Open source is going to win it all: Harvey proves it | E2328

Open source is going to win it all. Acme litigators and Delta litigators are both using Harvey, let's say. Harvey now is like, not only do we not trust OpenAI with this, we need to have our own model. Harvey was probably spending, if I had to guess, I'm gonna say $10 million a month with OpenAI.

Wow.

99% of their spend is coming off the frontier models. They don't want to give their intelligence to somebody who wants to build the This Week in Startups is brought to you by Lightfield.

Name one person who's ever enjoyed updating a CRM. Exactly. Lightfield's AI agent does it for you. It even prospects and books your meetings. Used by thousands of startups. Free at lightfield.app. Sentry. Your team should be focused on shipping features, not chasing down bugs. New users can get $240 in free credits when they go to sentry.io/twist and use the code TWISTSUPERHUMAN. Superhuman Go is an AI chat that's always there when you need it. Already aware of what you're doing and doesn't ask you to start from zero. Sign up to get the best in AI at superhuman.com.

All right, everybody. Welcome back to This Week in Startups. Boy, Jake Lyle here with Lon Harris. Episode twenty eight. How's that even possible? It's like year seventeen of the show.

We're in sci-fi years numbers now. We're in sci-fi years.

Getting crazy. Yeah, and just literally like yeah, it feels like this is Blade Runner or something like that, and it's appropriate given how great these bots are doing.

Wow!

I just want to start the show off.

You started playing with.

I got it. I owe Elon $200 now. I gotta pay. He's making me pony up.

All right, yeah, do it. Uh, you can sign up for a quarter.

I'm gonna do it. I'm gonna, I'm gonna do it.

Yeah, give us your explanation and what you're using it for, and then I'll, I'll tell you what I'm using it for.

To me, it feels like when OpenClaw was brand new and you had us all sign up and we were all in Slack, I— it was what the promise of that was. But OpenClaw in a virtual machine and like some data center somewhere. It couldn't connect to everything it needed to connect to.

Yeah.

It was too walled off. So it wasn't really that useful. You kept having to remind it things and log it in, this API key, that. GrokBot is automatic. You just tell it what you want. It can find whatever. It's got its own computer. Like just today, I asked it, you know, to look at the news and cross-reference it to X, which is a thing I always wanted Gaff, my old OpenClaw assistant, to do. but he couldn't ever get into X reliably. He would try to backdoor and it would never work. It was always very time-consuming and painful. GrokBot just is like, hey, here, I'm gonna bring up my computer. It actually shows you what it's seeing. It goes, just log into X for me. I logged in through our This Week in AI account.

Yeah. And it's just off to the right.

It just, it sees everything I could see. It's logged in, it's pulling tweets, it's referencing things. It's very intuitive. You just need to tell it in plain language what you want. It's, it's very impressive.

Yeah, the, um, interface is really interesting. I'll just share mine. Fuck, who cares?

Let's see your GrokBot.

Show us your GrokBot.

Well, here's my GrokBot, uh, and what I'm doing. So we have a thing, tactical and practical. If you make it rhyme, people will remember it. So I wanna do more tactical and practical talks here on the show. And, um, you know, I've been just having, um, a conversation with it. Um, looking for, looking for you to make talks, startup founders focused on tactical and practical advice for them to scale their businesses, find product market fit, and generally grow. hiring, sales, paid marketing, fundraising, content marketing, all that good stuff. And it just went to work finding these things and linking to them and building it out, building it out. And then I said, hey, well, give me other ideas for topics and then send it to the Tactical Practical Room on Slack every 48 hours. And then ask Ismael on my team who's in Tokyo right now representing us for Founder University. And I want him to be part of this motion.

Mm-hmm.

And then I said, also, when you find a great speaker, I want you to gimme their X account and LinkedIn accounts for the report. And then I said, please always send a Google Sheet, yada, yada, yada. Like you said, you easily connect it. And it's just doing this every day for me.

Yeah.

And then I asked it like, hey, do me a favor, find me things on Hacker News and Reddit, 'cause those are where founders are actually talking about their problems. Like for example, it's like, here's AI unit economics, pricing credits, token spend. And people really care about that. I hadn't thought about that as something that founders would be interested in.

Yeah.

And this thing is running over and over. Then I made an AI trends one every night for This Week in AI. Do a scan, tell me what the top stories are. It's nailing it. Then I said, give me some competitive intelligence, right?

Mm-hmm.

And so I said, for our newsletter business, I'm so out of it. I want to know what every other organization is doing. Politico, Economist, Forbes, Axios, Semafor. And it is going out, as you can see, And then I told it, follow all of the people who are the most important people in these organizations. And then you were talking about 404 Media the other day.

Sure. Yeah.

I just asked it like, tell me about this publication. And it gave me all the founders. It gave me their LinkedIn, their X, their Bluesky, their Mastodon. I followed them everywhere.

Yeah.

And then I said, just tell me what they're talking about publicly. Then tell me all the event staff, uh, because I wanna hire some event staff. So I was like, okay, let's just figure out who's doing events at other publications. And this thing is just working for me over and over again. And it makes a report on just what people are doing at Skift, which I'm an investor in, Newcomer, Punchbowl.

Yeah.

It found all these for me.

Right.

I didn't have to tell it to find these. And it really is, um, it feels less brittle, which is weird to say, than OpenClaw, which was super brittle, broke all the time.

Yeah.

Can't get on X, can't figure out LinkedIn.

Right.

This thing is just rock solid. So for me, I am a fan of systems over goals. The goal is to have the best news stories here. The goal is to understand what other people are doing for events and newsletters in case we want to incorporate it into what our plans are. Uh, my goal is to find another, uh, co-host here. Uh, Alex is, uh, you know, uh, working on his projects and he'll be back for sitting in for me when I'm on vacation from time to time. But I need another co-host for all these pods. You're doing great, but, you know, we just need to have a roster.

Sure.

So I was like, go find me everybody and then keep finding them. Figure out, you know, CNBC hosts who are getting laid off or who are no longer there. Find me people who are working for The Verge, other places, and no longer working there. Find me some talent and then figure out what their— I literally had it— figure out what their politics are.

Right.

Figure out their DEI status. Figure out all this stuff, their point of view. Because like literally when you co-host somebody, like you, Now, in today's environment, I have Molly Wood on. I lose like a third of the audience because they hate her lib stuff. I gain a third of the audience back because she's, you know, hates Trump and whatever. Same thing with Alex, back and forth, you. And I'm just like, just give me this overview. It's so great. It just gets it done. And then when you get back to your mobile phone, it's the same thing going on. It runs everything in the cloud. So unlike Claude cowork where you have to have your app open and whatever, it's just running all the time.

Yeah.

It works.

Claude, it can be a little kludgy trying to share things across networks, across people. A lot of times like, well, I can't see that. Yeah. I can't share that. So far with Grokbot, it is all very, very seamless and across multiple apps, collating, collecting things. I've been very impressed with it so far, so I'm excited to spend the $200.

So this is my little mini talk here. Systems over goals. What does it mean? You have a goal. You wanna have, uh, you wanna find another co-host. You, um, you wanna do more events and you wanna hire an events production person. Okay, you wanna have the best stories, uh, and you wanna find, uh, really interesting speakers for, you know, Founder University and Launch Accelerator. That's the goal.

Yeah.

Okay. What system can you build? And the system I built is for these to go do the research every day, real time, every other day, present it to somebody on my team, have that team member rank them, put it into a Google Sheet. And then we have a whole bunch of other steps we do, uh, which I won't get into cuz it gets kind of boring, but mm-hmm. If we can remove, let's say, 10 hours of research a week, 5 hours of research a week, and then start with research that's better than human, okay. Then if when we're inviting people, we can have it find them quicker, find their connections, invite them, book them on a date, that's typically 3 hours of work per guest.

Yeah.

So finding the guest, every guest equals 5 to 10 hours of research. Every guest equals 3 hours of inviting multiple guests to get them involved. You put all that together. When a guest comes on this program, it represents 5 to 10 hours. The end. And that's what this stuff— and that's why goal— defining the goal is great, but then it's fine to give your bot that, whatever one you're using, your agent, but you want to build a system that doesn't break and doesn't require a human. And the system then ensures it happens. So everybody has to become a systems thinker. What I found is creatives like yourself, Lan, Less systems thinking. More systems thinking, less taste. And so now I'm figuring out as a leader how to get the people with taste to learn systems thinking. And I've got to figure out, and I don't have a solution yet, how to have people who have systems to learn taste. And that's going to be the final piece for me is to have that taste algorithm.

Yeah.

You know, which you and I have by default.

From many years of experience. Yeah.

At my firm launch, the pace never slows down. I'm doing my best. To just try and keep up. And we've tried all the expected efficiency and productivity tools, and a lot of these apps end up adding friction and they slow you down. There's one tool that I am obsessed with and that every person in my company gets trained on and has to use from day one. It's called Superhuman Go. Superhuman Go is an AI chat that's always around when you need it. It's context aware, so it knows what you're working on, and it works inside the apps and tools and websites that we already use. So no matter where we are in Slack or Notion, it's in our workflow. If you're trying to just quickly scan some notes before a big meeting, check your travel dates before you set a meeting, or hey, you need to get your grammar correct, or you want to have your emails nice and tight, all of that, Superhuman Go can jump in and help without needing a ton of catch-up or permissions. You know, I was an investor in Superhuman, Raul's company from back in the day, Grammarly, Coda, that's all come together to become Superhuman Go. Find out more, you're gonna just go to superhuman.com. I love this product and I've loved it for years.

I think it's interesting. It's the same kind of conversation we have a lot of the time with AI founders, especially on the This Week in AI show. It's, it's, it's about where does the human belong in the loop? We all get that we need a human in the loop. AI can't do it all by itself yet. Maybe one day soon. But it's like, that's the strategy is like, where does the human need to be there and where can you just kind of sit back and let the bot go? And I think it changes all the time.

Yeah, it's far out. It's, it's in the reinforcement learning phase. And so like I have an SDR bot that looks for who's sponsoring other podcasts, other events in business, in categories. And then it looks at our CRM system. I think we use Pipedrive right now, although we'll probably move to just writing our own CRM system. Sorry to Pipedrive, but that's what my team is advising me to do right now. I'm like, okay. So I'm gonna have to have a conversation with Pipedrive. I don't think we're a major customer for them, but—

Nah, they'll be fine. Their Pipedrive's gonna be just fine. I believe in them.

No, I wanna talk to them about like, hey, Is there a way for us to keep using this? You know, and does it have a good enough API and agentic stuff? So putting all that aside, I need to get like the sales team to look at what the agents are doing and give them feedback. And I always get resistance. Why do I have to do this? Is this necessary? I don't like to work this way. I have another modality of work. And I said, guys, let me just stop everybody right now. I have a new idea. This is a new idea, Elon.

Get ready to quit.

Pretend I'm the founder and CEO of the company.

Pretend.

Pretend, yeah.

Just do it.

And that I get to decide some things. And one of the things in this, you know, what do they call it in the Marvel Universe? Like a different timeline?

The multiverse.

Yeah.

Like in this version of the multiverse, I actually have authority to make a strategic decision inside of the company I own 100% of. Yeah.

We're in Universe 420 or whatever.

Yeah, in 420. Yeah, Wolverine has the yellow costume and I'm the boss and I get to make decisions once in a while.

Yeah.

And I'm in charge of my own company. So they're old school.

Our sales team, they're old. They're Willy Loman guys. They just want to get out there with the briefcase and the hat. Door, door, door.

They're just literally going to the shoe repair shop.

Yeah.

And the guy's like, oh yeah, give me, give me the shoes. I got your other pair here.

Yeah.

I put, I put the steel on them this time. They're not going to wear out.

Yeah.

They're dumping the dirt on the carpet. and vacuuming it up, you know, putting their foot in. Yeah. Yeah.

Might I interest you?

Yeah.

All right. We got some news on the docket.

We got—

there was the Harvey story and then here we go with the IPOs again. But those are the 2 I wanted to cover.

All right. Yeah. Well, let's start with OpenAI. They apparently— we're now told they're going to be a public company by 2027 or even sooner. In an all-hands on Wednesday, CFO Sarah Friar told staffers the AI lab plans to go public next year. Maybe even this year. Here's— I thought her framing was really interesting. She frames it as the IPO, it's not a finish line, it is a milestone. It's another fundraise. She also noticed that the— noted that the company has flexibility as they raised $122 billion in cash back in March. OpenAI confidentially filed their prospectus in June, but they're in a tight race for users and attention with rivals Anthropic. The latest volley OpenAI now pledges not to retain data, Jason, from businesses that are using its models. Here's the quote.

Interesting.

We've heard very loud and clear from businesses that this is important. They often have their own commitments that they have made to their customers. If you think about somebody who is serving other enterprises who are trusting them with incredibly sensitive data, that's Aaliyah House, OpenAI's head of product policy.

We all know they're growing like weeds. What we actually need to know is what their churn is. Who's using it less?

Hmm.

Who is hitting the brakes and then why? And that data is not coming out, and it's not even coming out in the IPO. They're never going to share the churn data. When companies become very robust, they kind of get enough institutional holdings that a company like Netflix or Verizon has to start reporting on churn. I think these guys are going to be able to— Anthropics of the world, you know, Groks of the world, OpenAIs, Anthropics— they're not going to need to give us how many people are unsubscribing or how those trends are going for the major accounts.

Right.

That's where I want to know the truth. The second place I want to know is how much does it cost to serve this stuff up?

Yeah, well, that's—

You know that funny Italian guy who was on the show with the bad teeth laughing hysterically?

Right, laughing hysterically. Yeah, the meme. I saw you shared this meme.

Yeah.

And I shared this meme. You can pull it up.

I'll type it in.

It's literally the guy doing the bit, and he's laughing about— he's paying $20 a month or $200. He's paying $200 a month for Anthropic's pro tier, $2,400 a year. And you can just play it in the background here.

Yeah, I'm pulling it up right now.

It is—

here he is.

I run Claude code all day. Agents, the subagents, the works. Okay. End of the month, I check the meter. Check the meter. He slams a hand on the thing. $8,000 of computer.

Yeah, the audio is he's like die. He's almost like expiring from laughing so hard.

So crazy. $78,000. Sorry, that's how he's screeching. Yeah, he's like then he goes on and on.

Yeah.

I followed the money. Bottom rung, Anthropic. And then he goes on, I spent 71 cents of compute per dollar, $40 per dollar on me. And, and then he goes back to like Nvidia, Microsoft, and Nvidia, Microsoft.

Yes, all the way up the stack, all of them paying one another.

Yeah, that's when this whole thing is gonna have to face the music, right? Uh, we're getting close to the face the music moment. I think because they know they have to face the music and explain profitability, they'll have a reasonable story. But there's a gap here between how much, you know, people are using and how much they're spending.

Yeah.

And then there's some people token maxing. But I believe people now are watching the meter. People have started watching the meter. So this is exactly what I saw up close and personal with Uber. in the price wars between Lyft and Uber, and then there were other competitors like Sidecar, and then DoorDash, Grubhub, Postmates. They were, you know, talking about this incredible growth. They were moving to every city. But then we knew internally, hey, um, across these companies, we're losing $5 a ride.

Right.

And there was 2 ways to understand this. Like, we're paying drivers a $500 bonus if they hit 100 rides a week. So if they do 100 rides, we'll give them $500, an extra $5 per ride. But that wasn't in the riding data. It was like a marketing bonus.

Sure.

But then Lyft was giving $600 and then the drivers are smart. They're like, oh, you guys are idiots. I'm going to— next week I'm doing Lyft. The week after I'm doing Uber. I hit my incentive. I flip to the next person. Okay. Yeah. YOLO. I'll DoorDash.

If it's possible to figure out a system with these gig apps, people are going to work the system for sure. And one person figures it out and it's on Reddit and now everybody's working the system.

Exactly.

Bingo. So what happens is over time, there's this famous Deirdre Bosa moment, and we should play it in a future episode, whoever This Week in Startups archivist is, where I said to her— I'd love to hear that. Deirdre, you asked me the question, let me answer it. And she's great. And she's now doing her own spinoff and going to be a media entrepreneur. So congratulations to her. She is like, but they're money losing, they're money losing. I said, okay, let's look at the quarter. There was a billion rides. incredible. They lost $2 billion. That's 1 billion rides divided into $2 billion is $2 per ride. It was like $6 per ride or $7 per ride everybody was losing.

Yeah.

You pay $7 for a ride that was like, this should be $30 in a cab or, you know, $15 in a cab, whatever it was. They were losing that money to kill their competitors, build their base, attract investors, and essentially do marketing instead of giving money to the network TV shows. The concept was, well, why don't we just give the money to the customers in the form of a discount, have them become addicted to this, right? And then we can slowly raise the price up to what it actually needs to be. We'll lose 20% of the people who will only use this service because of the discount. We'll keep the other 80%. Game over, right? So venture capital funded it, and instead of giving it to radio, TV, cable, magazines, newspaper ads, The internal discussion was, yeah, just pass this, just give it directly to the customer.

Sure.

That's what's happening right now with these enterprise products. That's gonna unwind.

Yeah. I mean, with Uber, that makes a lot of sense. Like we all did get addicted and got used to like, if you go out night out, you don't wanna drive drunk home, just, you know, use Uber. We all, it got drilled into everybody's brains and now it's there permanently. With AI compute though, there are these alternatives models. There are other things you can do.

Highly competitive.

So it doesn't feel exactly the same. Like, they're not— Anthropic's not killing all of its competitors and embedding in like, you gotta use Claude. Like, I'm perfectly happy with GrokBot if it works better, you know? Like, I don't think there's—

And you will be perfectly hot, uh, perfectly okay in all likelihood when Zuck comes out with something similar to GrokBot, the easy-to-use assistant.

Sure.

That just abstracts everything, let alone when new competitors come out and you can go on Amazon Web Services, Google Cloud, and they will provision a bot. And a harness, and there'll be an open source harness like WordPress. It's just gonna all come apart.

Exactly.

And it's gonna drive the, the margin out of this. Logs are an essential part of just about any tech startup. You need to keep your eyes on how your product is being used, and you definitely need to understand what happened when things go wrong. But logs are notoriously messy. It's especially hard to gather the insights you need when your logs, errors, and performance data all live in different tools. But now there's a solution. Sentry. Sentry's logs are trace-connected and structured, allowing you to follow everything clearly and understand the context, even if you're a non-technical founder. Whether you're debugging your frontend, backend, your mobile app, whatever it is, Sentry will give you the context you and your team need to get the problem fixed and to get on with your day. You got other things you gotta focus on. That's why more than 4.5 million developers are already using Sentry, including Amazing high-profile teams like Disney+ and Anthropic. Learn more by going to sentry.io/twist and use the code TWIST to get $240 in Sentry credits. That's S-E-N-T-R-Y.io/twist. That's what I'm looking for. There was a moment in this story where you were talking about— give me the quote one more time about, we hear loud and clear.

Oh, yes.

That people are— and then segue into the Harvey story with that.

Yes. Here's the— it's from Aaliyah House. She's OpenAI's head of product policy. We've heard very loud and clear from businesses that this is important. They often have their own commitments that they have to make, that they have made to their customers. If you think about somebody who is serving other enterprises who are trusting them with incredibly sensitive data. And so that does segue very neatly into our other big story today, which is Harvey, the legal AI company, released their very first in-house proprietary model. It helps Harvey Software take on more tasks that were typically performed by lawyers. It is, Jason, a post-trained open weight model built on top of KIMI-K3. So we did, we did, as you, as you requested, we did pull some clips. This is sort of what you've been predicting all along, which is application layer companies, companies that have this incredibly valuable expert data. They don't want to turn that over to the Anthropics and OpenAIs of the world. They're going to use it themselves to build their own models.

Yes. And did you say you have a clip?

We have 2. You pick your pick. We have a twist clip on this and we have an all-in clip on this. Dealer's choice.

All right.

I'll take a victory fap. But let's keep it short here. Play it at like 1.5 speed. All right, Jake, let's get moving through this.

Let's take a look. Well, yeah, here we'll pull up the, the, the twist clip here.

I believe first. Thank you for putting it on 1.5. Need a little bit of a haircut there. You need to start working on frontier model. You need to get off the frontier models and use open source ones and own your content and not educate them to the extent you can. And I believe that'll be the trend of 2027 is startups are already doing it. But if you give Sam Altman, who is a, you know, sharp elbow guy, and he's got to figure out how to fill in a $1 trillion market cap, he's going to do exactly what Anthropic did or Microsoft or Facebook, which is he's going to look at the applications coming in, all of those Y Combinator companies who give who take that deal, they're studying every one of their token usage, they're studying what they're doing, and then they will pick the top 5 in terms of success and incorporate it as free product into their platform. This is your final warning. Don't trust the platforms. When somebody comes to you with free tokens, you know, free anything, there's no free in life. There's no free beer, there's no free pizza, there's always a price. And then the second one's from All In. Let me do a victory fat bear.

Oh yeah, this one was right before July 4th, actually. This was the July 4th All In, as you could tell from your casual, laid-back, uh, Honolulu vibes in the clip here.

Oh yeah, that's when I had my beard, and everybody loved this when I was in Hawaii.

I gotta say, I think I think this is a good look. I'm not just trying to butter you up. You do. I think this is okay. I like this. This is a good J-Cal. I think you should go with this one for a while.

What do I have, like a Tom Selleck type situation in Hawaii? Hawaii Five-O?

Yeah, it's a little bit. It's a little Magnum P.I.

Yeah. Okay. Magnum P.I. To the platform. There is no free pizza. There's no free beer. When somebody like Sam Altman comes to you and says, here's some free tokens, your alarm should go off. Zuckerberg did the same thing. He said, hey, I'm gonna give people a bunch of access, gonna give them money. Come to the Facebook platform. Nobody who went to bed with Microsoft in the '80s, Facebook in the 2000s, or Sam Altman now In the 2020s did not wake up with their throat slit. This is a message to founders. If you partner with any of these people, they will slit your throat and take your business wholesale. There is nothing to discuss here. Don't trust them. Use your own models. Yeah, we'll link to it. You guys can go check it out. Um, there's lots more.

These are just the 2 that we could dig up 10 minutes ago.

I mean, this is the, this is the system I want to build next is like, I need to have an archive of TWiST and an archive of All-In with all of my opinions. And then I want our agent to look at how right or wrong I was and give me 2 lists. Here's where you blew it. Here's your Professor Galloway moments. This is your Professor Galloway.

Your Matt Monee Jim Cramer moments.

Yeah.

This is my— yeah, my reverse Cramer. My Professor Cold Takes. You totally blew it. And then here's your best Jason.

Like the on fire.

Nostradamus. A Nostradamus Index.

Yeah.

And a Professor Galloway Index. Give me those 2.

I do. For This Week in AI, I do have a full, like, a clause skill that we've trained on every episode, and I can just go in and ask it anything, and it's brilliant. It is a little time-consuming to set up.

But I want the bot to go through and say, give— make a summary of Jason's opinions, his hot takes from each episode, and then have an index just of the takes, not the full transcript.

Right.

So then once you have the takes, then take the takes. How many times he said it? How did his take change over time?

Rate the takes. It's a take rater.

Rate the take. Then I want you to build, how did it change over time? So did my opinion on Apple, Facebook, whatever change over time? What were the key moments? You know, was I hypocritical? You know, how did it change? Right.

Yeah.

Then based on the last 6 months of news or 3 months of news of what actually happened with Apple's product, did they ever launch a new iPhone? Remember, I was very critical of Tim Cook for a long period of time. Stock went up. He bought back all the stock, bought back half the equity in the company, I understand, like $100 billion or $200 billion in buybacks. They never really came out with a killer product post the Steve Jobs roadmap. So, you know, right about that, no product ever emerged. Wrong about the stock price because it went. So I need that. So for my team, if you can do that, that would be amazing. And we're working on this kind of stuff. So anyway.

Yeah.

What you're seeing with the Harvey thing, I said this 2 years ago. If you can find the 2-year clip where I said when I was in Tahoe skiing, and I think you pulled it for a previous All In, and I said 2 years ago, there's a chance that open source is going to win the AI race. I said it 2 effing years ago. I think what we learned in 2023 was that the language models are starting to hit parity very quickly. And that the real value is going to be in— and it may even become commodities, and open source may win the day. So then I think the winner is folks who have the training data. Give myself a pat on the back that you heard it live. That clip was like— I was shocked by that. I was like, really? Like, in the age of ChatGPT 3.5, 2.5? I said on the pod, filibuster here a bit, I said on the pod, open source is going to win the day. And we, I think we talked about that 4 weeks ago on All In or 5 weeks ago. That was like a very interesting moment because that's what's happened now. The evidence. So back to my, like Jason's takes and just holding me accountable to my takes. Am I Professor Ice Cold Takes Prof G and Macy's is going to beat Amazon and Uber and Tesla and Robinhood are going to zero? And buy foreign stocks, not US stocks, like all those terrible takes. I need to know when I had a bad take and when I had a great take, like a Nostradamus-level take. That was a Nostradamus moment for me. Um, and anyway, we'll, we'll, we'll find it for them.

We'll find it for the VOD, folks. I don't know if I can pull it up here live.

So I had mentioned when I talked to Lovable and ElevenLabs, they're huge, huge customers of Frontier Labs. Frontier Labs can get competitive with them, just like Figma and Design Cursor and Claude Code, Figma and Claude Design. They are— I kind of asked them on those interviews when I was in Paris in July, hey, are you making your own models? Are you doing that kind of stuff? And they're like, yeah, you know, we're looking at that. Here it is. Harvey, pull up the story. Harvey is making their own model. They forked an open source model. They're doing their own training. And they are working on behalf of their clients. This is the story.

Over the past 6 months, Harvey's research agenda has focused on 2 goals. One, building frontier legal intelligence using open-weight models, and 2, creating systems to allow law firms to build their own specialized models and their own intelligence. Today, we're sharing an update on that research effort, including initial results from our first post-trained model, which we're calling Harvey Tennant. Harvey Tennant is a KIMI K3 base that we post-trained together with Fireworks Research for long horizon legal work. In addition, it incorporates Harness improvements to make training and test execution more effective. Our initial work shows promising results for both performance and cost efficiency.

I'm here with Keith Pierce, the founder of Lightfield. They are an AI-native CRM that builds and updates itself in real time from your email, your calendar, your Slack, your meeting notes, all those things that you keep misplacing. Welcome to the program, Keith.

Great to be here.

Let's talk about Lightfield, your CRM that is AI-first. What can I do with it now that I've got it installed? Now that it's pulled in all my data, what is it going to do for me? Because no CRM system ever had AI built in it.

It's true. So the first thing we'll do is handle all of your busy work. You know, all of the emails you promised customers, all of the things that you were supposed to get to. It'll draft all of your follow-ups, send them on your behalf. It's going to go and look at your happiest companies, your happiest customers, learn from your happiest customers through product usage, feedback, sentiment, and generate you lists of new customers. It'll find contacts for you too. And then third, it'll draft communication for you to send to those folks over email, LinkedIn, phone, whatever it is. So it'll sort of keep proactively building out your pipeline. And then last, I would say, this is maybe the most non-CRM thing. It'll keep your entire company informed and what your customers want, need. and are willing to buy. Um, so it was a tighter alignment across product and engineering.

So Lightfield.app is going to tell me what I'm doing right, it's going to tell me which customers I've forgotten about, and then everybody on the team doesn't have to be a slave to the CRM filling in all these fields. Go ahead and check out Lightfield.app. Okay, so now you have to ask yourself, um, who were the original backers? Who?

Oh, I've looked this up.

The original backers—

I've I've looked this up.

Of Harvey and who were their original frontier model providers? Who were? Who were Harvey's original backers?

I'm looking at this on Harmonic.ai. Thanks to our friends at Harmonic.ai for allowing me to so quickly pull that up. A $200 million investment. Oh, that's March. Yeah, March 25th, 2026. $200 million at $11 billion valuation, co-led by GIC and Sequoia Capital with a16z, Coatue, Conviction.

Go backwards.

Elad Gil and Vaxant. Go back to the early rounds. Oh, okay. Who are—

who are they? Early rounds is where we want to go to. That— see, this is what Harmonic's going to help us with.

Yes.

If you start getting early—

Who are their earliest backers?

Original capital and original models both ran through this company.

OpenAI was the day one believer. They led the $5 million seed round and also gave the founders early access to GPT-4, which was a massive head start in building legal-specific fine-tuning before anyone else had the model. There you go, Jason. Thanks, Harmonic AI.

So thank you, Harmonic. Yes. Thank you, Harmonic AI.

There you go.

$5 million. First investment.

Yes.

One of the OpenAI startup Yeah, they led the seed round.

They led the seed round. Now here we are.

It was the OpenAI Startup Fund. Uh, other early backers included Jeff Dean, head of Google AI, by the way.

Okay, now let's just pause here, put the 2 stories together. Sarah Friar, CFO of OpenAI, says, hey, we hear you loud and clear. Harvey's saying, hey, by the way, um, just want to let everybody know, 6 months ago, we started using these open-source models. We started building our own, and we're building our own for our clients so that our clients don't share the intelligence.

Mm-hmm.

So if you're a litigation firm A and your top competitor, litigation firm B, so ACME litigators and Delta litigators are both using Harvey, let's say. Harvey now is like, not only do we not trust OpenAI with this, We need to have our own model. We need to build a fortress. We're gonna— and we'll have our castle. It's gonna have a big wall around it. We're gonna build 2 keeps. One keep for ACME litigators, one keep for Delta litigators. So y'all are gonna get your own litigation data, all your internal memos, all your emails, all of your Zoom calls about this. We'll feed your local model for you, but it's not gonna make it to our castle. You're in your own keep. Defend your keep. And the 2 keeps will not share data.

Right.

This is what's going to happen to the frontier models when they go public. You're going to have this headwind of which of these customers— Harvey was probably spending, if I had to guess, I'm going to say $10 million a month with OpenAI.

Wow.

I would say Lovable or ElevenLabs or, you know, Cursor, those level of companies also probably were spending $10 or $20 or $30 million a month month with the OpenAIs and Anthropics of the world. Those customers, let me be clear, are going to move their spend. 99% of their spend is coming off the frontier models. And any spend they do have with the frontier models is likely, even if it's like 5% of their spend, to be them doing distillation. It's going to be them doing distillation. In other words, they're going to be spending $5 million a year or $1 million a month in order to pull information out of the frontier models Open source is going to win it all. Open source is going to win it all.

There you go.

I'm saying it now definitively. 2 and a half years ago, 3 years ago, I said, you know, open source could win it. Now I'm telling you, open source is going to win it all. What do I mean by win it all? The majority of tokens, the overwhelming majority of tokens in corporate America will not be on the frontier models. It will be inside Those enterprises. Why? They don't want to give their intelligence to somebody who wants to build the final company, as Gavin Baker discussed two weeks ago on All In.

Right.

If if Anthropic's gonna pop off at cocktail parties or internally telling people, and I believe it's true, they say it's not true. I believe Gavin.

It sounds like something they would say. To me, it sounds like something I can imagine them saying is like. Well, if we're being honest, there's a good chance we're gonna be the last private company on Earth. Like, I believe it. I believe it.

Sure. Why, why wouldn't there be a— I mean, there, there was a final company in WALL-E. It's a great science fiction premise.

Buy and something.

Yeah.

Buy and grow.

I don't remember what it's called. Yeah. Buy and grow, whatever. So it's like the Walmartification of everything. Buy and grow.

Thank you to Jacob for that. We got a guest. I wanna bring up our, our first guest.

Terrifying.

Uh, he's, he is the co-founder and CEO of the AI-powered free Speaking of free things that are free, the free AI-powered voice dictation app Willow. Please welcome Alan Guo. Thanks for joining us, Alan. Okay, we made we made Alan wait, but Alan is here now.

Welcome to the program, Alan. This is a whisper competitor. Free is the model. It's free to use. Well, then how do you make money?

Hmm.

Yeah, I'm happy to tell you more, Jason. First of all, I'm excited to be on this podcast because I. I heard that you're a big voice user.

Yes.

And I heard that you have a pedal.

I have a pedal. I'm addicted to, uh, WhisperFlow. I use WhisperFlow constantly. So that's your competitor. That's who you got to displace now. Let's see what you can do here, Alan.

Yeah, and let's throw a period of time. I'm gonna have to convince you to use Willow.

Well, I mean, the other thing that I will say, Lon, I don't know if you noticed this, my Grammarly and my Notion are all trying to get in on the dictation game. So they also have dictation offerings, and they're continuing to try to intercept this very important function. So, Alan, are you going to demo the product here for us and tell us, you know, your innovation?

Yeah, of course.

All right, demo or die. Here we go.

So as a quick intro on what Willow is, we're a voice dictation product. We're a voice-first interface for modern work. Uh, similar to Whisper, we have a dictation product. You can speak, you can press a hotkey, it works. You can type anywhere.

Yes.

The thing that's really, really magical, that's been really catching people's eyes, especially high communication teams, is our product called Scribe. And Scribe is a step above dictation. You can think of it like a writing assistant. You tell it what you want to say, and then Willow writes it for you using our style and context. So I'm going to share my screen here. Okay. Here's a demo. Someone said, you know, my dictation quality has gone down this week for some reason. Um, any idea what's happening with Wello?

Okay. And this is a Gmail and you're responding.

This could be like, um, you know, a customer support request.

Got it.

With dictation, as you know, you can press a hotkey and you can say, hey Sarah, this is a test. Best, Alan. And you can see you kind of dictate that. Oh, there it is.

Very snappy.

You know, less than 250 milliseconds, the most accurate dictation tool. But Scribe is, which was, which is what I really want to show you. I don't even have to word for word dictate. I can just say, respond to her with our common troubleshooting tips.

So this is like a snippet.

Oh wow, look at that.

Coming from the corporate side of the, the snippet collection. Yeah.

Yes.

And I press enter. And it magically appears. And what I can also do is, you know, let's say that didn't solve her problem. I could just highlight that again and say, can you also add my calendar link just in case those following tips didn't work for her so we can talk about it live? And there you go. And press enter there. And it even knows my calendar link.

Very nice. Very nice. Well done.

The power of Scribe here is that it has context personalization understanding. And this is something we've deployed across enterprises, airlines, customer support teams, sales teams, where Scribe has an understanding of your knowledge base, of your context. And so teams right now are no longer even typing. They're not even dictating. They're just saying, respond to this, respond to that. Um, tell them X, Y, Z, and it's boom, it's there for you.

Who manages the canonical level of truth? Because I have been having this problem for a long time. I used to use like a text expander. Like open source tool. It was like shareware. And then I put it on Groupon. It was, remember when I had the salon? It was like $10 a person in the company.

Absolutely.

Yes. So I was spending $120 and I would like, I would write, hey founders, thank you for applying to the Launch Accelerator. Um, you didn't make the cut, but I want to make sure you understand you can send us your updates to updates@launch.co. We want to encourage you to do this. You can watch This Week in Startups. Here's some codes for free credits from some of our partners. Yeah, I had a whole thing. But I wanted consistency. Then I told Raul at Superhuman, please build snippets into the product. He eventually did. And then I said, I need it to be team player because I don't want mixed messages. So the challenge with what you're doing, I think, is if you have 10 people working in sales and you want them all giving people the same language here, right? So how do you manage? Which one is the canonical perfect version?

Yes. So with customer support teams, they commonly have a number of documents that are the source of truth that we will understand for them. And then we also have another aspect of this, which is auto-learning. Over time, Willit picks up facts about yourself that you can write and then it learns them so you can rewrite. I also have a perfect example for you, Jason. Yes, perfect. It matches with what you just said. That you have to use a snippet for responding with who you are. Um, this person just happened to say, yes, uh, heard you might be joining This Week in Startups. Can you tell me a little bit more about this guy called Jason?

Oh boy.

What I'm gonna do here is I'm gonna say, yeah, um, that's great. You heard that. Can you, can you tell John more about who Jason is?

Here we go.

Yep. Here we go.

Wikipedia page written by all my haters and jaders.

Maybe it's just Wikipedia. You don't know.

A Silicon Valley entrepreneur, angel investor, podcaster, best known as a third investor in Uber.

Yeah, third or fourth. Perfect. Yeah, previously worked at Weblogs, sold it for $25 to $30 million. That's actually a true— that's a true number. Uh, later backed Robinhood, Thumbtack, and Trello. True numbers. Today he runs Launch House. This week he started Coast. Okay, yeah, perfect.

Very—

so yeah, this is great. Good job. So you, you put a, um, a level of interpretation and you still have to hit the send button Great job, great product. Um, and so free to use, but then if you want to do multiplayer mode, that's when you turn on pricing, I guess?

Yes. So what we've realized is that dictation is a commoditizing market.

Okay.

There's no— always difficult place to be as a founder.

Exactly. There's no doubt in my mind that within 2, 3 years, Apple is going to get 70, 80% as good as dictation tools like Willow Voice and other ones, AI dictation tools. And the reason is because, um, it's, it's, it's, it's going to be better built into your device, but it's only going to be 70, 80% as good because there's a level that we do that's more pro. Like we have auto learning, we have fine-tuned models that are smarter, that do corrections, that do personalization. So we're always going to be a bit better, but when it's 70, 80% as good, a lot of people might not pay for this.

Mm-hmm.

And so rather than, um, getting eaten by this commoditization, we want to commoditize it first.

Okay.

We made speech-to-text that's faster than Whisper Flow, more accurate than Whisper Flow and other dictation tools.

Okay.

Completely free.

And we actually launched a comparison video, one showing Willow, our free version, and one showing other dictation products. And Whisper was better at formatting, better at correction, faster, more accurate than all those other ones. And our paid product is Scribe. So Scribe again is like a writing assistant, writes for you as well as, you know, team plans, enterprise plans.

As they're saying, well, this is the playbook, Alan. Uh, is this your first startup?

It is.

Yeah. So you learned one of the most important lessons. Number one, fight up. You want to fight with a competitor who's bigger and established. And 2, your margin is my opportunity. So, you know, Whisper's charging for this, we'll make it free and we'll charge for something else, et cetera, et cetera. Great job. Good luck with it. Uh, and I will give it a try. And I will find out with my pedal if you are better than Whisper. And I'll let you know if you can displace Whisper. That's going to be hard. I'm kind of addicted.

You're a big Whisper user. Heavy Whisper user.

Heavy Whisper. All right. Thanks, Alan. Where can people find out more? It's willofvoice.com. Am I correct?

Yes. willofvoice.com is where they can try it out for free.

All right. Well done, Alan. Congratulations on your success here. Again, this is the best thing possible for the ecosystem, Lon, and for entrepreneurs. There's 2 tactical, practical things I'm going to tell you right now. Number one, you fight up. So, if you are Apple, you're not talking about Whisper Flow. If you're Whisper Flow, you're not talking about Willow Voice, but you can fight up, right? You can fight up. So, if you're Willow, you want to fight with Whisper. If you're Whisper, you want to fight with Apple, right?

Right.

Fight up. And then, if you're the bigger fish, never engage going down.

Yeah.

Because you're taking the bait. You're taking the bait. You're feeding into it.

Same rule.

It's the same rule in comedy. Don't punch down, you punch up. You go for—

You punch up.

You go for targets that are, that are higher and better established than you. You don't make fun of people who are already lower status than you.

Yes, a very simple concept here. Um, and then second, your margin, my opportunity. So you just look at the roadmap. It's very simple to do. If the roadmap for dictation starts with with crisp dictation, then it goes to interpretation, which is kind of what he's doing here. And you have this killer feature known as snippets. What should you do? And then you have team player mode. Just make everything free. So if Whisper's watching this, they should just say, oh yeah, by the way, we added to our free version, you know, snippets and text expander tools. And here's like our robust one. What's really interesting to me is when these are all going to be done locally, Sure. And built in with like very crisp local models. Right now, they still go to the cloud. So when I use Whisper, man, it is flawless. It is tremendous when I'm at home. When I'm driving in the backcountry, in Hill Country in Texas, the great state of Texas.

The great state.

And I'm going to watch the Knicks beat the Spurs in their own arena, win the NBA championship. And I started using Whisper and I had no connectivity.

Yeah.

You know, it's like, oh, sending it to the cloud, coming back.

I really think it's Hill Country. There's a lot— your connection isn't always perfect out there. It's not always perfect in Ransomware.

Well, you know what I'm gonna do now is, um, when I finally find my roadie, I'm getting a roadie to go with me on the road because I have so many road things that I do.

I get my roadie—

Heavy things.

I'm on a lot of the trips, but I'm not good with the heavy stuff.

I mean, it's— yeah, given where you're at with ro.co/twist, maybe, maybe, maybe sometimes. What I would like to do is I'm going to get the Starlink Mini, like on-the-road one. And I'm gonna— I want to get this executive van. I love the Alphard, which I think I drove with you when we were in Japan.

The Alphard. It's like, yeah, when you take a car, you take an Uber in Japan, they have these little commuter vans that they pick. Yeah, Jacob, put it up. Jake. Yeah, I'll find one. They're, they're so neat. Like, I feel like I would actually like drive one of these if you could. Yeah, the Toyota Alphard.

I mean, it is the The Toyota Alphard in the executive mode is like a minivan. There's Juan on his iPad Pro.

Exactly.

In his suit, looking good. Look at how big these seats are. Unbelievable how gorgeous.

It's very comfortable.

You feel like you're—

Rides smooth.

You've got a nice sort of distance from the driver. They're in their world doing their thing. You're in the back doing your thing. It's very comfortable. I felt right at home in the Toyota Alphard.

Why don't they have these in America? I literally tried to import one. You can't import them. I think they're getting there. Genesis GV90 EV just dropped. Literally, this just dropped, and people are losing their mind over it. Let me show you this. So I have been— I was literally going to import one of these and pay like a 50% import tax. You're not allowed to.

Yeah, that sucks.

But this is the Genesis GV90 that's coming out. So what this thing is going to do It's an EV, goes 300 miles, whatever. You see the doors look interesting, right?

Yes.

Boom.

You open the doors, you can have the captain's chairs face each other. See that?

Yeah, yeah.

They spin around. This is like the old G90, but this one is something slightly different. GV90. Genesis GV90 is what you're looking for. People are losing their minds over it. So that's what I want to get is one of these bad boys.

Yeah.

And then put a Starlink on it and then I could just recapture, you know, whatever, an hour a day. But look at that.

Oh wow, that's nice.

Look at that. By the way, this— if you were to build a, uh, like, uh, Becker or something automotive builds these for $400,000, literally $400,000 or $500,000. This is going to be like $100,000, $150,000, you know, to go crazy.

It was like $150K was the— on the page you were looking at.

Now this is deceiving. You, you have like chairs facing each other. Those 2 chairs are the captain's chairs for the driver and the copilot, and this is a parked car on the beach.

What I want is an extended version of these with the 4 chairs.

Yeah, so then you could have a 6-seater. See, there's the steering wheel back here. Yeah, anyway, if you're parked and you want to have a conversation— I don't know who's hanging out in their car.

If you're living in your van, that's perfect.

No, you know, see, they just— they got it wrong in terms of like what actually is gonna happen, but It's gonna— it's coming next year. So there's your off-duty from J-Cal. Let's do our final guest, and I gotta get the heck out of here.

Yeah, we got one more guest. I will, I will pull them up now. Here we go. We actually— we discussed our next guest when their video went viral, and we— you, you wanted to book them, so here they are. He's a Stanford computer science student and the designer of the self-driving robo-taxi golf cart. That you'll recall.

Finally.

All the, uh, all the high, uh, priests of AI and tech were riding around because they all hang out on the Stanford campus. Ethan Goodhart, thanks for being here, Ethan.

All right, Ethan, what's going on? You're, uh, you're an undergrad, you're grad, what are you doing over there in Stanford?

Uh, yeah, technically undergrad, so I'll be a junior next year.

All right, incredibly hard to get in. You're a sophomore, incredibly hard to get into Stanford. So you're either brilliant or your parents built a building or you, uh, Photoshopped yourself playing, uh, Lacrosse. Uh, which one of those 3? You're just a brilliant kid. Where are you from?

I can definitely confirm that it's one of those 3. Um, all right, no building, unfortunately.

No building. Okay, narrows it down. So did your parents Photoshop and do generative AI to put you in a lacrosse video?

Uh, I think the models were not—

not quite there yet. Okay, so you earned your way in. Where are you from originally?

I'm from Atlanta originally.

Oh, all right. Well, that's good. See, I actually know people who are moving to smaller states You want to know how crazy these parents are? And Ethan's nodding. Maybe his parents left Manhattan for Atlanta to get him into Stanford. Some of these lunatic parents look at the tables, Lon, of which states, because they want to have representation from every state, which state has the least number of—

Wow.

And they'll literally move to Nebraska, ruin their lives for 2 years, 3 years.

Georgia's got to be doing pretty good. I feel like Georgia's competitive in terms of like, West Virginia, that's where you'd go. Idaho. Those are—

Ethan, it's true though, right?

Those are your Stanford states.

Yeah, I mean, that's not the reason why we're in Atlanta.

No, but, uh, but you have heard this.

Yeah, yeah, I have.

Yeah, this, this is the lunacy. Okay, so, uh, what are you, computer science, electrical engineering double? What are you doing?

Yeah, pure computer science. Uh, this is my first, like, real hardware project.

Got it.

So, uh, self-driving Well, you know, Lon, my position has been 20 different companies of note are going to get there in the same 24-month period, putting Waymo out of it, just starting this year and next year into 2028. My belief, Ethan, was 20 people figure it out. I think I'm good. Like, in the same 24-month period, we'll see 20 people figure it out. And when I saw Ethan's project, it said to me, I might have underestimated it, by a magnitude. It might be 2,000 people figure it out in the same 36 months, just to extrapolate a little bit. Tell us what you built, Ethan, how long it took to build, and just the fidelity of it versus the brittleness of it.

Yeah, for sure. Uh, and it probably the best project I've ever worked on in my life. Uh, first hardware project, we got the first prototype up in like 2, 3 weeks. So we had a golf cart that could drive with some like very simple models. Um, obviously not as advanced as a Waymo or a Tesla, but, uh, yeah, we've kind of been working on it for a couple months after that. And then—

This is a standard golf cart and you built some kind of rig to go on, or a harness, but not a harness. Yeah, it's a physical harness that's going over the steering column.

Yeah.

Yeah, exactly. It's, it's so it's retrofit. Um, we didn't modify any of the like internal electronics. This is all just placed on top of the golf cart. Yeah. Um, so it's like 3 front cameras, 3 back cameras, and then—

No LiDAR, all cameras like Elon.

No LiDAR.

Okay.

Yeah.

Um, and so you take a stock one, you put cameras on it. What is the hardware stack? Because, you know, Elon's been working on this for a long time. And building essentially the compute locally on the car to do exactly what you're doing here. So yeah, I see there's kind of a box on the floorboard, so I'm assuming you have an NVIDIA Spark or something down there doing this.

Yeah, we, we have an NVIDIA Jetson Thor that's wired directly to the battery of the golf cart.

Got it. So that uses a decent amount of power. Um, and the golf carts are notorious for going, what, 50 miles in a day, 30 miles before charge? So how do you deal with that? Do you buy a bunch of Anker batteries and just bolt them in?

So right now, the, the range is a little bit limited, so it's definitely something we have to figure out, uh, in the future. But it's still very much in like testing phases.

On the top, you got a Starlink, I see.

Um, Starlink Mini, that's right.

Fantastic. Uh, what is that doing?

So that allows us to do 2 things. It allows us to test out cloud inference. So we're exploring running some DLAs and world models in the cloud. And then it also allows us to have good connectivity to write code and ship code live to the golf cart while we're driving around campus.

Got it.

So like, I often work on it outside of my dorm. I spent a lot of all-nighters in spring quarter working on this, and I can like ship code live While I'm on the golf cart instead of having to like go back into a dorm room or classroom somewhere.

Okay, I mean, listen, it is— autonomous is on golf courses. Uh, people don't know this, but a lot of the, uh, grass being cut right now, Lon, and even in, uh, snow country, a lot of the snow being moved is being done by autonomous robots. The team will go, um, find a video or two about this, but even here in the great state of Texas, Ethan and I drive around And I see in the planned communities robots cutting grass, um, and that's in like Hill Country. That's here today. They're doing a lot of this on golf courses. Uh, you made this as a proof of concept, but you know, you're right next to Sand Hill Road. There's so much money sloshing around. Is the idea to like go to golf courses and say, hey, we literally can help you bring the golf courses back? So here's the snow one, check this out. Anything This thing's incredible.

That's awesome.

Um, and so, you know, snow blowing is— one year when we had a lot of storms up in, uh, Tahoe at my ski house, I got an incremental $12,000 bill because they're paying the people who shovel snow $70 an hour. They're literally getting paid $150,000 a year, Lon. And we had so much snow that it was going to collapse the houses. Therefore, there was no choice but to outbid everybody to save our houses. Anyway, these things are becoming standard operating procedure. Here's one on a golf course using Toro, Tofpro, Anonymous. And these things are really easy to set up. People love them. What's your plan? I mean, are people asking you to build this for golf courses? Because you might have a situation where somebody stops And leaves their golf cart around, or they finish, you want to drive it back to get charged, whatever it is. Do you see this as a business?

I think it's uncertain what the future of this was. This was just a fun side project, which I think it's absolutely incredible that something with the scope now can just kind of be a side project. Uh, the future is uncertain, but there's definitely been a lot of interest from people in the golf world, also people just in the transportation world. So like people like Disney and other companies that have a lot of golf carts for transportation. Um, I think there's a lot of use cases for golf carts.

And yeah. What can you, uh, if you were to iterate on this twice and get, you know, obviously this is like a custom sensor pack, but let's say you had $10 million in funding, you do, you get to, you know, oh yeah, number 500. What do you think it would cost to retrofit these? These golf carts cost $10,000 to $20,000 depending on how fancy dancy they are. What do you think you can get the retrofit to, to have autonomous golf carts in the real world?

Yeah, even the current version is quite cheap. So I won't say the exact number, but—

Give us a ballpark.

Yeah, where you could see a retrofit version of a kit like this where you could sell for under a couple thousand dollars.

Amazing.

Wow.

Incredible.

Yeah.

What software did you use? There's been a lot of talk about NVIDIA and Uber and this open-source stack. that they're providing because they want a— Uber obviously strategically wants open-source driving. They want a fragmented market. They want every single car to have it so that Waymo or Uber or, you know, pick your company, Noro, Pony AI, Wave, AV Zoox, don't run away with it. They want it to be fragmented. And Nvidia wants to sell compute hardware. So it's good for them if the software Abstracted. Are you using their stack for software, or did you write your own, or are there other open-source ones? Tell us about the self-driving software, the brains.

Yeah, the software is by far the hardest part so far. Um, we've ended up using our own software, and so we found that a lot of the open-source models, surprisingly, were either way too large or didn't generalize well to the environment that we're driving on, um, because it's a little bit different than the environment that most cars are driving on, given pedestrian density, and temperatures are very different. So there's also the question of the models being too big. So most of the models that have been published by NVIDIA so far are actually quite large and hard to run on even the $4,000 GPU that they have.

Hmm.

And so we ended up using our own approach and custom Love it.

Listen, continued success. If you make it into a company, just save a slice for your unc, J.Cow. You know, I'm good for like, you know, if I get on the cap table, I'm always good for like a ticket to, I don't know, an All-In Summit or something. I mean, it might be good for a retweet.

You never know.

So this is what I got to do, Lon. I got to set— I'm like a 55-year-old man, unc. Talking with 20-year-olds now trying to get a slice on the cap table. No, I've got a way to— At what point is that desperate? Is it 65?

No.

Where it feels desperate, or—

You just graduate to becoming more and more senior, you know, like where you're—

Okay, like Vinod.

People want you on the cap table even more because of your incredible reputation. Reputation.

Yeah, right.

Okay.

All right, all right. So I get to— I'm in my Obi-Wan era right now, Ethan. That's a Star Wars reference. And then I guess Vinod is in the Yoda Kind of timeframe on the timeline. Ethan, is there a place people can find out more? Is there a name to this company or a website or an open source project?

Yeah, there's— the project name right now is Wind. So kind of like the wind that you feel blowing on your face while you're driving.

Okay.

I love it.

Great.

You understand branding. Wind self-driving golf cart. I'm going to try to find your domain. Do you have a domain?

There's no domain, but if you go to my Twitter, there should be a link to the test flight. And then if you're in the Stanford area.

Okay.

You can come check it out.

Yeah. How did you, uh, how did you get, uh, Jensen to take a ride? How did you, uh, finagle that? You had one of your props, uh, or you're haranguing?

Well, he kind of— he was speaking on campus, and so we basically, uh, waited outside the lecture and said, hey, we built this self-driving golf cart.

That's my guy. That's my guy. Ethan is learning the startup lessons.

All these guys like punk rock.

Ethan, punk rock was a style of music in the '70s. And what punk rock people did is they didn't give an F, and they would just go do stuff. The Clash. Have you heard of this band, The Clash? Be honest, Ethan.

I haven't, no.

Oh my God, this kid's killing me.

I'm gonna blow your mind, Ethan, right now.

He probably heard London Calling and just doesn't know that's The Clash. You—

and this is gonna score you a lot of points when you throw a party and you meet girls. At some point, this is gonna happen for you at Stanford. You're gonna throw a party Meet some girls or guys, whatever you into. They them, no judgments. I know it's a whole thing.

They I don't want.

But you're going to get an album and vinyl, and you're going to get The Clash London Calling. And just you just do a Google search, or you do your LLM search. Go down the rabbit hole.

The kids they want Pink Pantheress or something.

Then I want you to DJ, and I want you to just play Rock the Casbah, London Calling. And then when you meet a chick or a guy, whatever you're into, again, no judgments here. I know it's like. 2026, whatever you're into. I want you to just say like, yeah, just, I've been kind of in my class here. I've been reading. Then I want you to get the London Calling, put up London Calling. And I want you to get the original vinyl, spend like $30 on it. I want you to get a t-shirt. Don't play it. I wanted you to give me the COVID of London Calling.

We're not allowed. We're gonna get—

Well, we can show a little bit, but anyway.

It's gonna get us in trouble.

Yeah. I'm telling you, Ethan, you're going to listen to this while you're coding. It's going to break your fucking brain. You're going to be able to code 40% better listening to the Clash. Show the album cover, London Calling.

Does Jacob not know? Jacob might not know what you're talking about.

Jacob's like turning the—

I might, I might not. There it is.

Look at this cover.

There it is.

There's the Clash.

You see this, Ethan? You see this?

They're rocking in San Francisco.

Yes. You see the art direction of this? And he's smashing it. Yeah, that's you in your startup. That's you intercepting Jensen. You're punk rock, okay? You're willing to break the rules. You don't give an F. You're willing to smash your guitar on stage and just let out all your emotions. The Clash, London Calling, coming at you on a Two for Tuesday.

Yeah.

All right, Ethan, get back to work. And yeah, thanks. Is everybody cheating on their— with, uh, AI? Tell me about the cheating. Not you, but other people.

I think there needs to be a real working of the educational system. I think it's evolved to the point where people kind of do the work that they want to do. They kind of pick and choose what parts of education they want to get out of it. Um, so I think it's a lot different than how it was before.

So if they, if they're not into this course, they just AI speedrun it. If they love the topic, they're like, I'm gonna go acoustic and I'm gonna really learn it. Is that what you're telling me?

Yeah, I think for the most part that's—

Can I be honest, Jason? I was already like, by college I was sort of like that too. Like, it was pretty easy. Some of these courses you could sort of like— microbiology, I wasn't super into it. You could just kind of figure out what was gonna be—

You didn't have the tools to get an A in it. You would kind of cruise to a B.

Like a B, a B+.

Yeah, cruise to a B. These kids can just speed, you know, speedrun it either way.

It wasn't like having Claude. It was still like I had to do something. of it myself.

You know what I take from that, Ethan? The kids will be all right.

What's that?

I think the kids will be okay because here's the thing. That's what adults are doing. They're like, this job is completely effing boring. My boss wants me to do a TPS report. That's in Office Space, a really good movie that you should see. And they're just like, I'll just speedrun this, give it to my boss, write some sycophantic answer, ask my boss 3 questions, tell him it's brilliant, and then I'll actually do the work at work I actually like. Well, just everybody does that. You route around the stuff you don't like. You don't do your chores and, uh, you pay somebody to do your chores. In this case, AI. I like it. The kids will be all right, Ethan. Thanks for the honesty.

I think, I think we live in a world where every young person has access to a superhuman software engineer, mathematician, expert in medicine, physics, all at the same time. So if that doesn't fire you up as a young person, I don't know what will.

Yeah, it should fire you up. What are you gonna do with this $300,000 in debt you're going into? What's your plan there to pay it back? Well, you got some scholarships. You look smart enough to get a scholarship. How much in debt do you think you'll be when you get out of this ballpark?

It depends on how many self-driving golf carts they build, I guess.

I'm going to give you a secret. When you go raise your round, here's what I want you to say. I'm going to give you the script. Okay? Don't tell your Uncle Jason said this. Unc is going to give you some advice. You're closing that $3 million seed round. Here's what I want you to say. Listen, I want to give it all my energy. I got a little anxiety. This is like thing over here, my tuition, it's $140K. When we do the $3 million, would it be okay if I just sold $100K in secondary of my common shares to take it from $140K down to $40K so I'm not anxious about that? I can put my whole effort in. Boom, that becomes the ticket price to get in. To your, you know, there you go, your new company.

That's a tactical and practical tip, right?

There's a tactical and practical.

That's very tactical.

Here's how to manipulate your seed investors. All right, brother, good luck with everything. And, uh, yeah, when I'm next on campus— I'll be on campus in January, possibly.

Apparently this is where you have to hang out now, Jason.

Well, I mean, that's not— that's not now, that's always been.

If you want deal flow, you got to just be hanging out.

No, if you go to GSB If you wind up going GSB route, Professor Pfeiffer, Prof. Pfeiffer does a course called Power. And there's a case study in it about this podcaster, blogger, kid from Brooklyn who accumulated far too much power than he deserves.

No, no.

In Silicon Valley, and it's about me. And I go every year to talk to the students about accumulating power through media and other things. So good luck, Ethan. Hopefully I'll see you on campus.

Thanks, Ethan.

Yeah, likewise. See you guys around.

Peace out. That's like that Harvard class that Bill Gates's daughter was in that teaches you how to manipulate people and cheat.

Did you hear about that?

Oh, how to manipulate your son?

Yeah. No, apparently the rumor is there's a Harvard class that's about like how to rule the world through manipulation. It's like a secret, a secret class. And they say that, yeah, that Phoebe Gates was in it and that's where she learned.

I think she's being railroaded. I'll be honest.

Really? Wow.

Because I think there might be at first.

You were—

I think this might be like a convi— I think everybody's kind of trying to manipulate the attribution tag. Um, and so I think— Well, that's undeniable. I'll be honest, if she wasn't Bill Gates's daughter and she was a dude, I'm gonna be honest, like on 2 levels— nepo baby plus woman.

Yeah.

Plus, let's, uh, she's like a— I think she's kind of like an Instagram influencer type. So that has a little baggage with it. It's a gas table.

It's Khloé Kardashian. It's Sydney Sweeney.

It's Sydney's. People resent it.

Exactly.

People resent her for being a nepo baby, resent her for celebrity connections, resent her for her Instagram stuff. And if it had been a guy, same set of circumstances, I wonder. And I don't have all the details of it because it's a little bit like in—

We'll see. None of us do.

I'm guessing somebody might say, to the founder, that's a really clever growth hack. Trying to get the attribution tag so you get the commission. Wow, that's a good growth hack. Cookie stuffing, they call it. Well, they might be like, you should probably read the terms of services because you might be in a gray area. That might be how the conversation actually went down, in my experience, in a board meeting, because I've seen people go to the gray and have these kind of situations. And I always say to them, it seems a little gray. Maybe we should read the terms of services and maybe back off this, you know, a little bit. It might be, you know, we could get in trouble for this. But there might— is this like a criminal legal case?

I don't— I mean, I don't think so. No. And they stopped doing it as soon as the report came out that said they were doing it. They stopped doing it. So I— no, I don't think—

Anyway, I want to adjudicate this case. I'm going to adjudicate this next week. All right, we'll talk about it. I'll see you on Monday, and I will put on my robe. This is This Week in Startups, thisweekinstartups.com, @twistartups. We'll see you next time.

Bye-bye.

Bye-bye.