Sequoia's Abhishek Malani on the AI poker-bot tournament: 7,000 applicants, 700 competitors, and a Sunday final

Speaker 1: Good. And up next we have Abhishek Malani from Sequoia Capital. He's a partner on the No Limit Poker Bot Tournament. We've talked about this on the show. What's going on? Excited to dig in deeper. How are you doing? Welcome to the show.

Speaker 5: Thank you for having me. First time on TBPN and excited guys.

Speaker 1: Very excited to have

Speaker 2: you Fantastic to have you.

Speaker 1: Take us through the Poker bot challenge. How is this going? How did it come together? How many people are competing? Walk me through the high level.

Speaker 5: Definitely. So call it about a month and a half ago, was playing a lot of poker with friends, realizing these new models are coming out, and holy crap, they can do a lot of things.

Speaker 1: Yeah.

Speaker 5: And so we started testing over on the weekends. Can we build a poker bot? So poker bots have been around for over almost a decade now, but you've never been able to do it quickly. It used to be, like, a class for a full semester at MIT, and was, like, week long competitions. We started doing it in a weekend, and I was like, students way smarter than me can probably do it a lot faster. And so come into partner meeting at Sequoia one day is telling people, this is something interesting that's not possible. And immediately, Sean, Alfred, a lot of the partner are like, we need to do something about this.

Speaker 1: That's great.

Speaker 5: And so when you think about poker bots, the first person you think about is Noam Brown. He is the one who started this with his research at CMU in 2017. So first call was Noam and OpenAI. Yeah. And they were excited. They're like, let's see what our models can

Speaker 1: do. Yeah.

Speaker 5: And then we've been great partners with Citadel Securities for a while. A lot of their job is thinking about strategies with hidden information, and they've been incredible partners.

Speaker 1: Yeah.

Speaker 5: And from there, like, month and a half ago idea to finals is gonna be this Sunday. It has been an insane sprint, but over now, I think, 7,000 people have applied. We have over 700 competing. Finals is gonna have your top 20 teams from both Boston, The Bay, Pittsburgh competing, and we're pretty fired up at strategies.

Speaker 2: Yeah. What are... Like, what are all the different approaches that people are taking? I imagine some of the 700 is just like the most thin wrapper of all time.

Speaker 1: It's just what should I...

Speaker 2: But then other people are like, you know...

Speaker 1: Asking a GPT four. Like, hey. Yeah. What should I do?

Speaker 2: Give it your best

Speaker 1: shot, buddy. Give your best

Speaker 2: You're one... Most people would quit right here. They'd walk away from the table, but, you know, 90% of gamblers quit right

Speaker 1: before they're about to... The whole secret is a is a motivational prompt. Yeah.

Speaker 5: This is also true. I think the first versions of poker bots a couple years ago, you'd see these stories of the bot that shoves all in every hand

Speaker 1: and just I ends up winning remember that.

Speaker 5: No one knows.

Speaker 1: No one knows what to Unfortunately,

Speaker 5: if you push in all in every time

Speaker 1: Yeah.

Speaker 5: You almost always come close to last. Students have gotten a lot smarter, models have gotten better, but

Speaker 1: That's what 99% gamblers say

Speaker 5: just before the end it. At the lower end, you see some people trying to honestly, like, ask Codex

Speaker 3: or

Speaker 5: Claude, hey. Make me a really good bot. I have to win. Okay. Or download all this hand history. Sure. And then you get some really creative strategies, and that's the bulk of the competitors, honestly, where

Speaker 1: They're thinking about the meta game. Right? Talk through.

Speaker 5: Yeah. Yeah. The meta game. Exactly. And each variant and each regional is different because we want it to be a new challenge every time. Finals, it's gonna be its own thing. But at CMU, the team that won first place built this crazy classification model. So you come into a game and you start to figure out as fast as possible, is this player gonna be shoving all the time? Are they loose? Are they tight? And within about, like, 15 hands of data, they start to figure out what is the Geordie bot gonna play like. And if you're gonna be super loose, I'm gonna go super aggressive into you when I have a good hand and try to exploit you.

Speaker 1: Got it.

Speaker 5: And so you start to see a lot of these strategies. Some people make neural networks.

Speaker 1: Yeah.

Speaker 5: It's been pretty crazy the different types of strategies and exploits people have been making. Yeah. And it evolves over time. I forget is it...... Version is different.

Speaker 2: Are are there limitations on, like, on compute? I forget the original rule set. Or could somebody come in and just apply, a massive amount of test time compute to

Speaker 1: $10,000,000 of compute to create the Like, is

Speaker 2: there a situation where somebody spends, a $100 on on

Speaker 1: No one no one would spend $10,000,000 to win a $1,000,000 prize in the modern era. Oh, that's that's crazy. No company, no lab...

Speaker 5: There's a lot of aura. You don't know.

Speaker 1: That's true. That's true. That was the story of the Millennium Prizes. Right? Like those were those were EV negative but great aura.

Speaker 5: Exactly. But, no, there's a constraint where we have... Every bot has to respond within one millisecond.

Speaker 2: Oh, wow.

Speaker 5: This, like, prevents you doing, like, thirty minutes of test time compute before you make a decision, and you have to be able to be control under constraints.

Speaker 1: One millisecond. That seems really fast. Wow. Okay. Cool. Yeah. Yeah. But then it allows you to actually play the whole simulation out very quickly. That makes sense. Got it. What about tool... Who...

Speaker 2: So so I'm I'm curious, like, one of the one of the interesting parts of this initiative for you guys and and Citadel and and OpenAI that, you know, you're gonna have, you know, potentially hundreds of like, know, very very motivated high agency intelligent people. How do you guys divide and conquer on bringing them into, you know, funding them, hiring them, things like that? Or is it just gonna be a free for all at the end?

Speaker 5: I mean, I think at Sequoia, the core primitives of the business are people and markets. Mhmm. This is a great way to bring incredible people together.

Speaker 1: And Mhmm.

Speaker 5: Maybe I'll give a... Some of the students, like, when we're filtering on, like, who are competing. They're building Tesla coils that shoot three feet sparks. Some are building their own PCBs. Others are just developing their own, like, trading softwares that they've been doing in their high school PAs. It's pretty crazy, like, what students are capable of.

Speaker 1: Yeah.

Speaker 5: And then in terms of, like, Sequoia, we would love to be in business with founders from as early as possible. But whether that is founding a company, going to do research at OpenAI or trade at Citadel or join one of other portfolio companies, it's honestly trying to figure out how can we best support this new generation. And that's been the core mission

Speaker 1: from the beginning.

Speaker 2: Long term long term view.

Speaker 1: How do you think how do you think about tool use generally? I was watching, some viral chess Instagram account, do these very funny, chess examples where he plays against the models or has the models play each other and they're just hallucinating crazy things. So it'll just be like an extra queen spawns over here and then and then it moves like a like a like a knight all of a sudden and there's just like insane things happening and I was like this feels like a very odd model because in the modern era, if I go to any any frontier model with a proper harness, it'll just fire up Stockfish and just use a real chess engine. It doesn't need to hallucinate the board or even think about it like it's solved. And so in designing these, like, how do you think about... I mean, it sounds like you answered it with the one millisecond response time, but are there open source poker bots that the the models can build on top of? Is that relevant? Is that discouraged or encouraged? How do you think about that?

Speaker 5: Yeah. I think the one millisecond is one thing, but the other core part is the variants themselves.

Speaker 1: Mhmm.

Speaker 5: So if you wanna do standard no limit burger, there's a lot of resources available online that will help you get you a good starting point. Mhmm. But one, I think, crisp example in the Boston Regional, here's what the variant was. You start with your standard two cards, but then instead of a flop of three cards, you have a flop of five cards, and you're actually bidding against all the other bots on the table for which three out of those five cards are going to be the actual flop. And so this is a completely new mechanic Interesting. That changes how you're gonna play. If you just download a stockfish version equivalent for poker, it just doesn't know how to respond here.

Speaker 1: Sure.

Speaker 5: And choosing the flop is so important in the end outcome where you need to be able to develop the strategy in the tournament in order to really succeed.

Speaker 1: Are you competing or are you recusing yourself? Have you ever built a poker bot? Give us your...

Speaker 2: He said he did. Yeah. Yeah. Yeah. Weekend that inspired

Speaker 1: it. Yeah. Yeah.

Speaker 5: First time I built it but unfortunately not allowed to compete given I know all the secrets open.

Speaker 1: Okay. Yeah. Well, congratulations and we're very excited.

Speaker 2: Can't wait to can't wait to see the outcome. Yeah. We'd love to... For the winner, I mean, we'd love to have the winner on if it's possible.

Speaker 1: I I I wanna play the winner's spot. I I think I got a shot. I know what they've never seen anyone with so little experience in the training data. I'm terrible at poker. I never play. Anyway, a lot of fun. Thank you so much for coming on the show.

Speaker 2: Great to hang.

Speaker 1: Have a great weekend.

Speaker 2: Good luck this weekend.

Speaker 1: And we'll talk to you soon. Have a good one. Cheers. Goodbye. Let me tell you about Shopify. Shopify is the commerce platform that grows with your business and lets you sell in seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents. And let me also tell you about the New York Stock Exchange. Wanna change the world? Where's capital at the New York Stock Exchange? There we go. We gotta get some more of these soundboard going.

Speaker 2: Yes. Not on... Everyone always says, the shows... I enjoy the show but not enough soundboard.

Speaker 1: Not enough soundboard.

Speaker 2: I think that's it for today. Folks.

Speaker 1: Yes. Fantastic. Crazy week. Crazy week.

Speaker 2: Touch grass.

Speaker 1: We were sick. We got through it. We're very excited that you all are here with us and I'll hopefully have some merch. Thank you for tuning in to TBPN. We'll be back on Monday. I think we have an early show Monday. We're at tomorrow acts on Tuesday and Thursday. Tuesday and Wednesday. Thursday is GTA and I think the Anthropic IPO should be going out that day. That's going be a wild story. So tune in. Don't forget to

Speaker 2: Yeah. Crazy timing. Ring the bell. Anthropic IPO happening on International Mens Day.

Speaker 1: Yes. It will be

Speaker 2: a very surreal moment for I think Dario and you know, many of

Speaker 1: the other... You think it's about International Mens Day? I thought it was because it's Tyga's birthday on next Thursday. Think they were kind of like, happy birthday Tyga. Here's an IPO. Here's some liquidity. Maybe you're an early investor. You don't know. There's the hidden cap table. There's the... He could be in the line of successors. Right? Under the founders, there's a hidden list of people that succeed. Tiger could be one of them.

Speaker 2: You know, in a in a multi layered SPV.

Speaker 1: He also might be in a multi layered SPV.

Speaker 2: You never know. Could be a little... You never know where where that guy's gonna pop up. Tiger.

Speaker 1: You never know. But we'll be covering it. So leave us five stars on Apple Podcasts and Spotify. Sign up for our newsletter at tbpn.com and we will see you on Monday. Goodbye.... Goodbye. Have a great weekend. Good luck. Have the best weekend ever.