quick, for people who donât know, what is Listen? How do you describe it to people?
Alfred Wahlforss:
We built this AI agent that can understand what people want by talking to them. Marketers, PMs, user researchers will go to Listen and ask a question. Microsoft is one of our customers, and they can go and ask, âWhat do CIOs think of Azure versus GCP or AWS?â
Listen will go and find hundreds of CIOs. We have a database of 30 million people. Then it will run interviews, sort of like Zoom calls, with hundreds of people in parallel, and then give you recommendations of what youâve learned. Then you build this repository of all of the interviews in one place.
You can start to query that. Now weâre also building simulations, so you can actually use the interviews youâve collected to simulate how people will answer questions in the future. We can talk about that later. Weâve raised $100 million. Weâre used by a large portion of the Fortune 100, including Microsoft, Anthropic, Sweetgreen, P&G.
Turner Novak:
Anthropic is considered Fortune 100 now? I guess theyâre pretty big. Theyâve gotten pretty big pretty quick.
Alfred Wahlforss:
They probably would be up there, yes. We also are used by startups like Perplexity, Cursor. I think 20% of the Forbes AI 50 use Listen as well. Itâs really every company that wants to understand their users better.
Turner Novak:
So then how does it actually work? If I am a marketer, and I want to know more about what someone thinks about Azure versus GCP versus whatever, what kind of work do I have to do, and what does it look like when Iâm using the product? Just kind of talk me through how it would actually work practically.
Alfred Wahlforsss:
You first start by telling Listen what you want to find out, and then it creates this interview guide. Itâs a semi-structured discussion guide. Thatâs the technical term. Itâs basically allowing the AI to have some structure while also being able to ask follow-up questions, go on tangents.
Turner Novak:
The AI will go on a tangent? Really? Okay. It wonât tell you the whole story, or is it hallucinating?
Alfred Wahlforsss:
It knows your business question, the context, and then itâs able to ask follow-up questions. So if someone is giving you a bullshit answer or theyâre going off topic, itâs able to ask follow-up questions. Itâs like, âOh, thatâs interesting. Can you actually tell me a little bit more about that?â And it learns across all of the interviews to really dial into what is the core insight here.
Then it runs this, itâs all over video, so the interviewer itself is actually text-based. It can also speak, but we find that avatars are kind of janky right now. Expect that to be working at some point. We pay people to answer the interviews, thatâs why they answer them. You can also interview your own users by just sending an email.
Then it writes these reports, slide decks. You have a chat, so you can ask questions across the interviews. One example is Sweetgreen. They launched their new protein bowl based on insights from Listen. Manscaped, they tested their Super Bowl ad and radically changed their brand perception and positioning based on insights from Listen.
Cubbies, we work a lot with apparel brands. They kind of interviewed kids using Listen to figure out, you know, AI is great for these slightly uncomfortable topics, or if you want people to be able to share in an honest way. They were able to interview kids who talked about how the liner is uncomfortable, and they were able to launch a new product line that was really successful.
You can also use it to test products. You can have the AI share your, you share your screen to the AI agent, so it can actually see what you do on the screen as well. Thatâs a few of the examples.
Turner Novak:
You said that you have this network of 30 million participants. So what exactly is going on there?
Alfred Wahlforss:
The way weâve created that network is by partnering with over 200 different providers. There are these API partners that can provide interviews that can be really niche. WebMD is one example, where they have a unique way to access doctors, and so we can partner with them to find those doctors. Or you can have the expert networks like GLG and AlphaSights.
But then we also have our own participant pool. If you go to Listen and ask a question, itâs almost like a marketplace where multiple partners will bid on each query to say, âHey, I can find 100 doctors for this price with this level of quality.â
Then we have something we call a quality guard. Itâs able to check who someone is based on all the interviews weâve done in Listen to check consistency. A big problem in research overall, this is a financial transaction, there will be fraudulent actors. Someone might show up as a software engineer in one interview, and then they put on a hat, and all of a sudden theyâre a doctor or something, or like a fake mustache.
Turner Novak:
Yeah. I could see that being a problem.
Alfred Wahlforss:
Itâs a huge problem, and it was kind of shocking to us because we worked with one of the multi-billion dollar revenue market research companies. They sent us participants, and they were supposed to be B2B decision-makers, and they were clearly people from Sub-Saharan Africa that could barely speak English.
Because in surveys it looks really clean, you get these beautiful charts, and with us itâs over video, itâs open-ended. Itâs much harder to keep a high level of quality. We can actually check if you are who you claim you are. If you are not consistent across all of your interviews, you never get to do an interview with Listen again.
Turner Novak:
Oh, really? So if I signed up as a participant and I was saying I was a doctor, and I do the first one, whatever, maybe I pull it off, and then I sign up again with all the same information, and itâs for a mechanic or something thatâs completely unrelated, youâll start to flag like, âWait a second, this guyâs obviously not who he says he is.â
Alfred Wahlforss:
Exactly. Since weâre vertically integrated with the panel and the interview, weâre able to do that, which none of our competitors typically are. Thatâs a big value add.
Turner Novak:
Thereâs kind of this pretty big customer research, customer discovery market, right? I think I saw the number was like $140 billion that people spend on doing these surveys essentially.
Alfred Wahlforss:
Itâs an absolutely massive market. The software spend is there. If we talk to a Fortune 100, they typically spend about $10 million a year on Qualtrics. But then they will spend on top of that hundreds of millions to market research agencies. Thereâs this large services market because itâs historically been really hard to find the right audience and to analyze the data, and we can turn the services into software and automate a lot of the hard work.
When you think about it, every single company wants to understand their customers better, and thatâs why itâs such a large market.
Turner Novak:
Who are some of the legacy larger players in the space if people listening have maybe heard of them before? I feel like Nielsen is one. They do the TV. People might know them for TV ratings. Qualtrics is software. SurveyMonkey is another. I think those two were or are publicly traded.
Alfred Wahlforss:
Qualtrics went private for roughly $12 billion. There are these services firms, Kantar and Ipsos, that have billions in revenue. Thatâs another legacy player. Then thereâs a long tail of small kind of agencies.
Turner Novak:
So when you were coming across the space, was there ever a thinking of like, oh, they should make software to do this? Are they automating things?
Alfred Wahlforss:
We kind of got into the space by building Listen for ourselves in the beginning. Then we learned more about this market over time and realized that itâs just really hard to adapt your technology. First, the services firms, they donât have the capability to build it in-house. Itâs just really difficult to build.
Then if you already have a working software like Qualtrics, which is a survey platform, they have millions of people running through their interviews. They have the problem that if you add an LLM, the gross margin becomes worse. They also have to change the deterministic flows that they already have, which some of the customers that are already running the flows will find frustrating if theyâre just switching it over overnight.
So they kind of have to build an entire new product to do this. And large companies tend not to be so good at building new products from scratch.
Turner Novak:
So then what does the traditional process of running a survey, like a customer survey, kind of look like? Letâs say Iâm Microsoft, I want to do some research. Whatâs my process generally look like, maybe pre-Listen and then post-Listen? What does it look like before you guys and then how does it change when Iâm using you?
Alfred Wahlforss:
In the large enterprise, you typically work with an agency. It will be this back and forth process where you might have a question, but you donât know the methodology to answer the question. For example, if you want to understand pricing, you canât just ask, âHow much are you going to pay for this?â You have to use the right question methodology, and itâs actually an academic subject. Itâs really hard to learn how to design market research studies well.
Turner Novak:
My mother-in-law actually has a PhD in survey research methodology. She should come work at Listen. She used to work at the University of Michigan. They have this social research institute. And then a company called Westat, I think itâs called, is pretty big. They do a lot of government research stuff. And then I forget the name of the company she works at now, DLH or something. But she literally designs and runs surveys all day.
Alfred Wahlforss:
Right. You can ask her. Itâs not easy to get this stuff right. You typically have to go to an agency, then you have this process back and forth to design the discussion guide. That means you write it by hand. Okay, is it this question? No, itâs that question. Itâs like a long discussion to get that right.
Then you go and find the people, and that can take weeks, especially if you actually do interviews. You can imagine all the scheduling that you have to do if you want to do 50 interviews to make sure you have some kind of large scale. Then analyzing 50 transcripts is really difficult as well.
The process can take eight weeks to do and hundreds of thousands of dollars. One of these agency projects can be $300,000, $500,000. Thatâs literally talking to, I mean, we talked to a pharmaceutical company that week who said, âYeah, to talk to 20 doctors in eight markets, itâs $300,000.â
You can think of international work, it adds another layer of complexity, where now you have to find another agency thatâs like a layer on top that speaks this language and that can translate to the other agency, and itâs just very inefficient.
With Listen, you can get this done in 24 hours. You go to Listen, itâs very opinionated with the questions. It finds the audience very quickly. In five minutes, you can get 10 interviews done, depending on the length of the interview. Itâs a really magical experience when you see people just show up answering your questions immediately. And then obviously it analyzes the data very quickly as well.
Turner Novak:
If I wake up one day and Iâm just like, âI wonder what people think of this podcast. I want to get some feedback on it,â I spin up Listen, and I maybe set up a survey. It sounds like itâs not a survey. Itâs an interview. Oh, itâs an interview. Okay. Maybe thatâs something we should hit on in a second.
So I tell Listen what I want to get, and then I will click a button, and maybe my podcast listeners arenât on the Listen network, but how do you go and recruit people automatically, and then they click a link, and they do it within 10 minutes, and then I get out of my next call, and I have something sitting in front of me of, âHereâs all the data we collectedâ?
Alfred Wahlforss:
Exactly. The way we find people, we essentially put everyone in this embedding space based on all of the interviews theyâve done on Listen. So you know what are the questions they can answer, what is their expertise, and over time, this gets smarter and smarter. We send them an email saying, âHey, we think you would be a good fit for this interview. Do you want to take the question?â
You can imagine in the future where weâll have a phone number that you can just call if youâre ever feeling bored or maybe youâre driving, and you can just answer market research questions on demand, and you get paid per minute.
Turner Novak:
That could be good. Youâre driving to work every morning, you make 10 bucks, 20 bucks, answering questions. Do you use Tide? What do you think about Pepsi? I know you did something with Sweetgreen.
Alfred Wahlforss:
Maybe thatâs the last job for humans. Itâs a little bit dystopic, but as the models get better, as we get to AGI, I think the hard part will actually be knowing what to build, not how to build it. Thatâs what we want to do. And I think to do that right, you need human input, and humans are inherently irrational. So I think AGI will have a hard time predicting exactly how weâre going to answer.
Turner Novak:
Iâve always had a really hard time with this. Just AGI completely taking over the economy or whatever. Humans always need to do things. We will always be the reason that the computer and the software exists, right? Even when you read those dystopian books where the world is a simulation, itâs basically the computer is still serving humanity, like keeping us safe, creating simulations to keep us going.
I always have a really hard time with no oneâs gonna work and AGI is gonna take over everything. Itâs just a little far-fetched in my opinion.
Alfred Wahlforss:
Yeah. And I think whatever happens, weâll have to give input asking, what do we want the AI to do for us?
Turner Novak:
So you mentioned specifically surveys versus interviews. Can you just explain why thatâs a big deal? I think I might kind of get it, but I think maybe someone listening might be like, âWhatâs the point? Arenât they the same thing?â
Alfred Wahlforss:
We live in a very divided world. Thereâs wars going on. Everyone likes different brands like Pepsi versus Coke. But thereâs one thing that we can all align on, which is everyone hates surveys, because itâs so boring to answer a survey. Iâve never met anyone who said, âI love taking surveys.â
You have to answer these multiple-choice questions, and if you do that for more than three minutes, it just becomes super repetitive, and so you end up just clicking random buttons. In fact, weâve actually done research on this where we went back to the same person two weeks later asking survey questions, and they ended up being about 85% consistent per question.
If you then scale it up to 30 questions, the whole result becomes extremely noisy. People are not even paying attention when they answer surveys. When you do that with Listen, you have to actually think. You take a, you have another entity thatâs engaging with you. We find that people open up much more, and they are much closer to how they actually behave in the real world.
Turner Novak:
Itâs much more like theyâre having a conversation with someone versus sort of a one-way filling out a form, clicking buttons.
Alfred Wahlforss:
Yeah, exactly. And itâs much more engaging than doing that. We let people be human, and surveys turn them into robots.
Turner Novak:
Because if thereâs like, you know, if you ask someone, âDo you like Pepsi?â Someone might say, âYeah.â Right? Like, âYeah, whatever.â And maybe thatâs like a 10 out of 10 in a survey. But if I answered it that way, Iâm not very enthusiastic about it. But if I was like, âOh, I love Pepsi. I drink it three times a day. I donât even have blood. My blood is actually Pepsi because I drink so much Pepsi,â thatâs a way different answer than just a yes.
Alfred Wahlforss:
Exactly. Now these LLMs can also read your emotions. It can look at your video feed and say, âYeah, this person said, âYeah, this is great. Iâd love to have this. Iâd love to try this if I had more time.ââ But it can tell that this is someone whoâs never going to try this product, or maybe theyâre even sarcastic. You can really kind of translate human emotion into action.
Turner Novak:
That was a new feature you guys launched recently, right? This emotional intelligence, I think you called it.
Alfred Wahlforss:
Yeah, exactly. We have this model of human emotion. We can read six different emotions, and then we can use it for analyzing your responses. One good example is advertising testing. The holy grail of market research is to read someoneâs mind and see how did they actually react to this thing directly, and this is the next step in doing that.
Turner Novak:
Interesting. So what exactly is it doing? What kind of things can you pick up on? Is it like raised eyebrows? Is it like how their mouth moves to represent excitement or passion or something, or disgust?
Alfred Wahlforss:
Itâs not perfect, but itâs getting a lot better. I think itâs around 60% on our eval, and humans are around 80%. Itâs both audio and video, so it will pick up on your intonation. If you raise your eyebrows, it picks up on that. We try to train it to avoid hallucinations as well. Sometimes it will read into too much of the video. But we use Gemini and a couple of other models to do that.
Turner Novak:
You mentioned that you work with Microsoft, you work with Sweetgreen. I think I saw that VCs are using Listen to actually do diligence on companies. So how are people using it? What are some things that people are getting out of it? I think you mentioned Chubbies earlier too.
Alfred Wahlforss:
Things like ad campaigns, get product feedback, understand brand perception. Anthropic uses it for, if you churn from Claude Code, Listen will figure out why. In some cases, if there is a bug, Listen can actually send that to another agent, which will create a ticket or coding agent that will actually solve the bug.
VCs use it for diligence, so youâll have this whole process of talking to the customers of different products and understanding, do they actually like it? Thatâs another use case. You can imagine Procter & Gamble, theyâre constantly launching new products in new markets, and to launch one of these products is tens of millions of dollars in ad spend and also retail shelf life. If you can validate and understand how you should launch it, it can save a lot of money.
Turner Novak:
Itâs always interesting, like youâre Procter & Gamble, and itâs like, all right, weâre coming up with a new chocolate. Do people like chocolate? Or should we add dark chocolate? Should we make it 70% instead of 60%? And they do this whole research campaign. Theyâll talk to this process of hundreds of people to make change to some food or change the packaging.
It kind of seems a little bit silly, I guess, but thereâs just so much at stake that they definitely, theyâre like, âAll right, if we make the package green instead of blue, how will that change the perception, and whatâs the ROI on that?â So I guess it sounds a little bit ridiculous, but also it makes total sense that especially the more resources you have, the more youâd spend on this stuff.
Alfred Wahlforss:
It can have a huge impact, right? The package that you choose. It can even be, you constantly make decisions every day that in some way youâre not fully aligned with your customers. You donât know exactly what your customer would want in one case.
I use Listen myself. We have created a simulation of our customer base. We built this ability to interview one person and then create a digital twin of them by doing essentially a one-hour long-form interview. Then you can scale it up to a thousand people, so you have a representative sample.
The other day, I was figuring out whatâs the title of my talk for a conference with our customer base. Itâs a really small decision, but it actually does matter. Are 20 people going to show up, 50 people going to show up? By iterating with this synthetic panel, I was able to get to a much better result than I initially had. I think that if you can help improve every single one of those small decisions, you will have meaningful change in a large company.
Turner Novak:
You said something interesting about these synthetic personas or datasets. How does that work and how is it useful? Iâm just curious because Iâm thinking, do you run into different biases or, you know, itâs not actually real customer data because itâs synthetic or made up. How does that actually work?
Alfred Wahlforss:
Our core product is really focused on talking to real humans. We realized that weâve done more than a million interviews in the platform now, and itâs grown exponentially since we last reported it. We said, what if we train digital twins based on all of those interviews? That would be really powerful.
You could think of this as, if you have a partner, you spend a lot of time, then probably you can predict to some degree what theyâre going to like and not like.
Turner Novak:
Like if theyâll like a new food or if theyâll like a movie or something like that? You think you can do that?
Alfred Wahlforss:
Barely. I will say my wife can probably do that much better about me.
Turner Novak:
Than I could about her. But I could give it, depending on what it is, I could probably call it, but she knows me so well. Sheâd be able to, like, anything, be like, âOh yeah, Turner would or wouldnât like that.â
Alfred Wahlforss:
Okay. So your wife has a good model of you. It turns out that LLMs can build this model quite successfully. We have in some cases like 95% accuracy, and we measure that by just removing one of the questions from the training set and testing how well is the AI able to predict the answer to this question.
You can get very high accuracy. The problem is that obviously there are questions you canât predict. The model needs to know what it can answer, what it canât answer, and whatâs the confidence interval. The use cases are, I would say, brainstorming, the 99% of decisions where itâs too difficult to talk to real people, or you need answers really quickly, or itâs a really small decision, but it still matters, like the title of a talk.
Or if thereâs hard-to-reach audiences, like high net worth individuals, really expensive to talk to. Now you can create these simulations of them. If you just talk to a hundred of them, you can have some kind of simulation.
Turner Novak:
Iâm trying to think of what something could be. If Iâm like Doritos, or like Taco Bell, they always come up with these crazy new products. I could maybe say, âHey, should I make a strawberry-flavored Dorito?â I could probably go into the Listen dataset and a bunch of people have maybe mentioned how they like strawberries or something, or they donât. So Iâd be able to maybe get a little bit of feedback on, hey, it looks like people may actually be interested in strawberry-flavored Doritos, or you have enough history to say people probably wonât like that.
Alfred Wahlforss:
One of the best use cases I think is message testing. That is basically what is the title of this billboard? What should I name my product? These really difficult, vague decisions that may or may not have some reference that you donât know about, that went viral a few weeks back, and youâll be ridiculed by it, or a specific set of framing.
Also, even aligning yourself and aligning other people. Iâve actually created a synthetic version of myself, and sometimes when I have decision fatigue, Iâll throw that in, like, what should I have for lunch? And Iâll just let my synthetic AI choose for me. Itâs just easier to have someone else make the decision. Thereâs value in getting faster to decisions.
Turner Novak:
One thing maybe relevant is, for this podcast, when Iâm trying to think of, what do I title this thing? What should I put in the thumbnail on YouTube? I always just basically copy and paste the transcript, and I have a Claude skill thatâll just basically bang out a bunch of ideas, and nine out of 10 are pretty bad. But thereâs usually some in there that are pretty good. Iâm like, âOh, I did not think about framing it this way.â
Even when I asked it before for prepping, part of it was like, oh, you should give it this immigrant to successful founder type of framing, or you should give this AI unlocks the qualitative side of humanity, even though itâs a very quantitative or something. And I, again, it was like, I didnât feel like any of those really hit.
Iâm gonna see after this conversation, Iâm literally gonna throw it in, throw the transcript and be like, âWhat are some ideas?â But it always comes up with, usually thereâs a couple that are pretty good that I wasnât thinking of.
Alfred Wahlforss:
Whatâs interesting is that the taste of the models are trained on the average user. When we tried this, we asked Claude, ChatGPT, what do you think? Even if you tell it, âHey, you should act as a market researcher,â whatever, it has different opinions than our synthetic or our digital twin panel. Itâs not as aligned with your specific segment.
Imagine if you had created a simulation of your user, the people listening to this pod, you could then have that in as an MCP and let Claude kind of iterate together with that simulation to come up with the perfect title.
Turner Novak:
Interesting. I need to figure out a way to automate, because itâs all still kind of manual. I need to do probably some more like Cowork automation stuff. When it notices that Iâve recorded an episode, it will automatically go and run. I havenât gotten that far yet. I need to.
I think this begs maybe an interesting question of, if Iâm a brand, couldnât I just go to ChatGPT or Claude and just be like, âHey, hereâs what Iâm thinking. What do you think?â Whatâs the value of using something like Listen versus just a more general AI tool?
Alfred Wahlforss:
The value for simulation is that the results are different. If you ask Claude, it has much worse taste than the simulation, because itâs not based on your specific sub-segment. If you think of something like Sweetgreen, you would think that, okay, thatâs a general audience, but actually itâs high income, itâs urban, and by the way, they need to know what seed oils are, and all of a sudden itâs a very small subset of the population, and it lacks a bunch of the nuance that you get from the interview.
We see a very meaningful lift in the accuracy. When you look at pure Claude accuracy is around 40%, and we get 95% accuracy in some cases.
Turner Novak:
What is that accuracy like, that 40%? Whatâs at 40% and then whatâs at 95%? Is it like the success of an outcome?
Alfred Wahlforss:
Itâs the mean average error in answering a question. We will remove 10 questions from the training set and then predict how will we answer this question. We let Claude do that, and we have the real answer as well, and then we see whatâs the average error, and we get about 5% of error.
Turner Novak:
One of the things you mentioned a little bit earlier, that you work with Sweetgreen. I think itâd just be interesting, they actually developed a product using Listen. What did Sweetgreen use you for?
Alfred Wahlforss:
They came to us and said that the menu has had issues with protein, and we did a study where we interviewed Sweetgreen customers, and we ran hundreds of interviews. Listen came out with an idea that they should create a new bowl, called the Max Protein Bowl. They ended up actually launching that, and it became a huge viral hit, and a lot of people are buying it now.
Thatâs the kind of use cases that work really well when youâre trying to do ideation or concept testing, and you see how people react to these new ideas.
Turner Novak:
Yeah, I feel like Sweetgreenâs really good at being on the forefront of new technology that comes out. Actually, the very first guest of the podcast was Jonathan Neman, the CEO of Sweetgreen. I think at the time they had just launched their robotic kitchen, where they were using these autonomous robots to automate some of the preparation.
I feel like they were pretty early on mobile takeout. Mobile ordering and takeout, which obviously you can optimize the kitchen. Iâve looked at the stock price recently, but I know that, I feel like Iâve just generally, that categoryâs been struggling a little bit, just the pricing. Consumers are getting a little bit upset about the Chipotle slop bowl memes. Iâm sure youâve seen those.
Theyâre really great at testing new things.
Alfred Wahlforss:
Theyâve been an amazing partner from the beginning.
Turner Novak:
So I know in AI, thereâs kind of this Jevons paradox thing, where the better it gets, the more that you do. Is there a similar element going on with broadly customer research? Are you finding that people are doing more and more talking to their customers because you make it easier and faster?
Alfred Wahlforss:
I think there are these examples where thereâs no limit to how much value you can get out of a specific segment or a specific task, and customer research is one of those. You can always perfect whatever you do to make sure that itâs fully aligned. Our vision is to create a world that finally works the way people want, and thereâs so many small things that are misaligned with what people want.
We actually see that now that you can launch something, like, one of our customers, they used to do these things once a quarter. Now they do it every week.
Turner Novak:
Thereâs a new product or something? Or new event of some kind? Like customer research, essentially.
Alfred Wahlforss:
They used to work with one of these agencies once a quarter. That means they can fundamentally launch more marketing campaigns, more products. They can iterate much faster, and their products are more aligned with actually what their users want.
Turner Novak:
Youâre basically just tightening the feedback loops, speeding them up. Theyâre able to, I mean, really talking to your customers. If you go back to what is YC, the advice to when youâre starting your company, itâs just talk to your customers, build a product that theyâll pay you for. Thatâs basically what youâre helping people do at the end of the day.
Alfred Wahlforss:
What Iâm really excited about is, when the coding models get really good, the YC model is write code, talk to users, and I think the coding models are almost good enough for this. We can essentially give a Listen, set a like amount of capital, and then go and talk to users, figure out what they want, and build it, and run that in a loop, and you have an autonomous organization. Thatâs pretty interesting.
Turner Novak:
How, to what extent can you do that today? Are there certain points where it just doesnât quite work yet because the technologyâs not there yet?
Alfred Wahlforss:
There is still the judgment of when to ask and when to build, and the modelâs reliability is not quite there yet on the coding. But I think towards the end of this year, there will be huge improvements, and especially with the simulation where you can get really quick feedback. I can see the way we develop will be quite different.
Turner Novak:
You can basically have a product, like somebody in product who is talking using Listen. Itâs going out and talking to customers. Theyâre getting feedback on it, and then theyâre like, âOkay, Devin, just go make it.â Within the course of the day, maybe thereâs time windows for all these things, but youâre basically just kind of sitting there and youâre talking to customers, and then thereâs almost this triangle of product, customers, engineering. Itâs all in one maybe.
Alfred Wahlforss:
Because today the preference model is you as the builder. Youâre building it for yourself, and you kind of have to think, âOkay, what would our customers actually want? What do they actually care about?â But imagine if you could have a simulation of your real user, and thatâs just going to be so much more powerful.
Turner Novak:
Is there people that are doing that well today? Do you feel like thereâs any companies that are the closest to that? Or maybe how do you guys do it?
Alfred Wahlforss:
I donât think anyone has cracked that yet. I think we have an edge because weâre talking to real people all the time, so we have this extremely rich data set that we can train on, and thatâs why Iâm excited for this direction, and weâre hoping to launch this in a couple of months.
Turner Novak:
Oh, so itâs not out yet.
Alfred Wahlforss:
Yeah, itâs not out yet.
Turner Novak:
Oh, interesting. Okay. Whatâs the challenges in building this? Whatâs been the hardest part of actually making it practical and usable?
Alfred Wahlforss:
Making it accurate. The models have a bunch of, you know, theyâre super smart, right? So they will sometimes act in a way thatâs not in tune with how humans work. And a bunch of issues around that, basically. Thatâs the hardest part.
Turner Novak:
Interesting. Because itâs sort of like with anything AI, itâs like how do you quantify everything? Everything needs to be a data point in a sense, but this is still a very qualitative thing. How does something make someone feel? Itâs kind of like this weird balance of how, I donât know, how do you balance it? I donât know if thereâs an answer, but...
Alfred Wahlforss:
You will not be able to replace all of the work we do with simulation, because there is something about talking to real humans and seeing them react in ways that are just impossible to predict. Also being able to share highlight reels of how people actually feel when they see your product. The big value of research is aligning people, motivating them to actually go and fix the problems. Sometimes you know all the problems, itâs just thereâs no one actually going and fixing them.
But having real people react to how bad your experience is can be a really great catalyst to make that happen.
Turner Novak:
Interesting. Yeah, because I feel like, and maybe an example of that happening right now is, a lot of people are now starting to build products and software thatâs kind of agent first instead of human first, right? A year ago, that probably wasnât necessary, but weâve kind of, as more and more software moves to being more of agents interfacing with other agents, the human first software is not quite built correctly or in the same way more efficiently.
Itâs like this new problem that emerges where, a year ago, nobody wouldâve thought this was a thing, but then now as the industry shifts, as behavior shifts, demand, use cases shift, all of a sudden itâs like, oh, thereâs actually a need for this to exist that wasnât there six months ago.
Alfred Wahlforss:
Yeah. But those agents will always be doing things on behalf of their humans, right? Thatâs why it will always be very important to understand the humans behind the agents.
Turner Novak:
Maybe speaking about humans, like selling to humans. I know you mentioned that Microsoft was a customer. I think they were kind of one of the big first customers that you have. How did you get them on board so early?
Alfred Wahlforss:
We were really lucky. We ended up hearing about this pitch competition in a niche conference around market research, and we decided to hop in and do our pitch. We ended up winning that competition, and in the audience there were a bunch of enterprises. Product was barely not working at the time. We were extremely early. It was a couple of months in.
Turner Novak:
This was a startup pitch competition?
Alfred Wahlforss:
Yeah, but for market research companies. You would think like, oh, if you raised from Sequoia or whatever, youâre too cool to go to those pitch competitions. A lot of founders have that mentality. When we showed our giant check that we won, some of my founder friends were like, âOh, why did you do that? That must have been a waste of time.â
But it ended up validating us. Instead of having, the typical advice for founders is to start mid-market and then go to enterprise.
Turner Novak:
Sell to startups because theyâll be much faster to convert. Itâs easier to identify the problem and who needs to buy. Usually, itâs the founder, right? And theyâll just make a decision right there.
Alfred Wahlforss:
Exactly. But I think that can be a huge mistake, because you can just skip that step and sell to enterprise directly. Most of the revenue is in the enterprise. Of course it depends on what youâre building. But for us, we just built it enterprise ready from day one, and weâre able to start out with Microsoft, Google, P&G as one of our early customers.
A lot of very successful companies like Wiz have done that in the past, because you just grow so much faster, especially in AI, where the AI budgets are extremely large in the enterprise, specifically traditional enterprise. That would be a piece of advice to go and build for them first.
Turner Novak:
So then how did you convince Microsoft? Because itâs still a big company. You gotta prove the use case. How did you, they were in the audience, like what happened next?
Alfred Wahlforss:
We had printed out this traditional survey that I was sent by IKEA, and I kind of had it as a prop when I gave the talk. I dropped it down, and you see these pages and pages of surveys, and they just felt like, âWow, this is how we understand our customers. Weâre not treating them well enough.â
They were just really bought into that idea, and then we had to sprint and build really quickly. Luckily, my co-founder is the national champion in competitive programming in Germany, so we were able to quickly recruit these amazing engineers from all around the world, get them into SF, and build something that worked when we were ready.
The procurement process took almost a year, and so by then we actually had a working product that was pretty good.
Turner Novak:
What would you say, how many total people at Microsoft did you, like different people, like individuals, did you interface with in that process?
Alfred Wahlforss:
It was surprisingly simple to get the pilot done. It was just a couple, a handful of people. But now weâre working with, I think, 30 teams, and itâs growing relatively quickly as well in the org. Itâs an infinite amount of people that can use Listen at Microsoft. The key is, like, land and then expand.
Turner Novak:
Iâm assuming theyâre probably giving you feedback on the product. You probably added features based on feedback youâve gotten from them, all that kind of stuff.
Alfred Wahlforss:
That allowed us to be kind of building for other enterprises as well, at the same time. It is important to have multiple enterprise customers and not just one, because then you can be kind of get stuck with them. But we always had a couple in the similar segment.
Turner Novak:
So youâre basically telling founders, âDonât try to get one big enterprise customer. Try to get three or four. No big deal.â Thatâs easy, right? Were you able to use logos to then help you kind of ladder up and convince other people to take you seriously because you work with this other company? Is that maybe a benefit to doing the enterprise route?
Alfred Wahlforss:
Yeah, if you have Microsoft, then all the other security and compliance, those procurements, they have their own certification called SSPA. Forget about SOC 2 Type 2. You have to kind of get their own auditors to look at your stuff. It really needs to work. You canât use Delve or anything like that. That was a huge validation for the other enterprises that can be very slow-moving. And then you use them as customer references as well.
Turner Novak:
Oh, yeah. Thatâs gotta be helpful. Plus, itâs probably, they have a friend who works in a similar role at another company, an old coworker or something like, âHey, check these guys out.â So speaking of advice for other founders, I know you had a pretty interesting process for fundraising. What would you recommend other founders do, whatâs kind of the fundraising advice that you generally give people?
Alfred Wahlforss:
Yeah, less about fundraising, but more about the psychology of VCs. One thing I found is that VCs will work much harder before they invest than after they invest. Founders should really use that to their advantage, especially in these crazy times when fundraising is a very hot market.
You should actually ask VCs to go and make a bunch of customer intros for you before they invest. We systematized this. We created a leaderboard that we shared in our investor updates, where you can see which VC is performing the best in terms of intros made. Not just number of intros, but actually closed ones.
Ribbit ended up leading our Series B because they are true workhorses. A lot of, their brand is not really well known, but these name brand VCs, they end up being a little bit complacent, and they actually donât do the work that they promise that they can do. Theyâre great at giving advice, but if you can get 10 customer enterprise intros, that can be worth a lot more.
Ribbit closed almost $1 million in ARR for us before they led our Series B. You can actually get large amount of pipeline from this motion. A lot of VCs will probably get annoyed by this, but it does work. They also find it kind of fun and competitive because theyâre very competitive in nature.
Turner Novak:
Yeah. Hopefully the good ones. The good ones are probably competitive. How do you, like, actually do that in practicality though? Do you, is it a part of the fundraise, or is it like a, hey, or do you mention, âHey, we think we might be raising money in three monthsâ to plant the seeds and get them in the back of their head?
Theyâre like, âOh, I gotta start doing some work.â Or do you say, âHey, weâre specifically picking our investor based on customer introductionsâ? How do you actually tee that up in a way that lands correctly where the VCs will actually be motivated?
Alfred Wahlforss:
You have to be careful to not be too arrogant, but you can also be pretty upfront and say, âHey, youâll get a lot of VC inbound if you do a Series A.â I think it only works at Series A and beyond, because then it also becomes a very significant quantum of capital, and so a lot of people will try to fight to get into your deal.
Theyâll reach out and then youâll say, âHey, Iâm not fundraising right now, but when we do, weâre basically gonna look at this leaderboard, and weâre gonna pick the top folks that perform the best. Would love to get to work.â
You have to, of course, when they do the work, you then have to show that you are building trust with them, and you canât just use people, of course. But they also enjoy being competitive and helping out.
Turner Novak:
So did you build some kind of custom thing or is it literally just like a spreadsheet, itâs like an extension of the pipeline?
Alfred Wahlforss:
We have a vibe-coded app that we share.
Turner Novak:
So you raised money. I think you said you raised $100 million total. Youâre obviously trying to hire people now, Iâm assuming youâre trying to ramp up the team. What are you looking for in terms of types of people, roles youâre trying to fill? How do you think about adding to the team?
Alfred Wahlforss:
Hiring is one of the most competitive things in this market, especially in San Francisco. Iâm not from here, so I donât have a ton of friends. Itâs really been kind of a fistfight. Moved here from Sweden.
One of the ways that we have tried to differentiate, and generally how I think about how you can get top-tier talents if youâre a small startup, is by really having a distinct culture. As I mentioned, my co-founder is a competitive programmer, so we naturally have a bunch of engineers who are really into hard math problems and puzzles, and we kind of do problems on the weekends, like IMO problems.
We wanted to communicate that, so we created this billboard that we put up in San Francisco that is just a string of random numbers. That, if you were able to understand what that was, which by the way alienated most people, like no one had any idea, like most people had no idea what do these numbers mean.
Turner Novak:
Yeah, I wouldnât have known. Itâs literally like a URL, but itâs all numbers in the URL. Like, I was like, âAh, I donât know.â
Alfred Wahlforss:
But if you do know, itâs like it becomes this secret club, and you feel like, âWow, this is very interesting. Let me go and try to understand what this is.â You realize that it was AI tokens. You could tokenize that, and you were put into this other URL where you had to act as a Berghain bouncer.
We actually had one of these, one of the problems that you do in interviews is quotas. Itâs this optimization problem where you have to figure out who should be interviewed. It needs to be representative of the world. So thatâs actually quite similar to being a bouncer at a club. We kind of reframed this internal problem as a fun puzzle.
We ended up going, it ended up, we spent months working on our fundraising announcement, but this ended up going much more viral than that, which was unfortunate. We just took a picture with our iPhone, published it on X, and it got millions of views. We had 10,000 people actually do the puzzle and ended up, now, everyone who we interview knows about this thing. They donât know what our company does, but they know that we did the billboard at least.
Turner Novak:
Which, I mean, thatâs, 10,000 people that applied is like an early-stage startup. Thatâs pretty hard to do.
Alfred Wahlforss:
Yeah. It was really cool to just see everyone trickle in. We had people physically compete because if you won, you were able to, we would fly out to Berlin as well to go to Berghain. Itâs like this pretty legendary nightclub in Berlin, like an EDM...
Turner Novak:
Yeah. Itâs also really, itâs famous for being extremely hard to get into because theyâre very picky about who they select.
Alfred Wahlforss:
I donât think our engineer then, he did not actually go to Berghain, but he did go to Berlin.
Turner Novak:
If you were to just say, âHey, I want to hire a recruiting agency to help me out,â what do you typically pay from the recruiting agency and what do you kind of get? I think you paid about $25,000 for this billboard. You got 10,000 people that did the problem and applied. If you were to go to the recruiting agency route, what would you have gotten?
Alfred Wahlforss:
For one engineer you can pay $50,000, so itâs absurdly expensive using a recruiting agency. The big problem is that they just reach out cold with 50 other companies. Not only do you pay the recruiting agency, but you also end up being, getting the most competitive candidates that have, that are interviewing at Anthropic, OpenAI, and are getting million-dollar salaries.
With this, weâre able to get a bunch of folks that maybe the others donât know about, but theyâre just really excited about our culture and thatâs been an advantage.
Turner Novak:
Yeah. Thatâs why I think a lot of people donât always remember, when you see some startup thatâs doing some crazy thing, theyâre just like, âOh, why did they do that? That seems kind of a waste of time,â or whatever. But if youâre, Iâm assuming youâre not paying the same salary as Anthropic, so youâre not gonna beat them by just, âHey, weâll pay more money.â
You have to get people that are like, âHuh, this startup seems kind of interesting. Seems like a cool problem. Seems like itâd be fun to work there. I will, you know, Iâll make the jump. Seems like an interesting place to work. Seems like a cool problem to work on. Seems like a cool product.â
A lot of people, they maybe kind of glaze over that part. Itâs actually really hard to just get people to give you the time of day even when youâre trying to recruit your first 10, 50, even sometimes first 100, couple hundred employees, because just no one cares about you if youâre a super early stage startup just getting started.
Alfred Wahlforss:
I always start to think of it from the position of the engineer, right, where they have no idea you exist. Thereâs 50 other companies growing extremely quickly, and how are they gonna explain it when they talk to their friends? How can you make something that you give them a cool story to explain why they joined this company specifically?
Turner Novak:
Yeah. Because itâs like their friends, but itâs also their parents. Letâs say you have someone, they went to a really prestigious school, they got a job at McKinsey or Goldman Sachs or Facebook, whatever, and youâre trying to convince them to make this slightly crazy jump of, âHey, you were like the top 1% your whole life,â and youâre obviously really ambitious.
And your parents are like, âHey, why arenât you a doctor? Why are you doing this startup thing?â There can be a lot of external things that you kind of have to help them solve for too.
Alfred Wahlforss:
100%. Being able to make it kind of high status and also clear why this is a specific fit for them makes a huge difference.
When you think about hiring, we try to find people who are kind of a little bit obsessive. People who I find are great at something that could be even outside of work. Theyâre just really passionate about it. It often translates into being successful at Listen. We have one person, sheâs a race car driver. She has like eight race cars and does drifts in Tokyo. One of our engineers built a jet engine in high school.
I also look for this almost good version of arrogance where you take a lot of pride in your work, where whatever you put out in the world, it needs to meet a certain quality bar. I find that caring about what you do is kind of the most important, especially as the models are just getting smarter and it actually matters less about being smart and more about kind of having agency, being ambitious, and just caring about every single detail.
Turner Novak:
One interesting thread along that is, one of my, like I did an internship with this big corporation in college and the CFO was just talking about what he looks for, early in your career, what do you do to stand out? One of the things, you know, if you just spend that extra 10 minutes, like relook at the thing you did, think of it from my perspective, do the colors look good? Did you use the right font? Did you catch the last spelling error? Did you just spend the extra 10 or 15 or 20 minutes just like giving a shit about the thing youâre about to submit?
I think about that a lot, just in everything. Itâs just like, okay, I just want this all to look good, and spending an extra 10 minutes relooking at it, and maybe you redo something because you found a better way to do it. Super simple. AI wonât tell you to do that, and maybe youâll think of a different lens of looking at something or framing something that wasnât there before, help someone else understand it.
Alfred Wahlforss:
A great documentary about this is called Jiro Dreams of Sushi. I donât know if you have seen that one.
Turner Novak:
I actually havenât seen it, but heâs like a guy who runs a sushi restaurant or starts a sushi restaurant or something, and itâs super successful.
Alfred Wahlforss:
Itâs about this sushi chef who literally dreams of sushi, and heâs been doing it for 60 years, and heâs still obsessed with trying to refine every single part of the detail of how you cook the rice, how you make the omelet, and just has an insane quality bar.
With AI, and you can generate AI slop now, this becomes more and more important. We see this in our interviews as well. Thereâs a bunch of folks that will be like, âOh yeah, well, I generated this case study in 10 minutes with Claude. Itâs good.â But they actually donât look at the details. So loving the details, thatâs one of our values. Itâs more important than ever.
Turner Novak:
Interesting. Well, speaking of films, I know youâre really into old films. I think on your website you have a couple, a couple favorites that were released decades before we were both born, I think from what I saw. What are some of your favorite movies, and what do you like about them?
Alfred Wahlforss:
I wanted to be a filmmaker growing up. I think thereâs actually a lot of similarities with being a director, as with being a startup founder, because you have this kind of interdisciplinary group that you have to align on a mission, and you have to learn the technical aspects of editing as well as the creative stuff like writing great scripts, directing the actors, sort of like your employees. You have to kind of align on your mission.
At one point, I used to watch a film a day back in high school. One film I really like is called Toni Erdmann. Itâs actually newer, but itâs a German film which is about a management consultant and her relationship with her dad. Itâs really funny. It has Sandra HĂźller in it, who was in Project Hail Mary, I think was one of her films when she became kind of famous.
I think overall, watching the classic films or reading fiction is a really good way of understanding what people want. Storytelling is a really important skill if you are a startup founder. I recommend everyone to watch Ingmar Bergman. He is a Swedish film director.
Turner Novak:
Interesting. A movie about a management consultant. Okay. Well, Iâll throw a link in the description for people to find it. I guess I have to ask, because weâre talking about films and video, whatâs your opinion on this whole launch video kind of culture or phenomenon? It kind of feels like we moved past it a little bit maybe. We kind of, we got to the ironic stage where thereâs people making ironic launch videos. I donât know. Whatâs your opinion on all of it?
Alfred Wahlforss:
I think it is a sign that we are maybe in a bit of a bubble right now, and as you said, I think itâs gone full circle where people are like, âActually, weâve just raised some money. We donât even have to announce it anymore.â But itâs also fun. Because I wanted to be a filmmaker, I also see, okay, yes, now we can do another launch video, and I can do that as a small video project.
We had a very ambitious Series B launch video.
Turner Novak:
Yeah, you said that you spent months on it.
Alfred Wahlforss:
That was, like, our Series A, or actually, yeah, our Series B as well. We spent months on it, and it was very intricate, but then we ended up watching it when it finished, and it was really cringe. At last minute, we ended up scrapping the whole thing. We had just the beginning where I jumped through a survey. It was very, very painful. You have to kill your darlings. Maybe when we IPO, we can publish it.
I think trying to be funny is just really hard. You do a good job.
Turner Novak:
Oh, thank you. Yeah. Well, itâs true. Itâs like you just donât know whatâs actually funny sometimes. Thereâs, and itâs the interesting thing about the internet, is sometimes Iâll post something and I thought it was pretty funny, and Iâm like, âOh, this is gonna do good.â And it gets like 5,000 views, which is kind of not that good for me normally. Iâm like, âMan, I thought that was pretty funny. I canât believe people didnât like it.â
And then Iâll have some where I just kind of think of something, and Iâll just, Iâm about to go in to get dinner, and Iâll just tweet something, and Iâll forget about it, and Iâll come back two hours later and itâs at like 50,000 views already. The next day it has like a million views. Iâm like, âWow, I didnât think that much about that one,â but it just really hit, I guess. So you never know.
Alfred Wahlforss:
Yeah. Thatâs what I found as well. When you overthink the launch videos, it ends up being not good. But there are these entire consulting firms now that have perfected the launch video and how to launch on X, and it does actually matter a lot. The VCs will really index on it, which I think doesnât really make sense, but for some reason they do, so you kind of have to play the game.
Then thereâs a lot of people who will juice the numbers and actually pay people to use bots to boost their X videos. You can always look at the retweet and like ratio. When they now have the view count, it ends up being wildly misrepresentative. The view count can easily be gamed. The like count is much harder.
Turner Novak:
Yeah. Well, I get probably on average one or two offers a day to just, âHey,â from these agencies, âHey, weâre working with this company.â And they might be like, âListen, okay, weâre working with this company. Weâre announcing their Series B. Itâs a cool AI product.â
Honestly, Iâve gotten up to $5,000 is what people have proposed, and Iâm just like, I could probably, if all I did all day was just retweet launch videos, I could probably make over a million dollars a year just retweeting launch videos. Part of me is, Iâm like, âOh, Iâm an idiot for not just doing this.â I should just retweet launch videos all day.
You can literally make $2,000 for just replying to a launch video like, âWow, great video,â rocket emojis. A lot of times they tell you what to post too. Theyâll say, âHey, quote tweet this video,â and be like, you know, if youâre listening, it might be like, âWow, AI is changing the market research game. The survey companies are cooked,â or something. Theyâll pay me like $2,000 to post this.
If youâre on the other end, if youâre somebody that has a pretty big following, you can make quite a bit of money. For me, the big tie-up is, cool, I make $2,000, but then everyone kind of is like, âWait, why were you just reposting this slop video and a product that doesnât work properly?â
For me, thatâs where I usually get caught up with this. Itâs, this was a pretty big problem with crypto. It was pretty prominent in crypto where theyâll give you tokens in this crypto thing and then theyâre like, talk about it to your followers and make the price go up and you make money because we gave you the tokens for free.
This stuff is always kind of around. I think itâs just about being able to dissect it as an audience member, but then also somebody who has more of an influence, just being cognizant of what youâre promoting to people, because there will be people that theyâll see what you post, theyâll use the product, theyâll pay for it, and if itâs not a good product, then theyâll be like, âWait, why did Turner tell me to use this? It sucks. Iâm not listening to him anymore.â
Alfred Wahlforss:
Thereâs so much noise, and as a founder, you just have to focus on what do the customers want and just build that. You will see your competitors show up virally and a bunch of people will boost it, but you never know whatâs going on on the back end, and itâs easy to get lost in that.
Turner Novak:
I have a portfolio company who they have a competitor. Itâs like kind of an adjacent company that always does this. I just, I keep seeing the posts. Iâm like, it is so obvious that this company is just paying for fake influencer engagement. I donât actually know if itâs working or not, but I just keep telling them, âDude, I donât think itâs worth it. I donât think you should do it.â
I think people are starting to come around to, these guys, theyâre paying for this kind of fake engagement. But itâs really hard to be sitting on the other end of that and seeing other people do it. So I get it. I get the struggle.
So a different topic, but I remember hearing that you guys have a certain harness that you made for the agents at Listen. What exactly is that?
Alfred Wahlforss:
An agent harness is the framework that the agent can use to do tool calling and the knowledge management. What we found was that every other harness is built around a file system. Claude Code, for example, will use CLAUDE.md, and thatâs how it kind of has memory and figures that out.
For us, thatâs the wrong architecture, specifically for statistical analysis. Because we think that the right way of building a harness is a table, because you can kind of operate on it as a Pandas DataFrame, which is a tool in Python. Youâre able to, kind of every row is a response, and then every column is a feature.
So every row is like an interview, and then you can extract information for every single interview. You can tell our agent, if you want to quantify something, it can run a sub-agent for every single response and classify, does this person like my product or not? Even if you collected open-ended interviews. Does that make sense?
Then you can easily do aggregated stats. You can run correlations between columns. Thatâs much harder in a file system. Thatâs one thing that I see that these vertical AI companies can do, is essentially look at the job that youâre trying to do as an agent and really perfect the harness, perfect the workflow around that job, and you can get much more juice out of the models than the vanilla model companies.
Turner Novak:
Interesting. It would basically be like if Iâm Doritos and Iâm asking some questions about a new flavor, existing flavors, how I feel about, like, you donât specifically ask me, âTurner, do you like Doritos?â But you will be able to tell if I do like Doritos based on how I answered other questions, essentially.
Alfred Wahlforss:
Yeah, itâs all open-ended, and you feed all of that into the LLM, and then itâs able to predict how you like or not like something.
Turner Novak:
I actually wanted to ask you something. I know you do this fellowship where you bring, talking about yourself, like I know youâre from Sweden, moved to the US. You actually run this fellowship program for other Swedes, helping them move to San Francisco. Whatâs the program, and what do you guys do?
Alfred Wahlforss:
I run this program called Velocity Fellows. I always struggled being the only one obsessed with startups back in Stockholm and wanted to create a space where people like that can find other like-minded founders and then bring them to SF to scale their ambition.
The goal is not for them to move to SF, because I donât want to increase brain drain, but hopefully to bring the SF spirit back to Sweden. We had Max Junestrand, whoâs now the founder of Legora, used to be an intern at my company as well. He was part of batch one. A lot of them have now raised money.
We connect them with, thereâs a bunch of Swedish folks in Silicon Valley as well, like Ali Ghodsi, whoâs the CEO and founder of Databricks. Heâs Swedish. Erik Bernhardsson, whoâs at Modal, heâs Swedish as well. Weâre seeing a resurgence of the Swedes.
Turner Novak:
Nice. I was gonna say, I had Erik on the podcast a couple months ago. Heâs really fun.
Alfred Wahlforss:
Oh, great.
Turner Novak:
Talking about clips and stuff. He actually had one of the most viral clips of the podcast. It was about CO2 levels in the office, the most random topic. But it got like thousands of likes on Twitter, like a couple hundred thousand views. I think it was close to a billion views. People were chiming in, âYeah, CO2 levels. You need to manage the CO2 level in your office. It actually has a huge impact on your work productivity.â
I was like, âWow, did not know that this was such a big deal,â but itâs true, I guess. His, heâs very big on, they have CO2 monitors in the office and make sure that CO2 levels donât get too high because it impacts your brain and makes you less productive. Itâs like, huh, all right, interesting.
One other fun fact that I remember hearing about you is, I think your brother is the founder of SoundCloud?
Alfred Wahlforss:
Thatâs true. Heâs 16 years older than me.
Turner Novak:
Okay. Yeah. I used to be a pretty heavy SoundCloud user, just a lot of EDM remixes and stuff. Less so now. There are just less people post on SoundCloud, I feel like, but I definitely have fond memories of, my first job, I was an analyst at a bank just listening to Chainsmokers remixes and Avicii remixes on SoundCloud for like 10 hours a day.
Alfred Wahlforss:
Yeah. SoundCloud is obviously a big part of my childhood, being seeing him building that company, the things to do, the things not to do. Got to visit the office when I was very young. Iâm also, you know, competitive, so I want to try to build something thatâs bigger than my brotherâs company. But heâs moved now to the Bay Area as well, so we spend a lot of time together.
Turner Novak:
Oh, cool. One other thing I heard you say, youâve mentioned before that you guys have a no shoes policy in the office. Thereâs a lot of, you go on the internet, people have strong opinions of shoes versus no shoes. So whatâs the shoe policy?
Alfred Wahlforss:
Having no shoes makes it much more comfortable. It feels like youâre at home, and that allows for more of an academic environment, I think, which is one of our values, to have folks be able to have free discussions, and you can sit in the sofas and be more open.
We also have Listen branded slippers, so if you do need some shoes when you come in, we help you swap from your sneakers to our slippers. It seems to be a very controversial topic, which I donât fully understand why. Itâs obviously much better to not have shoes in the office.
Turner Novak:
Yeah. Well, I think the thing that I think is kind of crazy is if you walk through San Francisco, you know, not the cleanest city in the world, and then you go into an office, you are walking the same shoes that were on the ground that people are partaking in the external outside activities that happen on the streets in San Francisco that you then do in an office. I can see the value behind it.
Alfred Wahlforss:
Yeah. Itâs from my high school. We didnât have any shoes on there as well.
Turner Novak:
In high school?
Alfred Wahlforss:
Itâs this hippie high school in Sweden.
Turner Novak:
Whoa.
Alfred Wahlforss:
We also only had vegetarian food. It was a school that was controlled by the students. If the students voted for something in the majority, it would happen. They had to stop that after a while because students ended up abolishing homework and things like that.
Turner Novak:
Was that the craziest thing that happened? Was it the no homework?
Alfred Wahlforss:
I think thatâs when they had to pull it back. The vegetarian food was also a big one. It was amazing. It was so delicious. Thatâs the inspiration behind no shoes.
Turner Novak:
Interesting. Okay. Where can, I think if we wrap up right now, where can people find you? I think you post on Twitter. Are you pretty active on LinkedIn?
Alfred Wahlforss:
Yeah, you can follow me on X and on LinkedIn, Alfred Wahlforss, or go to Listen Labs AI and sign up for a demo. Weâre also hiring for engineers, salespeople. Weâre around 60 people and we want to be 150 by the end of the year, so trying to scale very quickly.
Turner Novak:
Nice. Well, weâll throw links to all those in the description, and people can find you. This was a lot of fun. Thanks for doing it.
Alfred Wahlforss:
Yeah. Thank you so much.
