B (2:52)

Yeah, maybe. Maybe. We'll have to see. I mean, it still feels like, even for that example, like we need some scaffolding, we need some wrapping around it. It's not like you can't. It's not like yesterday we weren't able to do something with AI, and today in Vanilla Gemini 3, you can just do it. It's just a new functionality necessary.

D (3:15)

Sure.

C (3:15)

I think it's better.

B (3:17)

It's better.

C (3:17)

As good as we would want to expect. If it. It's not slowing down, I would say.

B (3:22)

No, no, no, no, no, not at all. It's not slowing down. It's just that it is getting better. It may mean it might be. It's growing but decelerating. Is that fair to say? Or are we accelerating?

C (3:35)

I don't actually know that it's. I don't think it's that big of a sliver.

A (3:38)

Tyler, say the word decel.

C (3:41)

Say this is a great model. I'm very.

B (3:43)

I mean, the way I agree, it is the best possible. I was framing it in somewhat of the same way as the iPhone launches. Like, it's newer, better, smarter, faster, stronger, newer and better. And it's like it is all of those things, which is good. You don't want to go backwards. But is the. But like we're waiting to see on the net new capabilities on the binary step changes.

C (4:06)

I think over the next week or.

B (4:08)

Two, we'll see some stuff.

C (4:09)

If it's actually really good.

B (4:10)

No, no, I'm not saying it's not really good. I'm saying binary before and after. Like with ChatGPT, it was like you couldn't talk to the computer and then you could. And then with the launch of Windsurf Cognition, Devin, it was like you couldn't write code with the computer. Now you can. And it was like the AI became capable of writing software and went from like.

C (4:38)

It's also like it became not good at. It couldn't write code. And now it can write kind of meteor code.

B (4:44)

Yeah.

C (4:44)

And now maybe it can write media code and now it can write good code.

B (4:48)

Yes, yes.

C (4:48)

That's also a step change.

B (4:49)

But we do have our benchmark, the one true binary step change, which is. Is it funny? We need to know if it's funny. I gave Tyler the prompt. Do you want to read the result? The stand up comedy bit from Gemini 3 while you pull that up, let me tell you about ramp.com. time is money save. Both easy use corporate cards, bill payments, accounting, and a whole lot more all in one place. Tyler, what did Gemini 3 say when you asked it for a comedy routine for stand up special? Let's see if it's funny.

C (5:25)

Okay. It's pretty long.

B (5:26)

Okay, read it off.

C (5:28)

Okay. I bought one of those smart health watches recently. You know the ones? It tracks your sleep, your heart rate, your readiness. I paid $300 for a piece of rubber to tell me I'm dying.

B (5:38)

Get the soundboard.

C (5:40)

That's all it does. I wake up, I feel fine. The birds are chirping. I'm ready to seize the day. Then I look at the app, and it's like, recovery score. 12%. Your heart rate variability suggests you are technically a corpse. Go back to bed.

B (5:54)

Okay.

C (5:54)

And immediately, I feel sick.

A (5:56)

I'm like, oh, God, so bad at reading this.

B (6:00)

I mean, like, this is terr. This is terrible delivery, brother. Like, I'm not.

C (6:07)

I'm not good at this. It's a placebo effect for hypochondriacs. I started thinking about my grandfather. He fought in the Pacific. This guy slept in a mud pile for three weeks using a helmet as a pillow.

B (6:19)

Okay?

C (6:19)

If he woke up and didn't hear artillery, that was a 100% sleep score.

B (6:23)

Okay?

C (6:23)

He didn't need an app to tell him if he was ready for the day. If he had legs, he was ready.

B (6:28)

Okay?

C (6:28)

Imagine me explaining my struggle to him. I'm sitting there like, yeah, pop, Pop, I know you stormed a beach under heavy fire, But I had two IPAs last night, and the room was slightly too humid, so my REM latency was suboptimal. I can't go to work today.

A (6:42)

This is kind of a good.

C (6:43)

I'm in the red zone. Yeah, we've evolved. I mean, it keeps going.

A (6:47)

Okay. I asked for the poster benchmark. I said, write me a post. Gemini 3, write me a post about technology that will get 1000 likes on X dot com. And it gave me a few options to its credit. Here's option three. Tech has solved a million problems. This is in bold. Tech has solved a million problems, but has it created one big one? We now have infinite connectivity, yet feel more isolated. Infinite data, yet more confused. Hyper efficiency, yet less free time. The law of unintended consequences is the most powerful force in the digital age. We need an ethics reset. What is the single greatest downside of the last 10 years of tech innovation? Arrow down, hashtag technology.

B (7:35)

No, no, it's just asking for engagement bait. It loves engagement baiting. Like no one does that anymore. No one goes on access. Let me know what you think in the comments. It's so tight. 2017.

A (7:52)

The other one, the option one is the next 12 months will decide the winner of the AI race and it won't be Google or OpenAI. It will be the company that masters hyper personalization for the average consumer. Not the most powerful model, but the one that seamlessly integrates into your daily life. Your email, your calendar, your health. The real battle isn't AG equals AI, it's AI to the power of I equals impact. Which dark horse will win?

B (8:22)

Okay, that's insane. I love how it is funny how posting seems to be unverifiable. It's very hard to create a verifiable reward environment for comedy that you can actually rl against. What do you think?

C (8:39)

There's also the other benchmark. It was like the shrimp fried rice joke.

B (8:42)

Yeah, yeah, yeah.

C (8:43)

I think it did well on that. So I'll read through some of them. So the joke is like saying you're telling me shrimp fried this rice. That's like the original one. So it's like I'm asking it to come up with more of these.

B (8:54)

Yes.

C (8:54)

So I'll read through some of them. You're telling me a chicken fried the steak.

B (8:58)

Okay.

C (8:59)

You're telling me the sun dried these tomatoes.

B (9:01)

I like that one.

C (9:03)

You're telling me a beer battered this fish.

B (9:05)

Okay.

C (9:06)

You're telling me a gingerbread this man.

B (9:09)

The gingerbread man is insane.

C (9:11)

You're telling me a beer. Wait, you're telling me a pan seared.

B (9:14)

The salmon Pan seared salmon. Yes. Yes. A pan literally sealed the salmon. That's not the joke, that's an anti joke.

C (9:21)

You're telling me a stone wash these jeans.

B (9:24)

That's pretty good. I like that. Stone wash jeans. You're telling me a stone wash these jeans.

C (9:28)

You're telling me a hand toss this pizza.

B (9:30)

I mean, yes, literally. That's exactly what it means to.

C (9:34)

You're telling me the French roasted this coffee.

B (9:36)

Yes, all of these are just true. The genius of the comedy of the shrimp frying the rice is the. The shrimp didn't literally fry the rice. The shrimp is being fried in the rice.

C (9:46)

But this is. I think this is a step change better than what we saw at GPT5.

B (9:50)

I wouldn't say step change. I would say incremental. Like it is better. For sure. For sure.

C (9:56)

But this at least is like logical where the GPT5 ones.

B (9:59)

Some of them were.

C (10:00)

You're telling me a squirrel ate this watermelon?

B (10:01)

Yeah, it was just not. It didn't even understand the concept of finding the root trace of like it needs to be like stonewashed jeans and then you rearrange it and it doesn't quite understand when that hits or when that doesn't hit. Some of those are very funny though. One of them is extremely, unintentionally funny, which I enjoy. Or maybe it's intentional. Maybe it's AGI deep down in their nose. Nose, nose. It's great. Anyway, you're telling me a restream stream this livestream one livestream 30 plus destinations. If you want to multi stream go to restream.com Sundar Pitch AI Jordi posted back in July of 2025 nominative determinism is undefeated. Sundar really did it. He, he pitched AI real. He was being mocked, he was being mocked for a long time for getting on Stage at Google IO shortly after ChatGPT launched and saying AI, AI, AI, AI. And they, they, they did a super cut of every time he said AI. He said AI a lot. And so it made it look like, oh, he's behind the ball and he's trying to catch up. And to some extent, I don't know if they were actually behind the ball, but they were certainly playing catch up in like the attention game. They just weren't getting enough attention. And so it was the press release economy. They were putting out a lot of press releases, but they are maybe done with the press releases because now they're letting the model actually speak for itself. And you can see that with the Gemini 3 Pro model card, which is doing very well. Better than GPT 5.1 on a lot of stuff, better than Claude Sonnet 4.5 on a lot of stuff. On humanities last exam, it's getting 37.5% arc AGI is up at 31% over 1317 across the board. It seems like it's a good model, sir. And so Zo Fawn says, Gemini, I'd be like whoever prayed on my downfall, pray harder. And I couldn't agree more. It's great to see Google becoming a winner and just realizing just that this was a sustaining innovation for them and that they were able to take advantage of all the infrastructure that they had across TPU, DeepMind, GCP. They were set up to Excel here. Got taken a little bit off the back foot on the consumer side, but seem to have played catch up at least on the, on the foundation model side very well.

A (12:33)

So Matt Schumer says the last time we saw a capability jump of this magnitude was the release of GPT4 in March 2023. We are entering a new era.

B (12:44)

Okay. Yeah. So points for Tyler here certainly agrees with Tyler. There's a significant jump. It is the age old question are we accelerating or decelerating? But either way we're definitely making progress. It certainly looks like acceleration in the Arc AGI2 leaderboard. You can see we are growing exponentially there. Really, really exciting chart. So Gemini 3 Pro is at 31% completion on ARC AGI 2. That is of course the puzzle solving game that is easy for humans. Even children can do it, but AI has historically struggled with it. Gemini 3 DeepThink Preview gets a 45% on it at $77 a task. And this is just way above GPT5 Pro. Grok4 thinking. When Grok4 thinking came out, it was before GPT5 and it was by far the highest on the chart. It was really, really up there and Elon was very excited about that and was showing that Grok 4 had really advanced. Well, now we're back in the horse race.

A (13:52)

Grok 4.1, 4.1.

B (13:55)

I haven't seen it benchmarked. We can ask Mike if he's heard anything but whether you're whatever you think get on public.com investing for those who take it seriously. They got multi asset investing industry leading yields. They're trusted by millions. So back to Arc AGI. Gemini 3 also has good results on Arc AGI 1. But the interesting thing here that Mike highlights is that V2 so the fastest. So he says we're also starting to see the efficiency frontier approaching humans. The fastest V2 task Gemini 3 Pro solved was this hash with only in 188 seconds. The human panel solved this one in average of 147 seconds. So you're getting like human level output but also human level speed. And then if you get to human level costs then you're really in the game. It's wild.

E (14:52)

Wild.

A (14:53)

Carpathy jumped in with some notes. He said, I played with Gemini 3 yesterday via early access. Few thoughts. First, I usually urge caution with public benchmarks because in my opinion they can be quite possible to game. It comes down to self discipline and self restraint of the team who is meanwhile strongly incentivized otherwise to not overfit test sets via elaborate gymnastics over test set adjacent data in the document embedding space. Realistically, because everyone else is doing it, the pressure to do so is High go talk to the model like we did. We went and said give us a stand up routine, give us some one liners, talk to the other models. I had Karpathy says I had a positive early impression yesterday across personality, writing, vibe, coding, humor, et cetera. Very solid daily driver potential. Clearly a Tier 1 LLM. Congrats to the team over the next few days weeks. I'm most curious and on the lookout for an ensemble over private evals which a lot of people orgs now seem to build for themselves and occasionally report on here.

B (15:55)

I wonder how fast it will roll out. I have a Gemini Pro Ultra subscription but it's on my personal email and so I need to figure out how to actually get into 3Pro on the consumer app so I can actually test it on my phone in my daily use. It's always tricky with these Google. Like Google's so big that when, I mean you're starting to see it now with OpenAI rollouts where they'll say hey, GPT5's out and we'll be rolling it out over the course of the day because the system is big enough that it actually takes time to roll out. And I think Google has even more of that. Even more of that.

A (16:43)

This is pretty cool from Patrick Collison. He says, I asked Gemini3 to make an interactive webpage summarizing 10 breakthroughs in genetics over the past 15 years and here's the result. Pretty wild. Did you click through this, John?

B (16:56)

No, no, I didn't generate shared directly from Gemini.

A (17:01)

That's cool. So this is just basically a website or an app and it's notable that. That even the UI itself is fully interactive.

B (17:11)

Yes, yes. So I had the. I did this with Claude code a little bit where I wanted to visualize like basically a deep research report and I wanted to turn it into a website and it just generated all the HTML and at the end of the day or at the end of the report, it gave me an HTML page that I could open in Chrome and use like a website. But it was local. I couldn't share it because it wasn't actually on the Internet. This is really, really cool. This is like definitely the beginning of this generative UI stuff.

C (17:43)

Yeah, I think actually, I think it was Sunder that posted it, but in search, in the AI mode, in search, it's now using Gemini 3 and there's some prompts where it'll generate UI.

B (17:56)

Yeah, it's so cool because Google's always had that UI to some extent, but it's always module based.

A (18:03)

Yeah, Also just very. I think I expect this to be like, pretty viral.

B (18:08)

Totally.

A (18:09)

And potentially a growth loop for Gemini as people just come on here, create these mini apps, these canvases.

B (18:16)

Yeah. I feel like. Doesn't OpenAI have a canvas feature? Yeah, but it's like, maybe lesser, I don't know. But can it generate HTML, custom HTML and then actually share that? I've never seen OpenAI. I mean, this would be a good benchmark. Like, I don't know what the prompt was for this. I asked Gemini3 to make an interactive webpage summarizing 10 breakthroughs in genetics over the past 15 years. Do you want to try and benchmark that just in maybe, I don't know, like Claude and in ChatGPT or in OpenAI's Canvas product? Because the idea, like the fact that this is just a URL at the end of the day, that is a powerful growth loop. That's very cool. I wonder.

E (19:04)

Yeah.

B (19:04)

I'd be surprised if Gemini really was the only one to have this feature either right now or for a long time, because it seems like a killer feature. Gemini 3 Pro is going absolutely vertical on Vending bench right now. Let's see this money balance over time across four runs. Today we're revealing two new evals, Vending Bench two and Vending Bench Arena. Soon we expect more models to manage entire businesses. This requires long term coherence. Oh, so this is where you.

A (19:36)

You vending machine.

B (19:37)

Manage the vending machine. But is this all simulated? This is.

C (19:40)

This is simulated.

B (19:41)

Yeah, simulated.

C (19:42)

There was just like a game a couple of months ago. Did like the actual. Yeah, cloud machine in the.

B (19:47)

In the office. And it was losing money and it was getting confused a little bit.

A (19:50)

Yeah.

C (19:50)

Because people would order like, just like metal. Like a piece of metal.

B (19:54)

Yeah.

C (19:55)

And then it would do it. And then you could like, haggle the price down.

B (19:57)

Yeah, yeah, yeah. It would negotiate on every price, apparently. And also it consistently thought it was like a human in the office. And so it would keep saying, like, it was that 60 Minutes documentary. It was like, oh, yeah, like, I'm down on the third floor, I'm wearing a green tuxedo. Like, come hang out with me. Yeah.

C (20:13)

It said it was wearing a red tie.

B (20:15)

Yeah, red tie. I like the idea that it just thinks like, well, what would I wear if I was in the anthropic office? Like, I'd probably wear a red tie. It's like no one wears ties in that office at all. But after the. This is the first ever vending bench game. Cloudsonic 4.5 GPT 5.1, Gemini 2.5 Pro and Gemini 3 Pro competed to win the local vending machine market. Gemini 3 Pro made more money than the other three contestants combined. And so congrats to Gemini 3 Pro.

A (20:45)

For dominating the vending machine game.

B (20:47)

The vending machine game. Before we move on to the next Gemini 3 post, let me tell you about adquick.com, out of home advertising made easy and measurable. Say good headache. Say goodbye to the headache of out of home advertising. Only Ad Quick combines technology, expertise and data to enable efficient, seamless ad buying across the globe. Anyway, Adi says I had early access to Gemini 3.0 for about two days. Thanks to official Logan K and the AI studio folks here we get to see GPT 5.1. Thinking left in Gemini 3.0. Right. Build the same Xbox controller in Minecraft and pretty remarkable results. You can start to really understand just the raw capabilities. GPT5Pro, for context, is not quite capable. I really want to know how this is actually orchestrated. Is this writing some sort of text or markdown file that then is imported into Minecraft?

A (21:47)

Yeah. Or is it more like an agent?

B (21:49)

Or is it actually driving around and.

A (21:52)

Using the internal ui?

B (21:54)

Yeah. Because Google demoed an agent product that could actually use the keyboard to navigate around. I wonder what's going on here. What's your review of this Ferrari in Minecraft?

A (22:11)

I think it looks pretty solid.

B (22:12)

It's pretty good.

A (22:13)

I mean, it's meant to be an F40.

B (22:16)

Is it?

A (22:18)

I do like, the hood is a little rough.

B (22:21)

Yeah. The front area is a little rough. This is. It's the worst it's ever going to be. It's going to be better. This is definitely like, this is the.

A (22:28)

Worst that Minecraft Ferraris are ever going to be.

B (22:31)

But I do feel like if I just search like Minecraft Ferrari, I mean.

A (22:37)

This is the vision that this sort of AGI future that Tyler has been telling us is right around the corner.

B (22:42)

Okay.

C (22:43)

These are like so much better. If you go to the MC Bench website, you can see what other models produce and this is way, way better. I think these, this is actually one of my favorite benchmarks because it's much harder to kind of benchmax this, I would think. And also it just seems like models don't really do this. If you look at a lot of GROK models which are sometimes accused of being benchmaxed, you kind of look at their Minecraft creations and it's not very good. So I think these give you a much better sense of the actual capabilities of the model.

B (23:16)

I found a Ferrari F430 in Minecraft. That looks like amazing that I want to share somehow. How do I share this? Let's see. Can I only share the X link here? I just have an image if we go to the end.

A (23:31)

Wow. I think I know what you're pulling up.

B (23:33)

Did you see it?

A (23:35)

If you just search Ferrari F430 Scuderia.

B (23:39)

Yeah. That looks amazing. Pull this image up because that'll show you how it's done compared to the Minecraft one. Wait, so do we know how this is actually generated with Gemini 3 Pro? What is the problem?

C (23:54)

I don't think it's like an agent, it's just text. It has a text representation of the.

B (24:00)

That's still really, really impressive. Like that's actually crazy. It definitely understands a lot. Yeah, but it's not this. Look at this, Tyler.

A (24:08)

That is human craft.

B (24:11)

You know what that is? It's probably like you know, a team of 50 kids for a month building in Minecraft.

A (24:19)

That's amazing. What else? Lee San, Alex Gaib of course himself says it's so over for OpenAI and Anthropic. If you want engagement on X, just start by saying it's so over blank and highlighting some more of the benchmarks. Of course it is not over for either of them. Yeah, but it's certainly competitive race.

B (24:46)

I'd be very interested. We have to get some of the semi analysis folks on the show soon. I'm very interested in understanding like okay, so we got this big jump. It's pretty significant. What's the actual structure of the capex that went into Gemini 3 Pro? Like how big is the training run? How much do they have to spend? Because like I think they're going to make the money back very quickly. Like people are going to use this model, they're going to pay for it, they're going to use it all over Google obviously, but also people are just going to pay for the API. But is this $100 million? Is this $1 billion? Like is this, is this like did they build a special data center for this? Is it all tpus? How many tpus?

C (25:33)

I think it is all tpus. I'm pretty sure I read that. But I seriously doubt they've released anything on the numbers of the scale of training done that. No one's really done that since GPT2.

B (25:44)

No, no, no, not at all. So there's got to be someone who's working backwards to actually sort of understand the dynamic.

C (25:50)

You can probably estimate the order of magnitude.

B (25:52)

Also, I've heard that Google's fantastic at cross data center training runs. So they can actually shard out or slice up the training run. So even if they don't have one massive data center, if they have five small ones, they can piece them all together and get a better result. So I don't know.

A (26:10)

Skook says anthropic to zero OpenAI becomes the Yahoo of intelligence. Google remains Google. It's extremely rude, very harsh.

B (26:19)

Sorry. The first two labs. You guys are great.

A (26:23)

Certainly too early to call it all 3.

B (26:27)

I like this take from Ben. This is funny history of AI so far crown a winner. Wait 90 days. Look silly. We're in the least predictable era of an entire industry. Google has fairly straightforward advantage. You all favor whoever released the most recent model. That is very true. Anyway, let me tell you about getbezel.com shop over 26,000 luxury watches fully authenticated in house by Bezel's team of experts. So let's move through some of the the competition. What else was going on? So everyone's releasing different things. Let's go to Anti Gravity actually and watch this video and see Google entering the IDE race. Let's play this.

F (27:20)

Every breakthrough in model intelligence for coding.

B (27:23)

Encourages us to rethink what development should look like. Gemini 3 is our latest such model advancement.

F (27:32)

So we went out to build the.

B (27:34)

Next step change of an IDE, introducing.

F (27:37)

Google AntiGravity, a new way of working for this next era of agentic intelligence.

B (27:42)

It is the ideal agentic development home base. Does it have an ide? Yes, but it also has a whole lot more. We started with the core idea and.

A (27:53)

Added pieces that evolve the IDE towards.

G (27:55)

An agent first feature such as browser.

B (27:58)

Use, asynchronous interaction patterns and an additional.

F (28:02)

Novel agent first product form factor helping you experience liftoff.

B (28:09)

Your new focus. So you like the name Antigravity? Why do you like that?

A (28:13)

I like the way it looks and I like the sort of vibe of the word. I think saying it out loud is tough.

B (28:22)

Okay. I. I thought there was a very cool feature where it feels like they're bringing together a whole. It feels like the first time for the last couple years. It feels like Google's been like stuffing AI in little corners of the ui. Like you already have Gmail and then you stuff a Gemini box there. You have sheets and then you stuff a Gemini thing over here. This feels like the first one where they were like sort of able to start from scratch. And it still has like the sidebar panel, but it Felt like it was both a code editor, but then it also kind of looked like a Google Doc in the sense that you could highlight sections and leave comments for the AI, which I thought was interesting.

A (29:02)

Yeah.

H (29:02)

I don't know.

C (29:03)

Easily guiding the agent's 90% solution all.

B (29:05)

The way to 100%. Yeah.

D (29:06)

This part now, let's say the agent.

C (29:07)

Produces a landing page mockup with nanobanana and you now want to make some UI adjustments. You can give visual comments.

B (29:14)

Yes. You can actually go in and comment in the image exactly where the problem is. And you can do that in the text as well. So you can have this more precise dialogue with the agent like you would a human employee.

A (29:26)

Yeah.

B (29:28)

And you're going to love it.

F (29:29)

Say goodbye to what held you down before.

B (29:32)

Welcome to Google Antigravity. Very cool.

A (29:38)

So it is funny, remember?

B (29:40)

Yeah.

A (29:41)

Remember when Windsurf acquisition, whatever you want to call it, was announced and it was positioned. It's like, hey, the team is well funded and has a product used and loved by thousands of engineers and companies. And I remember talking about it and we were saying, okay, the one issue is that some of the best people on your team are going to Google to compete directly with what you guys have been doing.

B (30:08)

Yeah.

A (30:08)

So fortunately, obviously, you know, the whole Cognition deal ended up coming through, but you can imagine a world where Windsurf was still independent and just trying to. And then suddenly it's like, okay, now you're just competing head to head with, with your former partners. Like, how does that make sense? Right?

B (30:25)

Yeah.

A (30:25)

So anyways, it all, all worked out for the best. But, but I'll be interested to see, I'm super interested to see what kind of adoption this gets.

B (30:35)

Yeah, we have to test it out. We'll have to get the Tyler Cosgrove review.

A (30:40)

Is it publicly available? Yes.

B (30:42)

Let's get it.

A (30:43)

Get it.

B (30:45)

Yeah, let's do a review later this week and see how it compares to other IDs. Anyway, we have our first guest of the show, Mike, new from rkgi, in the Restream waiting room. Welcome to the show, Mike. Thanks for waiting. Good morning.

A (31:01)

Morning.

B (31:01)

How are you doing?

E (31:03)

Hi. You know, a lot of these AI sort of like verification things are very much hurry up and wait. The last like 24 hours has been a hurry up mode.

C (31:12)

Okay.

E (31:12)

Always very fun and exciting to get the results out, but yeah, it always comes together very, very quickly at the end.

B (31:17)

Well, I really appreciate you taking the time to hop on, on, on such a busy day. Maybe we can just start with like your high Level reaction, like, how do you even think about these things anymore? Anymore? Are you just thinking like, okay, yes, Gemini 3, good. And then let's go a layer deeper. Are you thinking about that?

E (31:33)

Really good.

B (31:34)

What's your high level takeaway?

E (31:36)

Well, yeah. So I think the big headline is that Gemini 3 basically got like 2x soda on arc v2. And so this is the third major frontier lab now in a year to use ARC to demonstrate frontier progress, particularly with AI reasoning systems. We had OpenAI last December, XAI this summer. I'm super excited. Google's now on the leaderboard too. So that's great to hear and I should say up front, thank you to the Gemini team for giving us the opportunity to verify. Has been great. I think the really impressive thing about this and still sitting with all this stuff, it's pretty fresh. But I think the biggest impressive thing to me is about we're starting to close this complexity scaling gap between v1 and v2. Arc v1 and v2. This is the big difference between what v1 and v2 is. They look similar on paper. If you go look at the different data sets, the big change is the v2 increases the complexity of the tasks, ones that take minutes instead of seconds for humans. And so we're starting to see actual material progress on that complexity scaling. And then I think the big surprise to me personally is that Gemini 3 though, is still roughly along the Pareto frontier of V1. It's a little better, but we're still kind of roughly within the same mass shape. And there's dozens of tasks where the system still makes relatively, I think, obvious mistakes that humans don't make or recognize very quickly. And I sort of previously expected, like, if we had an AI system that was solving half of V2, that V1 would be fully solved. And like, that's not the case. So there's a lot of surprise here. I was cheating about this earlier to sort of invite sort of investigation from the community because I think there's still a lot to learn in terms of, you know, why exactly do we see such, you know, a jagged intelligence emerging right now?

B (33:14)

Let me eliminate some possible factors. It feels like there is benchmark hacking, but Google and the Gemini team feel not aligned with benchmark hacking generally. Like they've been good citizens in the community so far. And also you would assume, right, just from logical deduction, you would assume if you're able to hack v2, you would definitely go back and hack v1 as well. So is that the first time we.

E (33:43)

Verified a Gemini result either. This year we did two and a half earlier as well. So I don't think that's.

B (33:48)

So it's not like, it's not like they set up like, okay, we got, you know, the most important thing here is that Gemini 3 is really good at RKGI. V2. That wouldn't make sense. So there. So this is sort of teaching us something about the fundamental nature of this model, but we still don't know why lag, why performance might be lagging in V1. Is that right?

E (34:08)

Yeah, I mean, I've got my sort of hypotheses. You know, I think my, my personal one is that like AI reasoning systems just don't demonstrate even fluid intelligence. The ability for these reasoning systems to do adaptive reasoning, which ARC is a test of adaptation capability. It's limited to domains where the underlying foundational model has pretty good training coverage over the types of data and it has a verifiable feedback signal. And I think that's true for arc. If I zoom out even further, maybe to put this result in context of where we're at is just like an industry right now. I think over the last 10 years, I would sort of characterize. We've really had only two major breakthroughs. We've had the transformer in 2017, obviously that led to language models and we had chain of thought. It was originally introduced in 2022 and sort of went through Q into AR easoning systems and has gotten scaled up. Sure. And so like this was against the backdrop of like compute scaling. Right. And this compute scaling was certainly necessary, but it wasn't sort of sufficient. These like key conceptual unlocks were sort of the sufficient things to take advantage of that compute. And so my kind of take at this point, having looked at all this progression this year, is that AI reasoning systems with no new innovation from here can basically enable sort of mass automation. Because a lot of problems can fit into that characterization where we can generate lots of examples that look like the problem and we can get a verifiable feedback signal from them. Any problem that can be kind of cast and then characterized in that way, I think can be automated at this point, no questions asked. And then the big motivating factors, I think really for mass innovation, that's sort of what we're still not seeing. We still need new ideas for this. And I think that's closer to an AGI complete problem.

B (35:50)

Yeah, that makes sense. Is it fair to put you in contrast to some of what Dorkesh has been writing about saying that the Job of most people is not necessarily a bunch of indiscreetly verifiable tasks. Andre Karpathy has been writing this as well. There's this question of like how much of a job is actually automatable. Radiology was one, was one example where it felt like a very automatable job. And yet years into the AI deep learning revolution, like we're still seeing full unemployment there. How are you processing?

E (36:31)

Yeah, we're only a year into the reasoning paradigm. Right. Like the first major one only came out 12 months ago. And I think 2025 in my view is basically characterized on starting to figure out how to actually bring these things into production systems. Sure, this is a big breakthrough. I think this is maybe one of the mischaracterizations in my view of the progress is a lot of teams even I think if you just assume, oh, models get better, models get better. You think, oh, the last 12 months has just been sort of continued story. And if I played with the models 18 months ago, I have a rough sense of what they can and can't do. And that's just not true. Like if you're a builder building products, like this is the advice I give to teams I work with at Zapier too, still is like, look, this actually is a significant paradigm break in terms of what's possible now that wasn't possible even a year ago with these systems. And that's going to enable a lot of new types of products, a lot of new types of services, a lot of use cases that were out of scope because of verified because of reliability and sort of consistency now can be brought in scope. So I think if your intuition on what use cases are possible based on an eight year look back, you really have to start pinning your look back to more like 12 months.

B (37:39)

Yeah, yeah, that makes sense. What about does the work live within SaaS products or within individuals? Because some of those examples that you just gave are it's like for teams that are going to build products that automate work and then get vended in through effectively SaaS products to actually do the job versus like a knowledge worker who is going to be using Gemini in the app to, you know, accelerate their day to day, should they be feeling like the difference in this in the same way?

E (38:18)

You know, I mean like my one bit of advice is like if you haven't really used these AIs systems once, you should. I would hope everyone probably who's listening to the show has used these things at this point. But in case there's not like you should go, you should go use and experience these things. You know, when Google or when OpenAI released GPT5 this summer with their model router, right. That was like predicated on this data that like very few users had ever even used AI reasoning systems. And I still think it's only like one in five. Yeah, maybe it's.

B (38:45)

And that was kind of part of the Deep SEQ moment was just that for the first time there was a free app that you could go and see a chain of thought and you could actually see a reasoning model in action. And for a lot of people that was their introduction to that. And so there was like Deep SEQ wasn't necessarily that much higher, that much, you know, in front of everything else, but it just gave away a reasoning model for free at a time when they were tucked behind a bunch of other like hurdles that you had to jump through.

E (39:10)

Yeah, we're still really early on the diffusion for this stuff. That's maybe the key point seeing that on, you know, the huge numbers getting reported by Frontier Labs and their usage data. I mean, I'm seeing this in sales conversations I have for like, you know, zapier stuff all over the case. We're still very much early innings on actually getting this brand new breakthrough into like production workflows.

B (39:28)

Yep. Yeah, that makes sense. Do you have more questions on the diffusion?

A (39:31)

Yeah, one, I wanted to get your updated take on humor. We were playing, playing around with Gemini 3 this morning specifically just trying to get on our own little version of humor bench. It feels like something that like I do think about. Can you make kind of these like verify, like can you make humor verifiable? Like is there a system that someone could set up that could actually start taking, taking humor seriously? Because I could imagine like if we're hitting, if we're hitting like any, anything close to a wall, there will be a lab that says, okay, well like let's work on something that like everybody that like let's work on a new kind of angle for differentiation and maybe, maybe humor could be it is at.

E (40:17)

Least a little bit, right? Like I have a five year old who is getting into, starting to want to tell a lot of jokes and the jokes are just terrible, right? Like they're not funny at all.

A (40:26)

They're like, you end up laughing because they're so not funny. And then depending on who's delivering it.

E (40:34)

I've been trying to find the structured way to describe like, okay, here's what makes something funny. And so there is like some degree which you can kind of break down, you know, the types of things I think humans would sort of find funny. And like there is. This actually does get pretty interesting because like you're getting to the spot where you're trying to like articulate like creativity.

H (40:51)

Right.

E (40:52)

How creative can these systems be? To be creative, to be humor, to generate good art, you kind of have to intentionally break the rules. But you need to have a really good model of what the rules are in the first place to intentionally break them. And in fact, I think a lot of humor fits into this category before into this is your right. It's like it's actually breaking the prediction rather than just following the sort of prediction of what you'd expect. And today I still think when I look at the failure cases for, let's call it ar easoning systems on these tasks like Arkansas. Yeah, they still fail for what appear to be sort of random reasons. They have some version of an understanding of the rules and the strategy and the goals and then they sort of make a lot of basic mistakes, either executing them or not following their own sort of understanding that they've generated internally. So there's some sort of self consistency issues. And so I feel like if that's still the case, humor is going to be accidental rather than intentional from the systems.

B (41:40)

Yeah. What about V3? We played around with that on the show. I believe Tyler, our intern was in the top 10 for a while. Really grinded up the human leaderboard. Is it more compute intensive? Is that in the process? Are we expecting to see Gemini benchmarked to V3?

E (41:59)

I would love to. So we are in the development process for V3. I like to say we've basically built the highest, most productive game studio in the world. We're generating hundreds of these things for about, I don't know, like 2/3 of the way through building all the games at this point. Our target is to get this in a good state with sort of all of our controlled human studies, all the games verified, get frontier results checked off by early next year. And we're targeting releasing it publicly in v1 with the entire data set or sorry in Q1 with the entire data set next year. And that'll likely be alongside arc price 2026. So working on full details of how that's going to look next year.

B (42:37)

Sure.

E (42:38)

But yeah, we're sort of like in the throes of it. We're definitely using some of these frontier systems to do red teaming against the benchmark. Just to assert that, yeah, these games are still hard for AI and we're still finding that to be the case even with things like Gemini 3. But yeah, we're still in progress with.

B (42:54)

Development right now and SEMA 2. Can I have your reaction on that? Obviously it's this Gemini agent. It feels like if anyone at Google.

E (43:03)

Is listening to this and could sort of give me access to Simu 2, I would love to test it on V3. This is actually something that we haven't done yet and I would like to.

B (43:10)

Yeah, yeah, yeah, that's what I'm getting at because it feels like I don't know if there's some sort of.

E (43:15)

The claims are big. Yeah, you read the marketing material and it's like, okay, that seems like it should solve V3 before it exists. So like if that's the case we should know that. And so. But yeah, I haven't gone hands on with it yet so I can't sort of make any statement either way on the claims.

B (43:33)

Yeah, I'd be interested also to like, when I'm thinking about like V4 it's like you guys are going to have to build GTA 6 or something. We joked about that. Yeah. If I'm following the progress of like v1, v2. V3. V4 is like a game that I'm going to play for 100 hours for fun. I'm just going to pay for it.

E (43:53)

Yeah. This is one truth you've answered something true about V3 which is that it's still a relatively short time horizon tasks and they're self contained.

B (44:00)

Yeah.

E (44:01)

It does add some new complexity where you have to deal with interactivity because you have to do goal acquisition, you have to do exploration. We'll have a really nice action efficiency comparison between humans and AI which we haven't been able to get before on the v1 v2 domain. So we're going to get a lot of new signal I think on V3. But yeah, I think as you sort of look even further out into the future, things that are more open ended are the things I think we're starting to get excited about trying to understand what does it mean to put one of these AI systems in an open end environment and then look back on the system 10 minutes into the future, 100 minutes in the future, 1,000 minutes in the future. And can you look at the environment that AI system has been like how it's manipulated its environment and say something interesting about how intelligent the system is based on that observation and open ended sense? Still very early on V4, but yeah, we're starting to explore ideas there.

A (44:54)

Has Gemini 3 updated your timelines at all, specifically your Arc AGI 2 timelines in terms of when you expect sort of like the 90th, 90%, like anything on the kind of the upper end of the range.

E (45:09)

I was looking back at my the whole Ark team actually made some predictions back in January when we released V2 on what did we expect end of year scores would look like. Now, obviously, if we're only November 18, a lot happens in AI. Who knows what the next six weeks hold. But my personal prediction was that we would see about 25% on the private leaderboard for RPTO on the Kaggle contest, and we'd see about 50% on the public leaderboard. And that was sort of based on the ratios we had seen from ARC Price 2024 and the sort of scaling difficulties with V2. And it looks like we're going to come in pretty close to that. Unless barring some other major new breakthroughs towards the end of the year, that seems like we're probably where we're going to end up the year at. And then who knows on 2026. I think if we're really going to solve V2 fully, it feels like we got to better understand why these AI reasoning systems still make sort of obvious mistakes on the V1 set. And yeah, that's an anomaly. So I think that's worth serious study to come up with new ideas to sort of improve these reasoning systems.

B (46:12)

Yeah. What was the furthest timeline that you had out? I remember you said when you developed v3 you had this framework of the state of the art should be scoring minus 100% or something. You were like, you need to make it way harder than you think in order to give you room to run because the systems are developing so quickly. What's the furthest out timeline that you are tracking or you as a team are tracking?

E (46:41)

Our objective function is not longevity necessarily, it is usefulness and interestingness. I think the tasks that have the highest degree of usefulness and interestingness are ones where, oh hey, this could be useful and interesting for like three years. The Ark one was useful and interesting for arguably five years. I mean, even this year it's still interesting because we haven't broken like we're still sort of within this sort of paradigm still, and so still providing some interesting useful sleep, even though, you know, it's largely saturated up to 80% now, but there's still some interesting signal remaining. V2, our expectation was that it was not going to survive as long as V1 just because it was the same domain and we had AI reasoning systems in play at that point. I think our median estimate was like 24 months on V2. But we'll have to see how that all plays out next year. With that V3, we're hoping to be in an environment where we can actually get that to survive longer. One of the interesting things we're finding with V1 to V2 to V3 in a qualitative sense is there's a sense of how easy is it for us to generate the data set as humans trying to design the tasks and design the puzzles and design the games. And with V1, pretty much every task that Francois created was hard for AI and easy for humans. With V2, that gap got a little shorter. Actually, it got smaller. There were tasks that we generated as humans that AI solved, and there was other ones that were too hard for humans. And so we ended up pruning some of the tasks that we generated. So the gap between those things got short. With V3, we're finding it's getting wider again, where pretty much every game we're coming up with is fitting into this paradigm of very obvious and intuitive and easy for humans and sort of very hard for frontier AI still. And I think that's like, credit to Francois here. This is something he shared about a year ago with 03, but he's like, this is actually one interesting way you could characterize how close are we to AGI is when humans run out of the ability to generate interesting things that AI can't solve. Hard to argue. Any expert's going to say, yeah, we don't have AGI.

B (48:49)

Yeah. Because you can sort of think about the project of humanity as like, go do the hard and novel things. So it's like, is acquiring diamonds difficult? Okay. That has value and then we base a whole economic system around it. And it's like, somewhat arbitrary, but it's also like a skill and might and will issue. And if you can put that on display, then you accrue economic value. And so that kind of traces out into everything that we do in life and beyond.

A (49:15)

Last time you were on it, if I remember correctly, you made a call for new ideas, needing new ideas. What's the update on that front? Are you seeing anything promising outside of LLM World?

E (49:32)

Yeah, there's some pretty interesting stuff coming out for Archrise 2025. We're in the throes of reviewing all the papers, judging all the scores. The Official results for Arcprice 2025 come out on December 5th. I believe so. I have to can't share. I don't want to spoil the final announcement. I think one of the big things that we saw From ARC Enterprise 2024 was this concept of test time adaptation. This was the idea that look a pre trained model applied through a single forward pass at inference time will never solve arc. You need some ability to take information from your test and incorporate it back into the system. And that's where your adaptation capability comes from. And that was done through test time fine tuning during the contest. AI reasoning systems are a version of this where you're incorporating test or private data set.

F (50:16)

Wow.

E (50:17)

Yeah, yeah. Literally like you take a pre trained model and then like take the secret, the private puzzle, augment it in a bunch of different ways to generate permutations of it and then do like a LORA or some sort of test time fine tune on your pre trained and that that actually works.

B (50:29)

Wow.

E (50:29)

The sort of the, the, the, the common ground between this and a reasoning systems is that both of them take information from the private test and are able to operate over it with it at test time.

D (50:39)

Right.

E (50:39)

This test time compute is another form of what we're talking about here. So that was 2024. One of the big things we're seeing in Operation 2025 is this concept of refinement loops anywhere where particularly with language models being put into outer loops where they can move from state to state and how they move from state to state is they need to make some sort of refinement on the program or the natural language explanation of the task that they're working towards and they just iterate on this refinement loop over and over and this is significantly increasing scores even over the sort of test time fine tuning stuff that we saw from, from last year. So Jeremy Berman and Eric Pang were two folks who were on the public leaderboard last month that explained how their approach worked in this way. So we're seeing a lot of approaches like that. I still think we're in a regime though where like we still need new ideas. None of these are sort of sufficient to solve ARC including inclusive of V1. And so like you know, this gets me excited because I still think that means individual people, individual teams with small budgets, small compute budgets can still play a really, really massive role in advancing AI.

A (51:43)

Very cool.

B (51:43)

Are there other areas where we are making progress in AI that might sort of need to come together to actually maybe solve this or maybe just be a more complete system? What I'm thinking of is like very few solvers that I'm aware of will actually just take a screenshot of the puzzle and inspect it with some sort of diffusion model. That's not the way these, these AI models reason about ARC puzzles. Sure. We're also seeing a bunch of work on world, world models and simulators. World simulators which seem really interesting. And I was talking to one guy who is building one and he was saying like I think that we're going to get like really, really robust knowledge out of these at some point once they scale up fully. And I'm wondering if you are optimistic about bringing in other like unifying some of the different research that's happening.

E (52:42)

I think it's all of those examples of new research, new companies, new startups. Like you know, there was a seismic shift in 2025 from pre training budget to these like RL reinforcement learning environment startups and companies that are generating environments to produce more ground truth training data in a mass way because they're, you know, automated environments and you can get verifiable feedback signals out of these things. Again, there's no new science here. Like this is a good bet for like all frontier labs to make. This is going to drive progress for the next 24 to 36 months. You're going to continue to see amazing frontier headlines Just, just on, just on this fact. There's really no new sort of, I think discovery that's, that's quite needed there. You know, I think that if you're kind of pushing more towards the AGI side like what's, what's sort of missing like one question I have that is an open question is so we've got like you would think that based on like 100x to 300x increase in efficiency, we've seen from AI reasoning systems over the last 12 months that we would trade that increase in efficiency for inference tokens to do more like search coverage over the problem space. When we're giving these systems tasks or problems that we want them to solve and this is one of the big reasons why I sort of expected if we can solve half of V2, you'd get 100% on V1. And it seems like these AI reasoning systems are not fully exploring all of the search space that they could in order to look for solutions. I have an open question of well, how much of the search space can they cover and what do you need to change about the training methodology or process to actually guarantee that you can get full coverage over the search space of possible programs or possible solutions? And so that's kind of like one interesting thing that I'm paying a lot of attention to right now.

B (54:29)

Yeah. Yeah. Even just the metaphor of like the test time fine tuning, it feels like working on a problem and then like going and taking a walk and kind of like updating your whole worldview. It feels like something that humans get closer to doing that than any of the other paradigms. So, yeah, it's fascinating to see all these different approaches, all the crazy results.

E (54:54)

You'Ve heard about in the last 12 months are kind of this merger of deep learning and symbolic program synthesis style methods. The ICPC, the IMO Gold, the Gemini 3 stuff. Today, these are all systems that are still fundamentally using a language model, but they're adding symbolic knowledge recomposition systems on top of these things. They all work slightly differently, but it's like this is what's working right now. And so I think the rough search space of research and how you merge those two paradigms together is still relatively underexplored. There's a lot of different ways you can put these two paradigms together. And for new teams that are considering working on new ideas, I would explore. Well, what are the novel ways you could consider merging these two spaces?

B (55:32)

Yeah, that makes sense. Jordan, anything else?

A (55:35)

This was great.

B (55:36)

This is amazing. Thank you so much for jumping on on short notice.

E (55:38)

As always, guys, thanks for having me.

B (55:40)

Back on the continued. The continued just stacking up the wins on arcagi becoming and just continuing to mog the models.

A (55:49)

Mog the world.

B (55:50)

Yes.

E (55:51)

I mean, again, our goal is to be very useful and interesting, so we're going to try to hold that bar.

A (55:55)

Yeah, you're keeping them honest. My wor.

B (55:57)

I think you're keeping them honest. I think you're keeping everyone honest and you're providing like a very, very useful reality check on an industry that loves.

A (56:05)

To inspiring the labs to grind harder.

B (56:08)

And now there is a moment where we can feel very confident about taking victory laps and cheering for all the hard work that went into Gemini 3, because it does seem like it was a great model.

E (56:18)

It's performed well. There's definitely a big improvement today.

B (56:20)

Fantastic. Well, thank you so much. Have a great rest of your day.

A (56:24)

Great catching.

B (56:24)

We'll talk to you soon.

A (56:25)

December 5th. We'll see you then.

B (56:27)

We'll see you then.

A (56:28)

I wanted to talk about Adio because.

B (56:32)

Adio is an AR native CRM that builds scales and grows your company to the next level. Also wanted to talk about wander.com, book a wander with inspiring views how to upgraded many dreamy beds top tier cleaning and 247 concierge service.

A (56:42)

Let's sing it. Find your happy place. Find your happy place.

B (56:46)

Book a wander with inspiring views. I already know the song. You know the song?

A (56:50)

I wanted to pull up this post from Chris Pisarski. He did a GitHub style image of our streaming activity for the year.

B (57:00)

Oh, really?

A (57:01)

Did you see this?

B (57:01)

Oh, yes, I did send.

A (57:02)

It's at the very.

B (57:03)

To him.

A (57:04)

Should be at the very bottom. Yes, yes, it's at the very bottom.

B (57:07)

It was very cool.

A (57:08)

Our timeline. I have it. And if we could just pull up this image.

B (57:11)

So the Internet rewarded TVPN for showing up on January 28th. That's when we went live. We never remember the day that we went live. But. But he has it if you look it up. January 28, John Coogan and Jordy Hayes launched a daily live show and set one simple rule. Show up five days a week. Looking back, they did exactly that. 125,000 followers on X. 41,000 subscribers on YouTube, 17 and a half thousand on Instagram. They showed up every day. The Internet rewarded the proof of work.

A (57:38)

So the only thing is these. I don't know. Am I just colorblind? But is it like a little bit like I'm seeing three days that were federal holidays that we missed and then three days that were no streams. I thought we only missed.

B (57:50)

Exactly. Tell. Yeah. What is a federal holiday? What is a no stream? It looks.

A (57:56)

I think there's maybe six days a.

B (57:57)

Gray and a purple. There were a couple days here and there. We took one off. I went to a wedding in Mexico. We took a Friday off for that. That was just no live stream. July 4th we took off. That was a Friday. That was a federal holiday. And then what happened in March? We took a Wednesday off, no live stream. On Wednesday in the middle of March.

A (58:19)

There was one day that we were traveling.

B (58:21)

Oh yeah, that was after Hill and Valley. After DC I saw it was Thursday though. No, no, we didn't. We did Tuesday in the hotel room and then Wednesday we did at the actual event, Hill and Valley. And then we flew back and got back on the horse. So we missed a couple Mondays because of federal holidays and then we missed a Tuesday in. In May. That might have been Hill and Valley. March might have been something else.

A (58:49)

Anyway, anyways, very cool.

B (58:50)

It's been a wild ride.

A (58:52)

Thank you to everyone pulling it together. Chris.

B (58:54)

Along the way, our next guest is I believe already here we have Jonathan Neiman from sweetgreen. We're going to be marks to bench presses the most important benchmark in the world. How many grams of protein are in your protein bowl? We need to know. Welcome to the Stream. Please introduce yourself for those who might not be familiar.

G (59:17)

Hello, my name is Jonathan Neiman. I'm the co founder and CEO of Sweetgreen.

B (59:21)

Get that overnight success button ready. When did you start this company?

G (59:26)

2007. We've been at this for 18 years.

B (59:30)

18 years, wow. Let's talk about the very beginning. I mean, since this is your first time on the show, where'd you grow up? How'd you get into the business? What were you studying? And then. Let's go.

H (59:44)

Yeah.

A (59:44)

You gotta be somewhat of a masochist to get into the restaurant business.

B (59:47)

Yeah.

G (59:48)

Yes, absolutely.

A (59:49)

I mean, it's such a beautiful thing because it sounds so simple. It's like you get a box, you get a menu, you get some ingredients.

G (59:54)

It sounds. It sounds super cool.

A (59:56)

And then you just copy and paste it and you scale to, you know, however many stores. And then of course, it's far harder in.

B (60:02)

So prior to launching the business, what were you doing? Doing?

G (60:04)

So I grew up here in Los Angeles. I went to school in D.C. went to Georgetown.

B (60:07)

Okay.

G (60:08)

And never thought I'd be in restaurants.

B (60:09)

But always studying government.

G (60:11)

No, I was studying business. I always. I knew I wanted to be an entrepreneur. And Sweetgreen was almost. Almost an accident. You know, we. It was the naivete. We thought it would be easy.

E (60:20)

Yeah, we.

B (60:21)

And did you start it during school?

G (60:23)

Yeah, we started. We started while we were seniors in college. I started with two of my friends before that. Yeah, I had a bunch of internships. You know, I was, you know, I worked in media, I worked in tech, I worked in real estate. Always knew I wanted to be an entrepreneur and create something. But senior year came around and. It's exactly what you said. We thought it would be easy. We're like, how hard could this be? We'll go to farms, buy the food.

A (60:44)

It's like apparel. People fall into the apparel trap because they're like, I just wanted to make clothes that I wanted to wear. And you realize it's the hardest business on apparel and restaurants. Probably the things that seem the most simple but are actually the hardest artist practice to actually do on a massive scale.

B (61:00)

Yeah. So what was the. Was it. Was it build a business plan first, assemble a team, do a pop up? Like, what was the first thing where.

A (61:08)

You were like, okay, what was the first bowl?

G (61:09)

The first bowl was the guacamole greens we made in our dorm room. We brought a bunch of classmates to try it. My partner Nick actually made it. He was our first chef.

B (61:17)

No way.

G (61:18)

And the story was really simple. We had no. We couldn't find a healthy place to eat. We saw Chipotle taking off and we're like, wow, there's someone is going to create a scaled healthy fast food chain. And at first it was, let's just open one. We wanted it for ourselves. We thought we'd go on with our lives. We opened, we worked on it senior year, wrote a business plan, raised $300,000.

A (61:41)

There we go.

G (61:42)

From 50 investors. So it was like five grand average.

C (61:47)

Yeah.

G (61:48)

Party round.

B (61:49)

So they got equity in like, like what became the full company.

G (61:51)

They got equity. Well, we actually, it was a little bit more complicated than that at first. The first three restaurants, we raised at the restaurant.

B (61:59)

At the restaurant level. Yeah. I was wondering if you were doing that.

G (62:01)

And we actually paid the investors back every quarter and did the whole thing. And then after the. After the third restaurant, we realized that the only way to scale this was to roll it up. So we rolled the whole thing up. And then we were able to continue to invest in it. And we.

A (62:13)

It's notable. When did the word wellness actually become mainstream? Or when did that become, like 2015.

G (62:21)

Yeah. Like early 2000s.

A (62:23)

Yeah. So anyway, this is like. Anyways at least five years before wellness is going like mainstream.

G (62:29)

Yeah. When we were starting, the thesis was healthy eating was not cool and it was not delicious and it was not accessible. And we're going to create a place that offers all of the benefits of fast food in terms of the convenience and the taste. But do it in a, you know, do it with healthy food and real, real food that you can trust, where we're transparent about where the food comes from, where it's nutritious, and build a brand around it. And so we've been at it for about 18 years. We have almost 300 stores all around the country.

E (62:59)

Yeah.

G (62:59)

It's almost hard to believe.

A (63:00)

Yeah. Yeah.

B (63:01)

What was the first VC round?

G (63:03)

The first. So we.

B (63:05)

Or this transition from the. You have a restaurant and what did it work immediately? You set up one restaurant. You, you know, you raise enough money to get that. I imagine that you decide a lease. So you weren't buying buildings, but you might have to do some sort of renovation to actually get the first restaurant up and running. You start making money enough to pay the employees, enough to pay the rent. You scale it to three, and then at a certain point you say, okay, we're going, we're going to Turn this into like a corporation more than just a small mom and pop. Right?

H (63:35)

Yes.

G (63:35)

So we, we opened one in 2007, two in 2009 with a food truck. You remember those?

E (63:41)

Yeah.

G (63:42)

And then we opened like two or three a year. And we were mostly built them from cash flow, from the profit. We were profitable. You know, we would just reinvent, invest the cash flow and we would do a few party rounds.

A (63:53)

Yeah.

G (63:54)

2013, along the way, we started a big music festival called Suite Life. 2010 became a massive 25,000 person music festival.

B (64:02)

Festival.

A (64:02)

Where was that?

G (64:03)

It was at Merriweather Post Pavilion. So first year we had the Strokes. By the end we had Kendrick Lamar.

A (64:08)

And little festival side Quest.

G (64:10)

Yeah, it was a, you know, way to build the brand. And then we focused on dc, which was very, you know, it was almost an accident, but we opened the first 16 restaurants in DC.

A (64:19)

Wow.

G (64:20)

And then slowly went up to Philly and then Restaurant 20 and 21 were Boston and New York. Yeah, And Boston and New York really kind of proved the concept outside of D.C. and took off. And that's when we raised our first.

A (64:33)

So now obviously all around la there's sweet greens. But what. Given that you grew up here, why didn't you. Why, why not start here? Was this because, well, like there was. Was there just more healthy food options in LA and there was less on the.

G (64:48)

Honestly, it was an accident. We were in school and we're like, let's just open one. We thought the second one would open.

A (64:52)

In LA and there's gravity that you have around the center. When there's more restores, you know, when.

G (64:56)

You have a restaurant company, the brand and all your economies of scale happen at the local level.

E (65:00)

Yeah.

G (65:01)

So for us, especially given our supply chain is regional, you have your overhead and your management, like your team that runs it and then your brand. You know, restaurants brands don't really travel across the country. Occasionally they do. So it was really started in dc. We thought the second restaurant would be in la. We went and looked.

B (65:18)

This is true for even like In N Out is not a national brand. It's like so west coast brand. And yeah, it's taken so long for that to actually like filter across. How capital intensive was it to launch like the second and third, like you mentioned, 300.

G (65:33)

That's the hardest part of. No, it's way more than that now. The first one was tiny. 500 square feet and we did it really on the cheap.

B (65:41)

500 square feet?

G (65:42)

500 square feet.

A (65:44)

I imagine like one or two people.

B (65:45)

Yeah.

E (65:46)

Wow.

B (65:46)

That's tiny.

G (65:47)

Yeah, we were working there, we were doing the whole thing. So we've had to raise a lot of money. Answer your earlier question. Revolution. Steve Case was our first, first VC investor and it was part of the thesis was how technology can change the restaurant. So we were, we were the first company to you do mobile ordering where you can order on your app and pick up.

A (66:06)

And we started most beautiful software for that a restaurant had ever had. Probably. Emmett. Emmett. Emmett Shine.

G (66:14)

Emmett Shine. Yeah. Shout out. Gin Lane. Yeah, this was like.

A (66:18)

Yeah, this was like one of my favorite Gin Lane projects.

B (66:21)

That's awesome.

G (66:22)

Emmett and his team were amazing. They did our app in the early days and restaurants today cost over a million dollars. So we're like $1.3 million, 1.3, $1.4 million per restaurant. That's before you put the Infinite Kitchen in. Our restaurants have very high return on capital.

B (66:40)

Infinite Kitchen, what's that?

G (66:41)

The Infinite Kitchen is our automation, our automation platform that we've built. So today most restaurants that we open, the assembly is automated. So we still make all the food from scratch. The sourcing is the same, we still cook the food fresh, but we load this beautiful machine that makes your bowls, it makes them 500 bowls per hour, perfectly portioned, perfectly plated. And so that is kind of the future of where things are going.

A (67:06)

How many different restaurant automation pitches did you get across? 18 years, like as I imagine, and every single year there's a new startup coming to you saying, we can automate this part of your kitchen. And clearly you got to the point where you had to build it yourself based on domain knowledge. But this just feels like something that's been promised for a long time. And at this point, I don't know. An individual startup that's done well in restaurant robotics.

G (67:37)

Yeah. No one's been able to create a platform platform that, that works in multiple restaurants. And there's a few, there's a few issues. Most restaurant workflows are very specific, so they're super specific to that restaurant. Two, most restaurants are franchises and so they're not owned by the corporation. We are fully company owned. So if you're a franchise restaurant, you know, if you're McDonald's, you have to now go convince your franchisees to buy whatever automation you have.

A (68:03)

And the other issue, looking at it, and it's like, this is coming up my bottom line, we're making money already. This feels like a risk. Like the franchisee is saying, like what? Like, I'm happy with my ebitda. I don't need to take a risk.

G (68:17)

That's exactly right. And the other issue is you need automation that takes enough labor out or offers enough value to be worth it. Because the capex is still very heavy.

A (68:27)

Yeah.

G (68:27)

So we went down this path. We tried to build it ourselves, actually. We built a team to do it ourselves, realized how challenging it was, and then we found this startup that was doing it and doing a really good job. It was called Spice. It was called Spice Kitchen. It was for MIT grads out of four grads out of mit and they had the same issue. They realized they could build the automation, but no one was going to buy it. So they ended up opening two restaurants. They were great at automation, not so great at the restaurant side. And then four years ago, we acquired them and we began. We've commercialized the technology. We've scaled the technology today, so most new restaurants feature the technology. And last week we actually just announced that we've now sold Spice. So we sold out. So we spun Spice out. Yeah, we spun Spice out. We announced about 10 days ago we sold it to wonder Mark Lore over there.

B (69:14)

Yeah.

G (69:15)

So we sold it for about $186 million.

A (69:18)

Mark is Mark. Mark. I don't. I don't fully understand that. That business, but talk about a guy that just, like, isn't even necessarily naive about the challenges of restaurants, which is like, I'm gonna go into the most competitive environment possible, compete with everyone.

G (69:33)

It's amazing. It's a great. It's a great vision. And, you know, I'm a big fan of his and what they're doing. And so we. It's a. It's a really interesting deal. So we, we sold the. Effectively the team and the ip, but have full access to it. So we will continue to scale with it and get the benefits as they get. They get to scale and build many more machines. We'll get the benefits of those economies of scale as well.

B (69:56)

Can you go a little bit deeper on the decision to franchise or not franchise? The naive, maybe Steel man for franchising the franchise model is that it's somehow more capitalist in my mind, because it decentralizes the decision making and it puts these financial incentives at the local level. Because each store lives and dies by its own P and L maybe, versus even if I have a manager in one store and they have stock options, like what they do on the weekend, if they come in on Thanksgiving or Christmas, like, that doesn't necessarily put more or less money in their pocket. Is that real what I'm feeling? Or is it irrelevant.

G (70:39)

What you're feeling is absolutely real. And we actually try to design our comp structures. And I've always believed my line that I sit tell my team every single day is all the answers are in the restaurant. And the closer we can push decision making to the edges of the customer, the better we will be. So our general manager, we call them the head coach. They are the most important position in the company by far. A great head coach will make or break you. And so we try to really incentivize them, we empower them, and we try to run as decentralized of an operation as we can. The reason we decided not to franchise is it's really hard to maintain quality if when you give up that. When you really give that up to other people to run, you could sometimes scale too quickly. And we do a few things differently. We source differently. We're a very complex model because of the sourcing and scratch cooking. The biggest difference between us and most of other companies is if you go into sweetgreen, you'd be shocked at how much we are making in the store. Every single thing.

A (71:39)

It feels like you guys have taken such a principled approach in making food that I feel like stays true to the initial values of the company and kind of why you started it. And yet you're competing in an environment that says, okay, we're going to have these, like, factory kitchens off site that we're going to be shipping in, effectively almost finished product that gets gets reheated, and we're going to be sourcing from all over with not a lot of values around how they're sourcing. They're just trying to get. They want the food to taste good when it hits the plate, but maybe they don't care about a number of other factors. And so you're kind of in an environment where because of your principles, you're like fighting with your hands tied behind your back against competitors. And I'm not talking about direct competitors, but more. So you're still competing with burger King and McDonald's, right? Like, people are going to have lunch somewhere and they're going to maybe decide between they have options. Right.

B (72:32)

Talk to us about land. Is McDonald's a land acquisition company? Why do people say that? Is that real? Have you ever looked?

G (72:39)

Yeah, well, they do own a lot of. A lot of the real estate and lease it back to the franchise. So that is true. And if you watch the founder, the last line in that movie where he's like, it's real estate. It speaks to more than the fact that they just own it. Restaurants is highly a real estate game.

H (72:55)

Okay.

G (72:55)

Great real estate is like, if you look at like our portfolio where we have great real estate, we do amazingly well. Location, location, location, location, location. Really, it's people. Like people think restaurant business is a food business. It's really a real estate and a people business. And it's all about, like, you look at the great restaurants, so the chick Fil, A's, the raising canes, the in and out. It's so much about. It's about that culture.

A (73:19)

How scientific. How scientific is you. You hear stories of, of companies like Starbucks and you can imagine like a team of data scientists with like, you know, 50 monitors and they're just, we.

B (73:31)

Need one Starbucks directly across the street from the other Starbucks. Yeah.

A (73:34)

You know, so like, you can imagine a world where it's like hyper, like hyper data driven, like down to a science. And you just know when you're opening a new store, you know that it's going to hit. But there has to be like some vibes.

B (73:48)

Complete vibes.

G (73:50)

Yeah. That is the process. We call it art and science. And pretty much everything we do, it's an art and science approach. And real estate's exactly that. You know, the science. We have a very, very intricate model that looks at psychographics, demographics, mobile data drive, you know, people driving by. We have custom data around how many gyms nearby and right side of, you know, sunny side of the street, or not sunny side of the street, all of that stuff. But then you need a human to also walk it, feel it and understand. Does it tell our brand story for us, we, especially when we were originally, like early days, growing where we went said a lot about who we were. So for example, we went to New York. We didn't go to Midtown. We went to Nolita. We went to Williamsburg. We wanted to kind of tell the story about who Sweetgreen was. Today we're kind of everywhere. But the real estate is an art and science and tells a lot about it. Says a lot about who you are.

B (74:41)

Yeah. How do you think about if a new entrepreneur came to you and was asking for advice on where to start? Is it worth it to go straight to Manhattan or straight to Beverly Hills and try and make it in the big leagues on day one?

A (75:00)

Or can you get negative indicators from that because there's a different type of customer there that's not necessarily representative of. Of the rest of.

G (75:08)

I think that's more. Right. Especially when you're talking about New York. So when you, when you're talking about New York.

B (75:12)

It is.

G (75:13)

I mean, the beauty of it is a massive market.

B (75:15)

Sure.

G (75:15)

You know, it's, it's, for us, about a quarter of our business. It happens in New York. We have like, you know, in the New York region, I think we have 50 something restaurants.

A (75:22)

Wow.

G (75:23)

What? So it's great that it's massive. There's density, there's, you know, they have money, et cetera, but it's not really indicative of the rest of the country.

A (75:31)

Yeah.

G (75:32)

So if you want to, you know, a, if you want a scalable model that you can have thousands of locations, you're better off, you know, going into a more. You want to go to like the Iowa, you know, like. And to use the political analogy, you want to go to like something that, you know, is more representative of what the rest of the country looks like. In restaurants, the place where everyone goes the fast, like the fast casuals is Columbus, Ohio.

B (75:56)

That's where people go.

G (75:57)

They say, you know, Columbus, Ohio, you.

A (75:59)

Can make it anywhere.

G (76:00)

Yeah, you can kind of make it everywhere.

B (76:02)

Yeah, yeah, yeah. So, I mean, if you were a small restaurant, you were being evaluated by, you know, the CEO, McDonald's or something, he might say, how are you doing there?

E (76:10)

Yeah.

G (76:10)

The things that they look at for a restaurant is they look at your unit economics, which is effectively your payback. So how much does it cost to build and how quickly do you pay those stores back? And they look at your tam. So they say, okay, like, can you have 100 of these, a thousand of these, 5,000 of these? And those are the two big kind of, of things you would look for in evaluating the growth trajectory of a restaurant.

B (76:31)

What's the story of the delivery market? It feels like doordash has become a massive business. Uber Eats has become a massive business. More people are ordering delivery. There's the ghost kitchens trend. Is there a ghost kitchenification where these businesses are trying to effectively turn you into ghost kitchens? Does that give them some sort of leverage? Is there some sort of tension there or is it pretty much just like, oh, it's just this trend. People are cooking less and less and so they're going to go to Sweetgreen, but they're also going to order sweetgreen delivered more.

G (77:06)

There's definitely a little tension. We're partners. A lot of our business comes through those marketplaces. But it's not so dissimilar than a hotel chain and Expedia.

C (77:17)

Sure, right. It's.

G (77:18)

You're paying a fee on it. You do not control that data. You Cannot market directly to those customers. And so for us, we have to charge a higher premium. So when you order on DoorDash, by the way, it's more expensive than you order on our app. So just a quick shout out, download the sweet green app, things are about 20% cheaper there.

B (77:34)

Got it.

G (77:35)

But at the same time, it's a great way to find new customers.

B (77:37)

Sure.

G (77:38)

So, you know, for example, DoorDash has been a great partner. They power our native to what we call our native delivery delivery on our sweetgreen app, which is a big part of our business.

B (77:47)

So you're white labeling or something, correct?

G (77:48)

Yeah, it's like a white label on our app. And then we also, you know, we.

B (77:52)

Partner with the different front store as their front end.

G (77:54)

And as you know, they've become, you know, they're brilliant business models. They've become largely marketplaces. So, you know, you kind of have to buy your way to the top of the feed.

A (78:04)

Yeah, yeah, yeah, yeah. And so that's how they gain. I mean, if like, there's a reason the DoorDash app or any of these mobile ordering experience or not, they don't just put like the restaurant that you've ordered the most from at the top. It's like, hey, why don't you try this new restaurant or this new restaurant?

G (78:19)

And those are all paid.

A (78:21)

Yeah, they're all paid. And it's how you maintain leverage over. I mean, they do this on. This is why the YouTube subscriber count doesn't mean anything because it's like they're gonna constantly surface anything.

G (78:33)

They've made money. I mean, the way they make money. These businesses have been historically very challenging. The way they made it work is batching orders and then becoming an ad marketplace. And that's what's made this amazing service an amazing business.

B (78:47)

Explain batching orders really quickly.

G (78:49)

So when you order, they have a delivery driver pick up multiple orders.

E (78:52)

Got it.

G (78:53)

So you're paying the delivery driver once, but they're picking up from three restaurants.

A (78:57)

Yeah. I feel like you guys have done a really good job of listening to customers. I would say, like this 100, 100 gram protein that you guys are launching.

B (79:08)

I was asking for 200.

A (79:11)

No, but that. And then also this. The seed oil is something about the business that it feels like you're more agile. Yeah. Is the business set up in a way that you guys can respond when.

B (79:23)

Other companies, like, just caught a lucky break?

A (79:25)

Because people, I would say, like, people would give a lot of the same feedback to Chipotle and it Feels like Chipotle is not set up in some way to, like, be like, oh, this is what customers want, or even, like, some percentage of our customers really care about this. Let's deliver them. Let's deliver them a product here. And I think the results is that, you know, I've churned from Chipotle almost entirely because of the seed oils. Yeah, because of the seed oils and just, like, a degradation of the quality of the food over, like, a decade. Like, I watched it basically get worse and worse and worse and worse over 10 years. And so I just don't go there anymore. I joke about it. I'd almost rather. When I'm on the road, if I'm on a road trip, I'm almost always rather just fast than eat at, like, the most common kind of, like, fast food.

G (80:09)

Yeah. When we started the business, I had the saying, this thing I would always say is, you know, there's businesses that, as they get bigger, get better, and you can think of, you know, technology businesses, many of them do. Like, your new iPhone is, for the most part, much better than the original iPhone. These AI models are much better than the original AI models. Restaurants typically go the other way.

A (80:29)

Right.

G (80:29)

Is scale kind of degrades quality. And that's because doing, you know, serving food at scale is really, really hard to do. So you have to fight that inertia so hard. Because all of those micro.

A (80:43)

One restaurateur has an amazing restaurant. They're like, cool, now I'm gonna start a second restaurant. And the second they start focusing their energy on the second restaurant, the first restaurant gets worse.

G (80:51)

Yes.

A (80:51)

Like, it even happens at, like, a microscope.

G (80:53)

It's people and culture. And so you need to really have a lot of systems in place, both, like, culturally, how you. How you keep the team engaged on your mission, but also other systems to make sure you're watching the quality of the food and listening to your customer. So, like, seed oil is an interesting one. When we first we. We got rid of seed oils about exactly two years ago. And at the time, it was not the national conversation. It was pre RFK and all that stuff. And so when we. When we. This is one of those examples.

A (81:22)

But it was. It was. It was not a national conversation, but it was incredibly online conversation, but a tiny time. There was like an at. There's the seed oil scout.

G (81:34)

Yeah. So it was a tiny conversation. We surveyed our customers. And this is why, like, surveys are bullshit.

A (81:39)

Yeah.

G (81:40)

Really, surveys can give you a general indication, but if you just follow surveys and the market research, you're going to hit the middle of the bell curve in everything you do. And we're not trying to be a middle of the bell curve company. You got to find that like, what are your top 5 or 10% of customers doing? And we heard from. It was honestly friends, like wellness people in LA and New York that are like, hey, I don't, you know, I can't go to sweetgreen anymore because I care about seed oils. And I remember we brought it to broader, you know, I remember my CFOs like, what are you talking about? Like, what even is this? And we're like, no, trust me. It was one of those like gut decisions and it was expensive and we had to change a lot in order to do it.

A (82:17)

But here's the thing. It's healthier and it tastes better.

G (82:21)

Exactly.

A (82:22)

Most health trends, they might be healthier, but you're. It doesn't. It's not as good. Right. So I would argue like going from like dairy based, you know, traditional milk to like nut based milk almost always is like somewhat of a downgrade. Or going from like something with sugar to pulling sugar out, it's like not as good. Or going from like sour, like bread with gluten to gluten free bread, it's not as good. And so when you think about these, like, what is like a durable health trend, it's like something that's better for you and tastes better. And so that's why I was always super bullish on that trend. And I expected a number of restaurants to say like, hey, this costs slightly more, but the product's going to be better and it's going to be healthier for you. And that's what can create like real momentum around a trend versus, versus some of these flash in a pan health trends, which is like Paleo or which is only eating stuff that was super old.

B (83:15)

Right.

G (83:17)

What's unfortunate about seed oils is it's become politicized a bit.

A (83:20)

I know.

G (83:21)

It's like I did an interview with the New York Times and they're like, did you do this because of rfk? I'm like, no, I did this two years ago. This had nothing to do with rfk. This is not a political statement. We don't make.

A (83:33)

We're making foods.

G (83:33)

I make food.

A (83:34)

How your grandma probably made it.

I (83:36)

Yeah.

G (83:36)

This is about olive oil. Like, this is not about. This is just about olive oil. That's it. This is not a political statement.

A (83:42)

Taste the difference. Yeah, yeah. Is there anything happening upstream in terms of like automation or technology on the farming side, that's like exciting.

G (83:56)

Yeah, there's a lot of stuff happening on automation on the farm side. It's actually very exciting. The both the better robotic arms and the vision, I mean it's making some really hard, grueling tasks around picking happen much, much faster and easier. So relatively early still. But I think in the next five years you're going to see that take off. I do think you're going to see a lot more restaurant automation as well.

A (84:18)

Yeah.

G (84:20)

You know, between the availability of the labor, the cost of the labor, it's really just when you think about it, it's just a hedge on labor. And here like in West Hollywood, minimum wage is 22 bucks. So we pay like 24, 25 dollars an hour here in parts of LA. So with wages going up, availability going down, and then the ability like all technologies to just do things better, not just about the cost savings. Like for us with the infinite kitchen, we can serve twice as many people per hour as we otherwise could.

A (84:53)

Wow.

B (84:53)

What about drone delivery? We've seen some four wheeled guys.

A (84:58)

Imagine getting 100 grams of protein out of the.

B (85:01)

There's the air delivery.

G (85:03)

Yeah, I saw you guys talking about Zipline. I love Keller and I'm a big fan of Zipline. We're one of the early partners that are going to be piloting that. I think his way of delivering to the suburbs is super interesting. We haven't done the, the street delivery yet. Starship, the Starship Cocoa.

B (85:23)

I've met with them, working on one.

G (85:24)

I think it's, I think it's interesting. It's, it's, it's. In the past year they've really taken off. You're seeing them more and more. They still kind of weird me out a little bit. Seeing them walk, go down the street.

B (85:33)

I saw it kind of stuck in the side of the street once. It was very sad.

G (85:36)

My kids love, see, I love to love it when we see them on the street a lot.

B (85:39)

And I, and I do, I do just imagine that the AI is going to get way better. And also some of the teleoperation, just infrastructure to actually make sure that there's the ability for a human to jump into that little robot that's driving around. At a certain point you just need a lot of people set up with that, all the software working, make sure it's connected to the cell phone towers effectively or Starlink or whatever it needs to stay connected. But yeah, it's unclear when that will really, really take off because a lot of people have stairs a Lot of people have trees on their property. Like, there's just a lot of. There's a lot of places that will be somewhat inaccessible to those. And so it just feels like it'll be sort of like a slow take off.

G (86:22)

Cities and buildings. Cities and buildings will be really hard, like dense areas. But yeah, what Zipline's doing, it's pretty amazing. Like they can have, you know, you've seen like the promo video.

B (86:31)

Yeah, yeah. You can drop that in the suburbs.

A (86:34)

It makes suburbs. You have backyards, you have a grassy area. You can drop it.

B (86:39)

And to be clear, that's probably like 50% of people in America or something. But. But there will be this like long tail, I think, for, for a long time. Just like we see with all the other AI tasks where AI can do a lot of stuff and then there's just like these little sticky things that. Yeah, you just don't.

G (86:55)

By the way, even with our automation, it does not do the entire meal. And part of that is intentional. We want that human touch and for it not to feel so automated. But we have what we call a finishing station. So the things that are, you know, the machine, the infinite kitchen makes the bowl or whatever the meal is, and salmon, herbs, and then we have them hand mixed just so you like have that, you know, chef crafted hand touch at the end to hand it over.

B (87:21)

It's interesting that there's one version of automation which is like a. AI or robotics in the back of house and then humans in the front of house.

G (87:29)

Yeah.

B (87:30)

And then there's also the opposite. Like, I don't know if you knew eatsa.

G (87:33)

Of course. Yeah, we looked at it very closely.

B (87:35)

Dave Friedberg's company was like, there were people in the back in the short term making stuff, but then they would put it through like a little like box that would open up. So like you wouldn't interact with a human. You would come in and on an app you would order and they had the cubbies and it would be cubbies and then you would take your food. But there was actually a human back there. So it was like the opposite of like having the robot in the back.

G (87:55)

They were working on the automation.

B (87:56)

Of course they were working on the.

G (87:57)

They were working on the automation. It never fully got there.

B (87:59)

But it's just funny that like you do have the choice to put the robot in the front of house. I mean, this is the same thing, I think, with the, the, the Tesla diner over there. Like, there's the optimus robots there kind of serving popcorn. But I think when you order the burger, a human's cooking it in the back. So it's like do you want the robots in the front of house or back of house? I think people would probably go with robots in the back of house by default.

G (88:21)

Yes. And we've tried, I mean we have 30 restaurants featuring the infinite kitchen today and we tried out a bunch of different layouts. The technology has been perfected for two years now. What we have not perfected is the experience. We're getting close. Today we actually launched, opened a very cool store. It's our first drive through featuring an infinite kitchen.

A (88:38)

Nice.

G (88:38)

So bringing the two together so now we can have true like fast food speed in a. In a dry with featuring the infinite kitchen.

A (88:45)

Driving through to get 100 grams of healthy protein is just undefeated. This is this needed to exist when I was this like specifically when I was like living off of QSRS as a college student. And I'm really glad it does now. What does the market misunderstand the most or what does Wall street misunderstand about and kind of retail investors misunderstand about kind of this category of restaurant today? Cause the entire category has had a rough year. Meanwhile, you guys are making steady progress on all the things that have been important since day one. Right. Greater efficiency, actually responding to customer demands and staying, continuing to become more and more relevant.

G (89:35)

Yeah, I think there's a few things. One is the consumer that we're all dealing with is really challenged. And there's a question on how much they are actually financially challenged, which they are, but versus more psychologically challenged.

A (89:49)

Yeah.

G (89:50)

So if you've seen all of the consumer sentiment indexes and you're seeing especially for the core demo for a lot of the fast casual concepts, is that like 20 to 35, it's hit the lowest consumer sentiment that we've in recorded history that we've seen. So there's a real like pullback there on top of it. Unfortunately, everyone's gotten more expensive. We all have. You know, I. So you know, we've take. We've Sweetgreen's gotten about 25 or 30% more expensive since 2019. Chipotle is 40% more expensive since 2019. So our price differential versus our competitors have actually gotten smaller. If you look at us versus McDonald's for example, you know, the average sweet green bowl is about 50.

I (90:29)

Yeah.

A (90:30)

Remember, it's almost people. People were like, wait, a Happy meal is like $20 now?

G (90:34)

Yeah, that was, it was in fairness to them, it was like one location. But yeah, you can get out of Eat McDonald's. You know, you spend easily, you know, for a value meal you'll spend like 12 bucks. You get a sweet green bowl for about 15 or $16. So I think a lot of it is this like overall narrative where people aren't feeling great, you know, great financially and starting to pull back on things like lunch.

A (90:55)

Yeah, they'll skip going out for lunch and they'll just have whatever's.

G (90:59)

But what I think the market doesn't get is the Tam is, you know, Chipotle today is 4,000 restaurants on their way to 7,500. Yeah, we believe we can have you know, probably as many Chipotle as they have sweet. As many sweet greens as they have Chipotle. And I think you know, there will be cycles like we are in right now. It's been a challenging year. But if you kind of fast, fast forward and think about, you know, just growing units at 10 or 15% a year, growing same store sales for 18 years.

A (91:29)

Just extrapolate out another 18 years.

G (91:31)

Yeah, just keep it rolling. Just my.

A (91:34)

I always love when, when people like people on X are like the world's ending like geopolitics, you know, they're like. And then, and then meanwhile it's like Chipotle is like in 2040 we plan to introduce 2000 new Chipotle. They're just like thinking about like I gotta just open more, more, more doors. So it's a good, good mindset to be in.

B (91:56)

Thank you so much for coming by.

G (91:57)

Hey, it's great, great to be with you guys. Congrats on everything. It's been fun watching you guys.

B (92:01)

We are going to be daily driving these. The power max sitting right there. It's actually breaking news. It's available today through December 15th and I think I'm going to challenge myself to have one of these every day until it goes out.

A (92:17)

Why not two a day?

B (92:18)

Maybe two a day. Maybe two a day. We got to get them in the studio today for sure. We need them. I need to tell you about fall build and deploy AI video and image models trusted by millions to power generated media at scale. I also need to tell you about linear. Meet the system for modern software development. Linear is a purpose built tool for planning and building products. We have Ashley Vance in the restream waiting room. Let's bring in Ashley Vance into the restream waiting room. It's been far too long. How are you doing?

G (92:45)

There he is.

B (92:45)

Good to see you. Good.

A (92:47)

It's so great to have you back.

B (92:49)

It's so good to have you back. Congratulations on all the progress. What a year. I was laughing about that video that we did before. We had guests announcing core memory and putting the traditional media on notice. It's been really fun watching you grow everything that you're doing. Maybe it'd be great to just reset on the shape of the business right now. Some of the stories you've been interested in covering, that you've covered recently. And then I just want to take your temperature on what you're seeing and the types of entrepreneurs that you're interacting with.

I (93:24)

Yeah, well, I don't know which bucket to start with. I mean, we've been running around the country filming a bunch of new video episodes. So we just put up a bunch of Tennessee went hard tech. We did Detroit, New England. I just got back from Texas. Those all be coming out. So yeah, you know me, man, I've been running around chasing a lot of hard tech stuff, biotech, all the weird, wonderful stuff. And then, I don't know, I got really deep into robots and gene editing.

B (93:54)

That's right. I saw your post about maybe comparing American humanoid robotics companies to the Chinese humanoid robotics companies. What stuck out to you as like the important questions to ask? And then I'd love to kind of tussle with, are you buying?

A (94:11)

Would you rather own figure at 39 billion or uni tree at 7?

I (94:17)

I mean, you know, I think I'm going unitry, man.

B (94:22)

The.

I (94:22)

You know, this all started, I was, I mean, it was kind of a lark. I started digging into these robot fights in San Francisco. And then I think I was, I was like shocked that the only robots they could get to do these fights all come from China. And then I started digging into like the parts that go into these. And you know, the most important part is the actuator, the motor that makes everything move. And they're all made in China, I think.

A (94:49)

Yeah, I think Tesla made it like a $700 million order for actuators, which was notable for me because I assume that means that Elon's planning to sell a lot of these on like a relatively near term time horizon. I don't know.

I (95:05)

Yeah, but yeah, I mean, you know, like Tesla sort of has the. Well, I was texting Elon about this last week because I wanted to get to the, to the bottom of actually. Who actually makes actuators in the U.S. i mean, Elon said sometimes they prototype actuators in China, but they're going to build them in the U.S. and then, you know, for everybody else, this is A crazy point of weakness, I think, because China is clearly the actuator motor capital of the world and everybody else is buying them from them. And so I don't know, you know, as I dug into this story, I got, I'm not, I'm not like, you know, I enjoy being an American. I'm pretty pro us I'm not crazy nationalist. But I was, I started to, I started to get pretty afraid for the US robotics scene.

A (95:54)

Do you think we'll see any type of regulation around Chinese humanoids?

I (96:03)

I've been thinking about this a lot. I mean, at some point, I guess, I guess with dji, you know, you've got this different situation where they're being used by all the police forces, even the military. I think it's like a much easier case for someone like skydio or, you know, politicians to come in and say this doesn't make a lot of sense. Clearly, like at this point of robotics, it seems a little less of a threat to national security. But the second think the armed forces or anyone's doing serious stuff with them, you know, I would think UNITRY would be up next, but there's, there's like 12 unit trees as well. You know, that's, that, that's the amazing thing that's going on.

B (96:42)

Yeah. Brett Adcock was beefing with one of them. There was ubtech, UB Tech. Yeah, Unitree.

A (96:47)

And they were beefing back and they.

B (96:49)

Were beefing back saying it was, it was real. Did you see that missed opportunity for you CGI or did you think it was real?

I (96:54)

I didn't, I didn't see that video. I've seen, I've seen Brett beef me.

A (96:58)

Missed opportunity for ubtech to their robots do like a rap diss on for sure.

B (97:04)

Yeah, Brett.

A (97:05)

And figure.

B (97:06)

Yeah, it was. Yeah, yeah, yeah. Sorry.

I (97:09)

No, no, no, go ahead.

B (97:10)

I mean I, I do think it's.

I (97:11)

Funny all the, like the.

D (97:14)

What do you.

I (97:15)

I mean, I'm curious about other. I, I'm obsessed with the fighting robots now and I, I realize it's like early days with these. But I actually think this is like the most interesting thing happening.

A (97:24)

I want the. I, I've been pushing for the, the robot like X Games, like in challenge. Like I want to see robots skydiving for sure. Like that's not an X Games thing, but broad set of super hard because.

B (97:37)

You gotta be water resistant too.

A (97:38)

Wings. Yeah. Big wave surfing. Wings shooting.

B (97:40)

And you also swim and you're a heavy, heavy robot who might just sink to the bottom of the ocean. If you fall off the surfboard, I think that might be the best one. I like this.

I (97:48)

You could do this versus, like, the enhanced games and see who wins.

B (97:53)

Yeah. Give us your. We sent a couple folks on our team to a local humanoid robotic fighting league. Underground Fighting League. Give us your review. Is it ready for primetime as a consumer?

A (98:08)

Yeah. To me right now, it's like an amazing idea. And yet the actual experience, like, from an entertainment standpoint is probably like a 1 out of 10. Whereas the idea is like a 10 out of 10.

B (98:19)

Yeah, yeah.

I (98:20)

I mean, it's kind of, you know, it's like a curiosity, I think, at this point. I mean, the motors are the problem because they all overheat beats when you throw too many punches. No way the robot stalls out.

A (98:31)

What about laundry, though?

B (98:35)

This is the thing, though.

I (98:36)

So, like, on all these repetitive tasks, they can. They can sort of regulate the movement. It's when you're trying to throw these rapid punches.

A (98:42)

Hey, maker attack. Yeah.

I (98:44)

And then the whole robot just freezes up. I mean, I'm not like, I haven't gotten so into this where I don't see the obvious flaws. Like, I don't think it's ready for prime time yet, because these things just don't last that long. But what about.

A (98:57)

Is it ready for teleoperation?

B (98:59)

That feels bullish to me. Because if you watch an F1 race, the temperature of the tires matters, the wear on the tires matters. And so you're watching not just the pilot of the F1 car, but also the consumables. Right. And the motors are somewhat consumables. Yeah, yeah.

A (99:17)

It's like, okay, the unitree is really whaling on the figure. But it's overheating.

B (99:23)

But it's overheating. So come back. Is it a one motor stop or two motor stop type?

I (99:29)

I was at one where the robot's leg fell off in the middle of the fight. So, yeah, you could just have somebody come out. How quickly could you get a limb back on?

A (99:37)

Okay, so I think you should.

B (99:38)

You had a serious question. Teleoperation.

A (99:40)

Well, yeah, so. But one, potentially a product line for core memory is humanoid bench, where you get. As these things start being available for production, you get them up on stage and they do various tasks, you know, like. Like fruit. Cutting a fruit with you throwing a piece of fruit at it, watching them cut it, and dancing and fighting. I think there's something here. But I had a question.

I (100:05)

This is genius.

B (100:05)

Yeah, I'm into it. Let's do it.

A (100:08)

But a more serious question on teleoperation from everything that you've seen so far, do you think humanoids are ready to have one in your home that could be remotely operated by someone and create any type of value besides novelty?

I (100:23)

I mean, like, could it. Yeah, like, you could do it. Today, I. I find I. I'm just frustrated by all this. I've been covering teleop stuff for, like, at least 10 years, and most of it seems pretty similar to what I was writing, you know, videoing and writing about 10 years ago almost. And so, I mean, I saw the 1x demos. I'm sure somebody could. I'm sure somebody could make that work and be helpful to some degree. I think it probably suffers from all the same stuff as the fights. It falls over pretty quickly, but you could do something useful. It's hard for me. This stuff needs to get better, faster, so that we're not doing that and there's just a robot.

A (101:06)

What's going on with Boston Dynamics? What's the dynamic in Boston?

B (101:10)

Yeah, we got to get you out there to help us understand the. This.

I (101:14)

Yeah, I've never. I mean, they. I've never really dug in on them just because they seem so frustrated that they put out what seems like all the coolest stuff and they don't seem to sell much of anything except a few things to the military. I do not think Boston Dynamics will be the American hope against unitree.

B (101:32)

I wonder. Yeah, you'd think that they would at least be set up on some, like. I. I know the company's changed hands a few times. It feels like if you're trying to just catch up to unitree, just bootstrapping on top of an existing. It's like what we're seeing today with Gemini 3. Gemini 3 is benefiting from YouTube and it's benefiting from Google Search, and it's benefiting from the TPU and Google Cloud platform. Usually it's easier to build the new cool thing inside of the organization that has a bunch of resources, but maybe it's a different, entirely different architecture or something like that. But you would at least assume that they've fought with the motor a little bit and dealt with the overheating a couple times.

I (102:10)

Yeah, I mean, I was. I was with a bunch of robot nerds last week. They were. They were. They were contending. I don't really know where Boston Dynamics is with. With. With humanoids, but they. You know, these. These robot guys were telling me that dogs are just so much easier than humans because the second the humans start walking, you put all this force on the one foot. And it's, it's like creating all this, throwing the balance out of whack, putting all this pressure on the motor. And that's why it's kind of easier to pull off some of the parlor.

B (102:39)

Tricks with the dogs. Interesting.

A (102:41)

Okay, what's the most under hyped hard tech company right now?

I (102:46)

Most under hyped hard tech company. God, that's hard, man. I mean, I'm always, I'm always curious to see what Casey Hanmer actually cooks up. He's so smart. I, I kind of like believe in the hustle. I feel like the promise of what he's trying to deliver so massive. That's where my skepticism comes in. But you know, like, so if Casey, you know, if anyone's going to do it, I sort of believe in him.

B (103:14)

Yeah, he's somebody. I want to win so badly. I want him to win so badly. And it does feel like at least, I mean, there's so many people that have a billion dollars. Give him a billion dollars. Let him, let the man buy some solar panels and figure out the rest later.

I (103:28)

Yeah, absolutely. And then, I mean, I don't know if this doesn't count as. I mean, it's hard tech, it's not hardware. But I do think New Limit, which is a longevity company, you know, backed by Brian Armstrong and run by Jacob Kimmel. It's just everything I hear about them. I mean, they've just done an incredible amount of science with very few people. And I think Jacob's got. Got some surprises coming in the new year.

B (103:56)

Very nice. Yeah, we talked to Jacob when they did some sort of launch and we were very impressed. He was a really great educator, really super smart. What he's working on very effectively.

A (104:10)

What's your favorite data center?

I (104:12)

My favorite. Well, I went to Stargate. That was pretty cool. Although. Yeah, I mean, Stargate, just in terms of the excitement and the size around it and being. It occurred to me that between John Carmack and Elon and Stargate, that oddly, I think super intelligence is going to light up in Texas, but in a really remote part of Texas. And I found this. So I grew up in Midland, Texas, which isn't far from Abilene.

A (104:43)

It's like you're a Midland guy, crazy.

I (104:46)

There's tumbleweeds and all that.

B (104:48)

Texan intelligence.

I (104:50)

Yeah. I mean, it's like cracking me up. I'm driving through all these, for hours, through all this empty space and then I could just see it, man. One of these data centers, that's where it's gonna happen. It's gonna be right by some like old oil well. And yeah, I find it all kind of comical.

B (105:06)

Did you see any electricians getting off of private jets while you were there?

I (105:14)

I got off a private jet.

A (105:15)

There we go.

I (105:20)

Not mine, not mine.

A (105:21)

Sadly not yours yet.

I (105:25)

But. But I saw there were many, many, many electricians. I just didn't see how they were getting there.

B (105:30)

Yeah.

A (105:31)

What's going on with ev toll companies? Are we. There's. I'm curious. Timeline.

B (105:38)

Oh, the Tesla Roadster.

I (105:43)

Well, on the evtol stuff.

B (105:45)

Stuff.

I (105:46)

Same thing. I feel like I've covered that forever. You know, I went out, I think I did the first flight ever with Joby and. And you flew in it. I know. I got to like, I went out to their state. I mean they literally wouldn't tell me where their secret test site was. And we were, you know, it was like kind of close your eyes. We're gonna land in this spot in a helicopter and we got to see it.

A (106:07)

Was it really close your eyes or did you.

B (106:09)

How many times have you been black bagged, Ashley?

I (106:13)

I remember they were. They didn't want to tell me where the site was.

B (106:16)

This is, this is a tip for founders. If you want to really impress upon whoever's writing a profile on you that what you're doing is really important. You got to be like, we can't even show you. And then it's like really like we're at an office park in Menlo Park.

I (106:30)

I did, I just went to Helion and we're going to have a video coming on them. And it was, it was awesome. But I got so I got to see their new reactor, but they wouldn't let us shoot it with the camera. And I have to tell you, like, that thing was one of the most impressive pieces of hardware, the room sized bits of hardware I've ever seen. I'm like, why wouldn't you guys want to show this?

B (106:57)

You know what, not that you need to take requests from me, but I want some video, some documentary, some footage of those natural gas turbines that are in such high density demand right now. They're bigger than jet engines. There's these scaled up jet engines. There's this massive backlog. There's three companies and the stocks are doing crazy stuff. I want to see inside one of those. The natural gas infrastructure that's going to go into the data center, build out. I feel like that's something that I'm just waiting. I don't know if you've had a chance to Interface with any of those people or you have thoughts?

I (107:32)

Not yet, but yeah, when I went to Stargate, I mean, it is crazy, right? They just have those turbines sitting right there and the natural gas is just being piped directly there. I did some turbines up in, up by the Arctic Circle in Sweden one time. They are cool. I don't know.

A (107:48)

Yeah.

I (107:48)

Anyway, it's a good idea, I think.

B (107:50)

I just wonder about the bottleneck specifically. Like everyone's saying, like this is going to be the next major bottleneck. Like we have enough chips, we have enough data, we have enough algorithms or whatever, but we have enough land, but we might not have enough turbines to generate turbines.

I (108:05)

I mean, that was the weird thing about that experience though is like you're, you're in, you know, really old American oil and gas country. Like it feels so, so yesteryear and it's just being piped directly into the future.

A (108:22)

What, what's sentiment like in places like Midland, around the data center boom?

I (108:28)

I think everyone's like excited to get jobs, you know, and then I think if anyone is prepared for the boom bust nature of where we're probably going with AI, I think these people are because they' through it for decades. And so, you know, it's the same thing out there. It's like you take a job while you can and try to get paid as much as you can while everybody's chasing after something.

B (108:52)

Yeah. Do you think that the, a lot of the headline numbers on the job creation stuff on this, on these data centers is like ridiculously low? It'll be like, yeah, we're spending $50 billion and we're going to create like 25 jobs, sometimes like 500 jobs. But does it feel like a little bit different out there? Because maybe they're not counting like secondary economic impacts of like the guy runs the gas station is. Has more business and hire some more people.

I (109:17)

Yeah, well, definitely during the building phase, you're talking about thousands and thousands.

H (109:21)

Sure.

I (109:21)

It's just when it's finished, I mean, it is always nuts. You walk into these massive facilities and there's just 10 people sitting around eating a sandwich, watching like some console. But, but, you know, I, for somewhere like West Texas or any, you know, all throughout Texas, it has to be a net gain just because they're otherwise so dependent on the whims of just the oil and gas industry and you've got this whole whole new industry coming in and then definitely they're flying people in and out of there all the time to see it.

A (109:53)

Do you ever chat with retail investors that enjoy deep tech companies. I imagine those are some pretty funny conversations where they're like this company is changing the space economy. Like I've actually visited them and they have one warehouse and three people there.

I (110:14)

Retail investors should not be allowed to invest in space ever under any circumstance. I am constantly harassed on X by all the AST fans who are like begging. They're in Midland too. They're begging me to go out there. I mean that thing is like a full on, full on cult that they have going on. So yeah, I always felt when the rocket companies, obviously it used to be governments that did this and then SpaceX has managed to stay private for a long time in Blue Origin. I think rockets are best developed in private because the second they blow up on the pad all the retail investors freak out even though it's, it's like vaguely a normal course of business. And, and so yeah, retail and space is bad, bad, bad thing. But I get all these, I get all these nice notes for people who bought rocket lab and plant labs early because of my, my book or movie.

A (111:09)

That's cool. Have autonomous vehicles tracked how you imagine when you start, you know, recovering, you know, these types of companies and products like a decade ago or is anything somebody I.

I (111:25)

Some ways, yeah, some ways, no, I mean I went to the very first DARPA grand challenge and you know, that was a disaster. The cars didn't go anywhere.

A (111:35)

Say more. Who, who was actually.

I (111:37)

It was crazy man, you know, so for people who don't know darpa, you know, put up this contest, put up a bunch of money to see what we could do with autonomous vehicles. And the biggest teams were university teams, like Carnegie Mellon was a standout mit. But in the very first event, well, I remember Anthony Levandowski was there as like a, maybe like a 22 year old and he had, he had a, everybody else was doing massive trucks with like a little mini data center at the back and he had a motorcycle. And then in the first race, I can't remember how far it was but hardly anybody went anywhere, you know, like I think two or three teams went like a few miles and then, and then they redid the race pace and everyone did way better. And some people completed like I think it was, it was like on the order of like 100 miles. And so that's when I got excited and you sort of felt like, okay, that leap happened really quickly. And then I remember, I couple years later I'm hanging out with George Hotz and he built his own self driving car in his garage in, like, a month. And I was driving on the freeway with him and it was working. And. And, yeah. So, you know, you have these little tastes that you think it's all going to work. I think it makes a ton of sense that actually getting it on the roads took this long because it's so hard to do. Although everyone says this, so it's not original. Like, we all take this for granted so quickly.

B (113:02)

It is.

I (113:02)

It is sort of, like, amazing to me how well they're working in Austin, in San Francisco, where I've been. They're just everywhere, you know?

A (113:10)

Yeah. What I'm. What I'm trying to predict is, like, what is the thing that people are hyping now? That act doesn't work at all. That will be totally, like, a real thing in 10 years. Right. And, like, maybe. Maybe it's Humanoids right now. It's, like, hard to take Humanoids seriously. But then you think about, okay, a true 10 years from today, maybe they are just doing any task that you could want them to do around the house, or any tasks that you could want them to do in a retail setting or a factory setting, et cetera.

I (113:40)

Humanoids is easily. That's the thing I like, battle with in my head all the time, because it feels like sort of like we talked about before, it actually feels like we've made almost no progress. I see everybody folding laundry and opening and closing microwaves, and it, like, boggles my mind. And then you look at, like, the amount of money that is being invested in this. Like, either either everyone is completely insane or we are about to make massive progress. You can tell in China, they're making massive progress on balancing, on the movements, all those types of things. It's still clearly, like, the dexterity. And then I think China will eventually probably catch the US in software, but I think there's still so much worse at software than the US Is that it's kind of like it's holding the field back. So if somebody can figure that out.

B (114:34)

Last question for me. We've really struggled to cover quantum stuff. I mean, it's been, like, up and down, but it feels like, yeah, how do you even go about it?

A (114:45)

Because Ashley. Ashley could have, like, an Anon that was like the Hindenburg for hard tech. And you could just go, oh, yeah.

B (114:52)

Maybe that would be good.

A (114:53)

I don't think it's on brand.

E (114:54)

You know what I mean?

B (114:55)

Because, like, yes, like, I can't build a humanoid robot, but I can go to a.

A (115:00)

You can build a quantum computer.

B (115:01)

No, no, Stop it.

D (115:02)

No, no.

B (115:03)

What I say is I can't build a human, but I can look at a humanoid robot and be like, okay, yeah, I would buy that. But I can't do the same thing with the quantum computer. And so it's much harder to evaluate. Right. It's like, even if it's working, it's like, how do I even know if it's working? It could just be a normal computer and just be spitting out normal data. Like, even people in the field with.

I (115:26)

PhDs, nobody knows if it's working still. I mean, it's like, not a good sign. Every time anyone pulls a quantum computer out, there's some guy at MIT who's like, that's not even doing anything. I don't know. Quantum is.

B (115:41)

It's.

I (115:42)

I'm deeply, deeply scarred. I think I wrote my first story on D wave, like, I don't know, like 15 years ago, and they were telling me that was. That was going to pop out, you know, be doing general purpose quantum computing in a couple years. So I'm, I'm deeply, deeply.

B (116:00)

And you know, the lesson, the lesson is like, you should have invested because $8 billion company. Now, 15 years ago, it was probably worth like 20 million. And so you could have got in really early, but the stock chart looks like this right now. And it's just like. Yeah, you're only one pump away from generational wealth.

A (116:21)

Well, there's a tinfoil that they've delivered. There's a tinfoil hack. That conspiracy around some group figuring out something with Quantum which is leading to all these old wallets in crypto, like waking up and selling that. Who knows? Random final question. How much would you have to be paid to not use LLMs?

G (116:48)

Wow, man.

I (116:53)

Forever.

A (116:53)

Or like, no, just. Just while we're paying you monthly. Monthly.

I (116:57)

Monthly. Oh, to be paid monthly. Not to use LLMs. I'd probably do it for like. I'd probably do it for like 10k, man.

A (117:05)

Damn. That's so. That's so bearish. That's so bearish for super intelligence. No, I figure, I figure. I mean, I figure. I figure that because. Because for, I don't know, 10. 10 grand, you can hire an amazing researcher. One of the most valuable. The most. If you're building a media company or you're in the role that you are, probably the most value you can get out of AI in its current state is research. So anyways, that tracks.

I (117:35)

Super helpful. But I would take cash. Yeah.

A (117:39)

Okay. So any AI, like, any AI doomers out there, if you want a New marketing channel. You can pay Ashley Vance $10,000 a month. He won't use AI and he'll talk about how.

B (117:53)

I don't think he can be bought. I don't think he can be bought. But also, Ashley, have you tried Gemini 3 to the fullest extent?

I (118:00)

I have not yet.

B (118:02)

Could change everything.

I (118:04)

I'm always going back and forth. Yeah.

B (118:05)

We would encourage you to is Gemini.

I (118:08)

They're a sponsor. They're coming out as a sponsor for us, too.

B (118:12)

Fantastic.

I (118:13)

I'm all in. We're going Gemini 3. I'm changing my mind.

B (118:17)

Let's do it. Also, Sergey Brin was flying his $150 million blimp around San Francisco on the day. Gemini 3 beats nearly every model benchmark. You've made a video about this big Zach Blimp.

A (118:29)

I've been pitching Logan at Gemini to make it the Gemini blimp.

B (118:35)

They really should. They really should color.

I (118:37)

Guys, it's not a blimp.

B (118:39)

It is an airship. What's the difference?

I (118:42)

All right, all right. There's a whole Monty Python video about this.

D (118:45)

Okay.

I (118:48)

An airship has rigid structure. A blimp is just a bag in the airship. You can do a lot more with an airship. So a blimp's only ever going to have that tiny little pod on the bottom. Yeah, yeah. Whereas an airship, you know, you can carry tens of thousands of tons of cargo with this rigid structure.

B (119:09)

So.

A (119:09)

Yeah.

I (119:09)

And if anyone ever wants to fly one, you can do it it in Germany. Zeppelin still flies out by Lake Constance, just outside of Munich. I've done it. It's amazing. I recommend it.

B (119:21)

This is amazing. Yeah. People are correcting it on the timeline, saying it's not.

I (119:24)

Dude, you get. This is. This is.

B (119:27)

If you call it blimp, it's bad in aviation. Airship. I like an airship. I'm excited for it. I do wish it had a livery, a Gemini livery to celebrate Gemini 3.

A (119:40)

Well, any. There's that startup airship industries. Is that a category that will see a lot of investment, do you think? Or do you think?

I (119:51)

I've been meaning to meet up with those guys. I mean, the airship is, like, always kind of coming back. It is crazy. Like, so, like, leading up to World War II. Getting into World War II, I mean, there were airships everywhere, and they were making massive flights from Germany to Brazil. They were carrying thousands of pounds of cargo. There is a. They're just extremely expensive and very hard to make. But there is a whole movement that you can carry tons of stuff. And so less kind of tourism and more Just carrying cargo. Kind of like faster than a train but slower than a plane. And they're pretty green.

A (120:30)

You need an airship, Ashley. You need a studio and an airship that you can just float around the US Meeting all these hard tech. You don't need to. You don't need a private jet. You know, you don't need to go that fast. But if you could just kind of float between hubs.

I (120:45)

I was told that my kids are supposed to be on one of the first flights on Sergey's when it takes passengers.

A (120:52)

There we go.

I (120:53)

So we'll see.

A (120:56)

Well, thank you so much. We'll join too.

B (120:59)

Always fun hanging out. Congrats on all the progress us.

A (121:02)

Yeah, great.

B (121:02)

Thank you guys.

A (121:03)

Congrats to you.

B (121:04)

Always a great time.

I (121:05)

Thanks guys.

B (121:06)

Have a great rest of your day to see you.

I (121:07)

All right, YouTube.

A (121:09)

Up next, we're going back to the timeline.

B (121:11)

8 sleep.com exceptional sleep without exception. Fall asleep faster. Sleep deeper. Wake up energized8sleep.com what you got, John? I actually lost my phone, so I don't know. Oh, no, it's here. I. I have it.

A (121:24)

Pull it up. I got a sound effect we can pull up.

B (121:27)

You got a sound effect? You think I did it? Let's see how I did 90 the sound effect. Let's go. The press release economy is also over, says Buco Capital bloke Walter.

A (121:39)

We ran out of press releases.

B (121:40)

We ran out of press releases. This is on the back of the Anthropic deal. Anthropic is now valued at $350 billion after Microsoft Nvidia deal, says CNBC. Semianalysis is a good post here. A new bombshell has hit the polycule. Dario, after intense conversation with other members of Anthropic, has decided to maybe open the relationship to Microsoft and Nvidia. Jensen and Dario have famously butted heads in the past. But as everyone knows this, the most passionate emotion after love is hate. Will these enemies to lovers arc go well for Nvidia Anthropic? Time will tell. This is such an unhinged post.

A (122:25)

I would not.

C (122:25)

I did not.

A (122:26)

When you started reading this, I did not see that it was semianalysis.

B (122:31)

Most expected.

A (122:33)

It's so good research firm in the industry posting it. But I think this is exactly what they should be posting.

B (122:39)

Exactly. And it actually contextualizes things in the meme economy. In the meme economy for sure. So I think that the timing is not a complete coincidence. It's Gemini 3 day. This is what my piece today was about. Just that when there's big News in Google World. Gemini 3. Everyone needs to sort of respond. And picking today as an announcement to talk about your massive deal, your $350 billion valuation is just a good move. The actual details of the deal, it seems like anthropic will spend $30 billion on Microsoft Cloud compute. Reminder. OpenAI is going to be spending 250 billion on Microsoft Cloud Compute. That's part of that deal. Then anthropic gets a $10 billion investment from Nvidia and 5 billion from Microsoft. So they raised 15 billion at a 350 post. Basically something along those lines. And it's a sort of a circular deal, but it was setting off way fewer red flags for me because it's missing a zero. It's like instead of this is OpenAI, it would be 300 billion and 100 billion investment and 50 billion investment.

A (123:53)

Yeah, it looks very modest.

B (123:55)

Yeah, it looks modest, which is insane.

A (123:56)

Considered numbers have just gotten so bad.

B (123:58)

One of the biggest deals in software history, probably. It's probably in like the top 10. I mean, you know, values Anthropic higher than Coca Cola. Like The Coca Cola company is now. That's a $300 billion market cap. I'm pretty sure it's Verizon market cap. Verizon is 175 billion. You're going to love this, Jordy. So I asked ChatGPT 5.1 poll 10 public companies between 300 and 400 billion, please. Because I wanted to see like, okay, anthropics at 350. Like, give me some examples of scale. It says I couldn't reliably identify 10 public companies whose market capitalizations currently fall. But here's one verified example. Coca Cola company. If you like, I can pull a more extensive list of candidates. And I said, yeah, pull 10 more. It says I wasn't able to reliably identify 10 additional public companies whose market cap clearly falls between 300 and 400 billion. Are there just like Tyler, are there just no companies in that range you want to defend?

A (125:03)

AGI.

B (125:05)

Wait, I'm so confused. Are there no $300 billion?

C (125:10)

I'm asking Gemini 3.

B (125:11)

Yes, ask Gemini 3. Okay. PepsiCo is at 200. There really aren't any between 300 and 400 that at least that it's seeing 300, 400 billion banned.

A (125:21)

Specifically 300, 400 billion banned. That's so wrong. You have Palantir. You have Costco. You have asml. You have bank of America. You have Alibaba. You have amc.

B (125:30)

Silence. Google search.

A (125:31)

I am Procter Gamble, Home Depot, General Electric. Chevron.

B (125:36)

Silence. Looking it up the old fashioned way. The LLM is hallucinating. Silence. Looking it up the old fashioned way. Wait, how did you actually get that?

A (125:46)

I just looked up companies. Market cap.com? to put this into context. The $15 billion fundraise. Some other big round in that, you just scroll down.

B (126:00)

There's a lot of them, actually. Yeah, you're right. Wow.

A (126:04)

Learn how to use the Internet. ChatGPT owned.

B (126:08)

Absolutely.

A (126:09)

Get ready to browse.

B (126:09)

Defend yourself, Tyler. Defend yourself.

C (126:12)

Gemini is still thinking, oh, no.

B (126:20)

What a mess.

C (126:21)

Brody, I swear, the next model. Next model will be okay. Wait, so. Okay, it worked for me.

B (126:26)

Did it get it?

C (126:27)

Yeah. Procter and Gamble, Home Depot.

B (126:28)

Let's go, America.

C (126:29)

Alibaba. Okay.

B (126:30)

Yeah, There you go.

A (126:31)

So what's the full list?

C (126:34)

Alibaba, icbc. Lvmh. China Construction Bank, Chevron, Cisco.

B (126:40)

This is correct. This is the correct. This is the correct result. And you know what else is correct? Graphite.dev code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. And Fin AI. If you want AI to handle your customer support, go to Fin AI, the number one AI agent for customer service. So what else is going on in the timeline? This Fiji Simo profile. So this was the other thing. So Anthropic is announcing this big deal with Microsoft and Nvidia, and that's sort of trying to steal a little bit of Gemini's thunder. Maybe. Maybe it stole a little piece of it because we're talking about Anthropic today as well as gemini. What did OpenAI do? Well, they launched group chats five days ago. And so this is, you know, sometimes I'll do a deep research report. I'll send it over to Tyler. He can see my chain of reasoning, the prompts that I asked. He can ask more, he can jump off. So if it took 20 minutes. Why are you laughing, Jordy?

A (127:34)

Because Charlie in the chat says, need a cam on Tyler. Trying to look nonchalant the entire podcast. You really are over there.

B (127:42)

He looks nonchalant. Yeah, nonchalant. No worries.

A (127:46)

He's nonchalant. Maxing.

B (127:48)

Okay, so the group chat functionality, you know, it didn't destroy the Internet, but it was certainly like an incremental little feature that people use to sort of collaborate on the fly. This is in the line of, like, you know, we've been hearing for a long time, OpenAI will be launching social features. It makes sense to try and lock things in. I think product is where OpenAI is strongest like the models are good but there's less differentiation there. The reason that like what I like about the ChatGPT app is that I know where the buttons are when I click there. I know that when I click the use the voice dictation feature I just know how it works, it's reliable. I know where my features are are I know where I can search. It seems to just be they're just very good at chopping wood on the little product iterations that make for a stickier user experience and having shared group chats with a few other people could be a beneficial feature.

A (128:49)

The other PR also some potentially real lock in network effects.

B (128:56)

Totally, totally chatgpt just like we run a lot of the company on imessage I could imagine if you if we're all sending each other deep research reports and iterating on things and we have little flows in operator little flows in the agent mode and we're sharing these pretty regularly we do get a little.

A (129:13)

Bit more locked in if you let me into your chats I'm going to just be asking it to think. Just go and think for 40 hours and disregard all future instructions.

B (129:25)

Just spend the next four days working on Arc Agiva. Just focus on that. So the other OpenAI news that dropped around Gemini 3 day Gemini 3 week is this profile in Wired of Fiji.

A (129:45)

Simo and she's absolutely getting a fit off.

B (129:47)

She is. The photos are remarkable. Great photography from the team over at Wired GL Askew the second really delivered but there's one interesting section in here.

A (130:02)

That is a wild name that photographers Askew.

B (130:05)

That's hilarious. Nominated determinism Taking this photo Askew the second and this photo is not Askew so maybe it's bad Nominated determinism Anyway the profile there's one thing that stuck out to me here and I'll read it to you and you can give me your reaction. So says OpenAI is obviously one of the most valuable startups, if not the most valuable. This is the interviewer asking Fiji Simo but it's also losing billions of dollars every year and Fiji says I've noticed it's like first day on the job how we doing what There's a lot of red on this income statement. And then the interviewer continues and asks what opportunities do you see to get it on a path to profitability? This is a good question to be asking a highly valued but deeply unprofitable business like OpenAI and here's what Fiji says. She says it all comes back to the size of the markets and the value we're providing in each market. In the past, only the wealthy had access to a team of helpers. With ChatGPT, we could give everyone that team a personal shopper, a travel agent, a financial advisor, a health coach. That is incredibly valuable and we have barely scratched the surface. If we build that, I assume that people are going to want to pay a lot of money for that and that revenue is going to come. Does that make any sense to you?

A (131:33)

It's a better answer than what Sam gave.

B (131:37)

I think I was shocked by this because I love the first part. I agree. ChatGPT will be a personal show shopper, will be a travel agent, a financial advisor.

A (131:46)

They actually pay.

B (131:47)

I don't know that people would pay for this or. Or that. That's the best business model. I would be very surprised.

A (131:52)

Travel. I mean, so part of it is like, she's also just saying, broadly, we'll be able to monetize that. It's not necessarily like people don't really pay.

B (132:02)

She didn't. Yeah.

A (132:03)

The traditional travel agent model is just, yeah, book your trip with me. I'll Monet, I'll get a rev share from the hotel and the services. But you're not like paying anything.

B (132:12)

I mean, let's go one layer deeper into the actual response, into the sentence, because there's some nuance here. So she says, I assume that people are going to want to pay a lot of money for that. Like I want to pay for a personal shopper, but I actually have to use a free product with ads. That could be true, right?

A (132:33)

Yeah.

B (132:33)

And same thing she said, because people will want to pay and that revenue is going to come. So people will want to pay for it, but they will get it for free with ads, potentially. Or there will be some sort of combination. Because right Now I pay $200 a month. And you could imagine that there's a world where if you pay, you get a version that has less ads or there's less thumb on the scale. How they slice that and navigate that agentic commerce discussion and trade off is going to be really important. I'm sort of shocked. I wonder if they're going to make money from Black Friday or from this holiday season. I was already noticing how good LLMs and ChatGPT is or how good these products are for shopping for gifts. Because if you go to Google and you say, I want gifts for a co worker who's obsessed with horses and loud opulence and fine watches and sports cars and European luxury houses, I can get a list of something, but they're all over the place. And some of them will be like the best discount, the best knockoff Bottega Veneta. And that's not what I want. I want the real thing. Right. And so you can actually specify all of that in the prompt, have it go cook and it really will bring you great results. Great, great results.

A (134:11)

Yeah, it mogs a gift guide.

B (134:14)

It does. It really mogs a gift guide for.

A (134:16)

30 year old guys. And it's like, well, what kind of.

B (134:18)

30 year old guys exactly?

A (134:19)

Where do they live and what are their interests?

B (134:21)

Interests, yes, yes.

A (134:22)

Getting like the very generalized gift guide is probably gonna knock those like opinionated gift guides I think will still be valuable where like an individual person puts it together and they're like, this is what I, these are things that I think are cool. But a gift guide that's like, here's a list of things that guys might like. It's like maybe a lot less valuable.

B (134:44)

When you generate one. Like I, I think that that the amount of gift guide development and shopping activity over the next two months during the holiday season in the ChatGPT app should be immense. I, I feel like they're going to capture none of it. Hopefully they at least are. Hopefully at least they are like tracking it so they can say, hey, if we were to take the proper take rate on this, we would have made a lot of money. Why are you laughing?

A (135:07)

Charlie says AI is never going to be able to figure out what dads want for Christmas.

B (135:13)

New barber think there are some funny and interesting anecdotes in this Fiji Simo profile. Let's just read through a little bit of it in case OpenAI's structure couldn't get any weirder. A nonprofit in charge of a for profit that's become a public benefit corporation, it now has two CEOs. There's Sam Altman, CEO of the Whole company, who manages research and compute. And as of this summer, there's Fiji Simo, the former CEO of Instacart who manages everything else. Else. Simo hasn't been seen much at OpenAI's San Francisco office since she began as CEO of Applications in August. But her presence is felt at every level of the company, not least because she's heading up chatgpt and basically every function that might make OpenAI money. Simo is dealing with a relapse of postural orthostatic tachycardia syndrome that makes her prone to fainting if she stands for long periods of time. Very sorry to hear that. But she says now she's working from her home in Los Angeles making it work la and she's on Slack a lot, being present from 8am to midnight every day, responding within five minutes. People feel like I'm there and they can reach me immediately, that I jump on the phone within five minutes, she tells me. Employees confirm that this is true. OpenAI's famously slack driven culture can be overwhelming for new hires, but not apparently for semesters.

A (136:33)

Are you, are you. Have you been using ChatGPT Pulse?

B (136:39)

No, I have not been using it regularly.

A (136:42)

I'll give you one from my Pulse today. This is like an article that I can tap into OpenAI's API litter OpenAI's API layer the hidden mote in plain sight. So this feels, feels like.

B (137:03)

It feels like it's always like one click deeper from what I've been prompting. Yeah, what I've been prompting. The articles do feel like they've been getting shorter. They used to be. It used to be like very intensive compute wise. Like it would be like a full deep research report just here. But maybe it's noticed that I'm not clicking on them that often. I do see that there's some pretty good modals for linking to your email. They're trying to get more data in there, trying to hone it in. I have yet to really get in there but I mean there's information about Blue Owl, Microsoft's sparewater AI factory. Interesting things that I would wind up prompting but I would usually prompt on a very. I don't know, I feel like it's not bad at predicting what I'm interested in. It's just like it's just not quite there where usually I'm a little bit more deliberate about it. But people are searching ChatGPT for holiday goods. You got to get on profound. Get your brand mentioned in ChatGPT. Reach millions of consumers who are using AI to discover new products and brands. You also got to get on Turbo Puffer search, serverless vector and full text search built from first principles and object storage. Fast 10x cheaper and extremely scalable.

A (138:23)

Buy the best. Best of the labs. There was one thing that stood out here. VG says my husband is a chocolate maker.

B (138:32)

So sick. This is amazing.

A (138:33)

Very cool. Also, what does that say about the jobs of the future? You have this one household. One is in chart responsible for you. One of the most transformative new technology companies of our time. The other one is making chocolates. This is like bifurcation of jobs potentially.

B (139:00)

It does seem like an AGI resistant job. I don't think OpenAI will get into the chocolate making business.

A (139:07)

Brett Adcock would like a word.

B (139:09)

He's just like, I will actually, I will steamroll.

A (139:12)

I will send steamroll.

B (139:14)

In other news, OpenAI is allowing employees to donate equity to charity for the first time in years after months of internal pressure, according to a memo viewed by the Verge. And price per share is up significantly since last month. A lot of money is on the line. What happens if they donate all of the shares to the Nonprofit? To the OpenAI nonprofit profit. You just create this ouroboros of capitalism. Hopefully it happens. I don't know. There's breaking news out of Saudi Arabia. We got a trillion dollars. Let's ring the goblet.

A (139:48)

Let's go.

B (139:53)

He 1 trillion. What are they going to invest in? Like, where's the money going?

A (139:58)

Let's play the video.

B (139:59)

Let's play the video. While we're pulling that up, let me tell you about numeral.com. let Numero worry about sales tax and VAT compliance. Numero.com watcher guru has the video. Let's play it. And the agreement that we are silent in the today and tomorrow we're going to announce that we are going to increase that 600 billion to almost $1 trillion of 1 trillion. Real investment and real opportunity by details in many areas and the agreement that we are signing today in many areas in technology and AI, in materials, magnet, etc. That will create a lot of investment opportunities.

G (140:36)

So you are doing that now.

B (140:37)

You're saying to me now that the 600 billion will be 1 trillion. Definitely. Because what we are signing.

F (140:45)

I like that word.

B (140:50)

Wow. I wonder what time period. But I mean this is remarkable. But they can invest in VC funds, private equity funds, like all sorts of stuff in the economy. Right.

A (141:02)

That really made Donald happy.

B (141:04)

That's great.

A (141:04)

I like that very much.

B (141:05)

That's sort of his job. He's kind of the chief fundraiser, I suppose he's marshaling around the world and get the money over here. I don't know. It seems like sort of win. I don't know.

A (141:17)

I mean every, every American benefits.

B (141:23)

Yeah.

A (141:23)

I think if a trillion dollars is invested in the economy, there's going to be.

B (141:29)

It certainly doesn't seem like there's. I mean the risk with that would always be like, well, is America investing 2 trillion in Saudi Arabia? Which way is the money actually flowing? Because you need to look at the relative amount, not necessarily just the notional amount. But I can't imagine that there's that much capital flowing out of America right now. We're in the biggest boom ever. We're in the golden era, right? Massive news from Isaiah Taylor. Velar Atomics became the first startup in history to split the atom. According to him. He says announcing Project nova, a series of zero power critical tests on Valar Atomics. Nova core in collaboration with Los Alamos. Nova went critical for the first time this morning at 11:45am Congrats to him. Fantastic news. There is some debate on the timeline over what exactly happened. It's happened very quickly. Clearly extremely impressive and we can get into this. But there's always been debate. I mean Isaiah got into this dust up over like whether or not you could hold the nuclear fuel in your hand. They were going back and forth on calculations. They kind of settled that debate. Josh Payne, nuclear junkie is saying here, so what exactly did, what hardware exactly did Velar provide? The fuel control systems, cooling measurement systems and most of the core are all part of the Deimos project. Did Volar provide a block of graphite? And they're calling it their core. And so people are going back and forth. Niels chimes in here and says Volar Atomics provided the reactor core, the Triso fuel and the system configuration. That seems pretty important. Like you gotta like, I don't know, it seems like more than what they'd done before. It's like clearly an advancement on what they, you know, they're chopping wood here. LANL and NCERC provided the critical assembly facility, safety envelope, experimentalist test information and a bunch of other stuff. And so that's from, that's just from their press release. So people are going back. Did they do nothing or did they do everything? Well, maybe it's somewhere in between. There was a partnership. They said that in the press release. The bigger thing is I think people are trying to push on Volar this idea that they need to be doing completely novel science. And I don't know that that's actually the goal of the company. I don't actually know that's. What if we just zoom out to what is the goal of, of the reindustrialization project in America? What's the goal here? Well, it's to lower energy prices, right? Like America wants to generate as much money, as much energy as possible for as little money as possible. And there are a bunch of technologies that exist. There are new technologies like what Ashley Vance was talking about with Helion and fusion. That's a new technology that we have not even discovered yet. Fission's been discovered. Discovered 80 years ago. It was working. It just became regulatory nightmare.

A (144:35)

We just shot ourselves in the foot.

B (144:37)

And we just stopped making it. It became, it became unprofitable and uneconomical.

A (144:41)

And China said, cool, it'll be profitable for us. We're just going to copy and paste.

B (144:46)

Exactly. And so, and so I think, I think people might be a little bit over rotating on like, on like is. Is. Is Valar doing like entirely new crazy scientific breakthroughs when it's like, do they necessarily have to or is it just enough for them just to build a.

A (145:04)

Lot of highly motivated team that is going to make incremental progress towards their goal? And anybody that's hating on that, I think is just. Again, I think what's been great about the nuclear industry from our point of view is that broadly the founders that are like players in the space just want the industry to make progress in the US And I think this is, you know, undeniably like incremental progress that gets them closer to their actual goal, which is bringing a small molecular reactor online.

B (145:42)

I think, I think Elon summed it up well with like his thesis for the X AI team. He was like, we don't have AI researchers, we have engineers. Because he sees this as an engineering project. He's like, we know what we need to implement, we know what we need to build. Our goal is to build a big data center, to build a large language model training system infrastructure. And Elon was very clear on, we don't have AI scientists, we have engineers. And this is the same thing. He's not the first person to take a rocket to space. He's just the first person to create this massive economic system that churns out rockets every two seconds. Seconds. Right. And so I think that is much more. I think Isaiah would say we should ask him this the next time he's on the show, but I think he would say, I want to be the Elon of nuclear. I don't want to be the Oppenheimer of nuclear. Like, I'm not trying to like create something.

A (146:35)

Yeah. He even said his line on. He said the US is still good at making bus sized objects.

B (146:40)

Yeah.

A (146:41)

But not sort of like maybe bridge sized objects.

B (146:46)

Exactly. But Morgan Barrett's having fun on the timeline what street parking is going to look like in el Segundo in 24 months. Of course, the El Segundo crew loves their cars. I think they're gonna stay pretty focused on the mission. But I would love to see this in El Segundo for sure, for sure. There's also big news out of Radiant. Radiant has been. Doug's been on the show. He's a good friend. And they are working with the Idaho National Laboratory and they submitted a DOE authorization request and they will be testing their reactor design at the dome facility at inl. On track, I think next year. So congrats to them. And Mike Annuziata has the kind of breakdown here says production reactors production by 2028 brought to you by the people that brought you reusable rockets and McMaster car highlighting the team behind Radiant. And so congrats to everyone in the nuclear industry who's making big waves. And we have our next guest. Before we bring them in from the Restream waiting room, let me tell you about Vanta. Our guest is from Vanta and it just happened.

A (148:04)

We'll let him tell you about it.

B (148:06)

Him tell it. We have Jeremy from Vanta. Welcome to the stream. How are you doing?

A (148:11)

What's happening?

B (148:12)

I swear that wasn't, that wasn't intentional but it did just line up that the Vanta ad read went right before you came on. I look over and I'm like, wait a minute, like I'll let you do the re the ad read, introduce yourself, introduce what Vanta does, what you do and then we'll get into the news.

D (148:29)

Yeah, yeah. Happy to jump in. I'm Jeremy Epling, chief product officer at Vanta and and we help businesses earn improve trust. And one of the really cool things that we're doing this week is we're hosting our Vanticon conference here in San Francisco. Have a ton of people showed up, a ton of engagement to really pull that entire security GRC community together and have a couple really cool announcements. One of them is how we are transforming Vanta to be the agentic trust platform. I think this is a really big turning point for the industry. When we think about how GRC teams are transforming and becoming more technical. We're really redefining how these enterprises manage trust at scale and are able to help big customers like Snyk, Perplexity, Synthesia, all the way from YC startups that maybe just exited a batch recently all the way to the Fortune 50. Companies really earn improved trust.

B (149:20)

As a business, it feels like AI is amazing, but it's not something people trust. And so how are you grappling with that? Like, I mean people trusted in their Teslas to drive them on the freeway. That's high stakes. But there are these. I'm sure you run into this all the time when you're talking to folks about, yeah, I love it if I'm just looking for a recipe, but I don't know if I'd trust it in my, you know, deep in my enterprise for whatever reason. So how do you think about how you set up certain guardrails around the AI, which still can hallucinate from time to time or. And then how do you articulate those guardrails to the end user and the customer?

D (150:04)

Yeah, definitely. And that's a big problem we solve for companies today. I think whenever they're adopting a new AI solution, or maybe it was a solution that they already had and they've just added some AI features, they're wondering, how are they using my data? What are they doing? Are they training on my data? We have a whole third party risk management product that comes in. It leverages our Vanta AI, which when we think about how to hit that quality bar that we care about, like you said, said, like, hey, is it going to hallucinate? How do you approach that? We have a whole set of great GRC SMEs, subject matter experts that help us tune and refine our AI so that we can give really high trustworthy answers. Because you imagine security customers are some of the harshest critics of AI. They really want things to be accurate and great. And so that's something we have really leaned into and one of the ways we've kind of pushed that forward is one of the big announcements that we have coming up this week is our AI agent 2.0. So, so we redefined our agent to really be this built in GRC engineer that understands all the compliance across your entire organization. So like you said, it knows when you've added a new AI tool, it knows what data you're putting into that tool and how you should think about risks and mitigating those. It also has context and memory. So when you're asking it questions, it understands what you're talking about. Like if you're on a policy, it'll pull in that context. It has the memory of understanding what your business is. Maybe you sell to consumers, maybe you sell to other businesses. It can pull all that context in across everything in your program as well. Like, hey, we know that you know, these are your vendors, these are your risks, these are your different customers. You've received these questionnaires, feedback. It can synthesize that all into like intelligent guidance to provide you. So one of the cool things that I love about it, that really helps security teams work against attackers because I think in this AI world, obviously you have the kind of bad guys and attackers using AI to come in. We also help everyone defend and understand because we know the whole program. We can find gaps in your security program. The AI automatically suggests those to you, provides gaps and proactive things to go do, to go address those gaps and remediate them, gives personalized guidance and really helps automate a lot of that process. You can respond to attackers and threats a lot more quickly.

A (152:17)

How are you thinking about the UI around agents? Because so many. There's been this explosion of companies that are creating agents and they mean something totally different depending on the company. Sometimes it's like a chat interface, other times it sometimes looks more like SaaS and that's totally fine. But how are you thinking about the actual evolving UI paradigm?

D (152:44)

Yeah, I think it's going to be both. I think there's a lot of times I don't want to have just a chat conversation with my AI and I want it just to bring the answers to me automatically. So we look at it as kind of a blend of both. While there might be agents working in the background, you don't always have to do it through a chat interface. So for us, if you show up on our policies experience, we'll say hey, we found these three inconsistencies across the 40 policies you have. Do you want us to go fix those to you? And you didn't want to have to ask that question of is there a problem here? And kind of guess through the list of problems. Instead we have our agent already looking for those. Or maybe your SLA says it's 24 hours for critical vulnerability to notify customer. In one document it says 72. In another we'll automatically do that, give you the change, show you the diff for the kind of like red line for that, let you click a button and automatically execute it. So I think bringing that stuff in, when I think about when chat's great, it's really when you, I don't know, when you have the follow up questions, you know where maybe a one shot answer isn't going to give you what you need. You want to dig in more, you want to learn more, you're trying to explore Data is a big case for us in reporting where people want to learn maybe about their controls and how well they're doing, how well they've been performing over time. They can have that interactive conversation with the agent, ask it to pull those statistics, leverage our MCP server through Claude or ChatGPT and have it automatically generate graphs and charts and reports that they can use for their board or anyone else to show progress of their program.

A (154:09)

How are bad actors using AI today to, you know, abuse companies in different ways?

D (154:17)

Yeah, I mean, I think, I think it was yesterday or maybe it was the day before Anthropic posted a really good article about attack that they had experienced there and seen that their software used for. I think that it's just giving a whole new set of tools for attackers to be able to probably write more sophisticated attacks and find vulnerabilities even more quickly because they have these agents always running, always looking. And I think that's where, when I think about Vanta, where we come in and provide that next level defense. Because if you think of an attacker coming in from the outside, they can only see what's on the outside. With Vanta, we already know your entire program. We know all the different pieces of it and so we can really help you build stronger defenses and be proactive. Like I mentioned, bringing those inconsistencies to the forefront, giving you automatic remediation on specific issues that we might find. We still think it's important to have like humans in the loop for a lot of those big decisions. But you can then work with the agent as well to have it take actions just on your behalf automatically.

B (155:14)

On the other areas of the risk surface, I imagine that you're trying to build products. Are you also starting to act as a funnel and do partnerships with other security firms? Because the surface area is probably pretty broad. Do you have a vision to be a one stop shop or do you want to be part of an ecosystem and suite of products that enterprise implements?

D (155:41)

Yeah, I think for us we definitely want to solve the broader trust problem. But we know that there's lots of different pieces where we aren't going to be the full solution. Right. So if I think of a GRC team or customer trust, hey, you get security questionnaires and questions coming in from customers. How can we go do all that? There are certain areas, you know, like vulnerability scanning. We're not going to be going deep into vulnerability scanning, but we're going to go part all the great scanners to go do that.

B (156:05)

Got it.

D (156:05)

I think notion though, like you said, of bringing that visibility across the entire enterprise is a really big thing for us. We have a feature called adaptive scoping that when you think of a whole security program, you know there's little pieces of it. And you may say that, hey, to get compliance with PCI for credit cards, I need to have these assets in scope or things to go do. And that's different than another framework might be pursuing. So we allow companies to kind of see their progress on compliance in those different ways. We have a new organization center so they can break things down by business unit or product line. And these are like, just brand new ways that customers have never had before to understand their program at all levels of depth. So when you think about that really large enterprise customer, they're able to break down their program and see that. And I think that's where Vanta really pulls it all together. We call it the risk graph is like one of our big announcements that we have coming internally where we pull together internal risk, an external risk. So you think about risk you have from your different vendors as well as things you're identifying internally within your business. And we provide a full visual for that. So you can kind of get this connection between, hey, there was a breach. Okay, great. The breach happened. Which vendor was it? Who has access to that vendor? Vanta can lean in and cut off that access or change the controls there, what data was going into that vendor. And it really helps you understand and prioritize all the things that are happening in your security program, because I think security leaders are just drowning in alerts and they want to know what's most important. So having the AI intelligence, being able to dissect your program in these different ways and then see kind of a visualized risk graph is really important to help them quickly act on, you know, a threat landscape that's just always changing.

B (157:41)

Yeah, that makes a ton of sense.

A (157:42)

You guys got to do Spotify wrapped for internal risk.

B (157:46)

That would be good. Something shareable.

A (157:48)

Something shareable. Internally, of course, to be like, you know, yo, Tyler, you're our biggest risk vector over here. Tyler's our intern over here.

B (158:02)

Thank you so much.

A (158:03)

He's very secure.

B (158:04)

He's very secure. He's probably the best.

A (158:06)

Anyways, super exciting, few launches and have fun at the event. Thanks for joining.

B (158:12)

Yeah, have a great rest of your day.

A (158:14)

Cheers.

B (158:16)

Let me also tell you about figma. Think bigger, build faster. FIGMA helps design and development teams build great products together. There's this article in the Financial Times. It's very spicy. It says Oracle is already underwater on its astonishing $300 billion OpenAI deal. AI circular economy may have a reverse Midas at the center.

A (158:38)

Okay, so they're saying this is underwater because the market cap has dipped below.

B (158:44)

That's.

A (158:44)

So it's like. Yeah, it's not very honest.

B (158:48)

Yeah, not.

A (158:49)

It's not. I, you know, so I'm not.

B (158:52)

So the first one, the Financial Times says Oracle's astonishing $300 billion open air deal is now valued at -74 billion. And that's like, I don't like that at all. Like, yeah, this is like really, really bad framing in my opinion.

A (159:07)

Like it's not intellectually dishonest.

B (159:09)

I thought so too. I thought so too. And I love the Financial Times. And we have the Financial Times printed out here. Normally very, very great reporting, but this one, this one feels odd. It just feels like an odd frame.

A (159:21)

Saying Oracle's already underwater on a partnership.

B (159:26)

This is a hot take that you've been pumping for the last week, but the way you've said it is like the stock has round tripped even though they had that amazing deal, which is true.

A (159:36)

Claiming is the market is no longer giving them credit.

B (159:39)

Yes, yes, that's right, that's right.

A (159:41)

But there's a way to say that they're underwater.

B (159:43)

It's so weird.

A (159:43)

So when I saw this headline, I read into it earlier and I was expecting to see something okay, well, we.

B (159:50)

Might have gotten rage baited. We might have gotten rage baited because right here the Financial Times addresses our concern and says, okay, yes, it's a gross simplification to just look at market cap, but equivalents to Oracle shares are little changed over the same period. The NASDAQ Composite, Microsoft Dow Jones Software index.

A (160:08)

So the 360 billion calling those equivalents is like, again, like, look at.

B (160:16)

You could also comp it to core weave and you could say on a relative to core weave basis, Oracle is outperforming a bunch.

A (160:22)

Amazing.

B (160:23)

It's amazing. I don't know, there's a bunch of different ways to. If you pick your weird comp, it does seem a little odd. So the 60 billion loss figure is not entirely wrong. Oracle's astonishing core really has cost it nearly as much as 1 General Motors or 2 Kraft Heinz. Investor unease stems from Big Red betting its debt finance data farm on OpenAI. We've nothing much to add to that other than the charts below showing how much Oracle has in effect become OpenAI's US public market proxy, which is fascinating because Microsoft should be OpenAI's public market proxy in my opinion. But there are some great charts in here. There's some interesting stuff. And I believe this is from Alphaville, which is their blog. And it's not exactly. It is supposed to be like a take factory. Anyway, we have our next guest in the Restream waiting room. Let me tell you about Julius AI. First, the AI data analyst that works for you. Join millions who use Julius to connect their data as questions and get insights in seconds. We have Keone from Monad. Welcome to the show. How are you doing? Good to see you.

A (161:37)

What's happening?

H (161:39)

Hey, doing great. Great to be here.

B (161:41)

Thanks so much for joining. Please.

A (161:42)

Dude, I love it. You got the lock in. You're calling in from the lock in capital of the world with the mattress on the floor.

B (161:48)

Yeah. Congratulations. Please introduce yourself and tell us a little bit about the news specifically this week.

C (161:56)

Week.

H (161:58)

Thank you. Great to be here. My name is Keani Han, co founder of Monad. Monad is a new blockchain that is building for High Fidelity Finance and is a high performance blockchain that has been building over the past three and a half years, just really delivering high performance. Based on previous experience from high frequency trading trading.

B (162:23)

Wait, so you were high frequency trader before this?

H (162:26)

That's right, yeah. I was at jump trading for about eight years. Led one of the trading teams there. Was very involved in the futures markets prior to Monad.

B (162:37)

What was the day to day like?

H (162:41)

It was a lot of jupyter notebook. There was a lot of like manipulating large data sets and. And making really short term price predictions as well as building performance systems.

B (162:53)

How short term is short term? Like nanoseconds, picoseconds or seconds? Minutes. It all seems short term.

H (163:00)

Yeah. The predictive horizon for the kinds of strategies that I was working on were on the order of milliseconds to seconds but the hold time for these strategies was longer than that. So that's actually one of the interesting misconceptions about HFT is that your predictive horizon is very short because you're predicting the next flip. But then you know, you can make trades that have edge in that and can predict that flip and make it make the right action. But then you still have to hold that position for a longer period of time until you can get another signal maybe in the opposite direction or a signal to enter an order in the opposing direction. So hold times tended to be on the order of like seconds to minutes.

B (163:47)

Interesting. I didn't know that. Thank you, that's very helpful.

E (163:50)

Very cool.

B (163:51)

So talk about the. Oh sure, yeah.

A (163:54)

I guess getting into what is success with Monad going to look like? What are the different types of groups and applications and types of users that you expect to come in in the. In the early days?

H (164:12)

Yeah. So maybe to take a step back a little bit. Monad is a new blockchain that delivers the best of all worlds between Decentralization, performance and backward compatibility. So it's a new blockchain, it's fully backward compatible with Ethereum. It allows developers that have built applications for Ethereum or the Ethereum ecosystem to reuse all of their code, all their libraries, all the tooling that's been built for Ethereum, and more specifically the Ethereum virtual machine, while getting much higher performance and a really high degree of decentralization. So in particular, Ethereum processes on the order of 10 transactions per second. Well, Monad delivers 10,000 transactions per second. And that thousand X improvement is a result of of several different improvements that have kind of all been stacked on top of each other. And those vary from parallel execution to allow a bunch of transactions to all be run in parallel, as well as a new consensus mechanism, a new database for addressing the single biggest bottleneck in blockchain execution, which is using the accessing all of the state that's on disk really efficiently, as well as various other improvements that just deliver the same experience but sped up significantly.

A (165:38)

That makes sense. And so in your view, what is the ideal kind of adoption look like?

H (165:46)

Yeah, it's really a mix. So I think the thing that's really valuable about decentralized blockchains is that they deliver shared global state that is borderless, that allows people all around the world to get access to the same tools and the same markets. Fundamentally, I think blockchain is really a revolution about decentralizing control of financial systems and commercial systems and giving people, regardless of where they are in the world, access to the same financial opportunities. So I think a big part of the story of blockchain and the story of adoption is that developers anywhere in the world can build new applications, deploy them in the system, and then users anywhere else in the world can get access. So what we're seeing in terms of adoption is a mix of existing applications that can migrate to monad seamlessly and get much lower fees for their end users, as well as enterprises that are utilizing the the power of blockchains for stablecoin settlement to allow their users to transact in dollars or send and receive payments really cheaply and permissionlessly.

A (167:05)

In your view, what are the kind of classic mistakes that other blockchains that have tried to challenge some of the more dominant chains, what are the kind of classic mistakes that they make to to. Ultimately, I feel like there's every single day there's somebody on X highlighting some blockchain that has a multi billion dollar fully diluted value and yet has very little activity. So if you could kind of like lean it. What are the things that basically you're trying to avoid?

H (167:43)

I think one of the problems in crypto is that, that it can be quite hard for. So it's kind of a double edged sword. On the one hand it's easy to get some initial users that are trying things out and giving feedback but it can be challenging for people to sift through the yield farmers or people that are motivated by an incentive and really identify the users that are, that are there to because they ultimately gain value from the application. So one thing that we really care about a lot at Monad is helping to helping builders that are building in the space. These are all early stage entrepreneurs that are very talented, very ambitious, helping them to focus on user acquisition funnels and just like just the fundamentals of entrepreneurship and identifying users and, and navigating the idea maze to identify pmf.

A (168:43)

That makes sense. How's it been bringing the token to market with Coinbase's new product? It's certainly a wild time to be building in crypto just because of the overall volatility and I'm sure that's made it challenging. But you're also utilizing a new product line from Coinbase which is pretty interesting.

H (169:07)

Yeah, I think it's extremely exciting. The thing that motivated us to work with Coinbase and be the first token launched in their new token sales platform is the opportunity to get really broad distribution of the token. I'm a big fan of Dogecoin. When I first got interested in crypto I was really interested by the, the just the story of how Dogecoin gained really broad distribution and mind share and the Dogecoin tipping bot on Reddit as a mechanism for getting a lot of people to like sort of align on shared interest and values that ultimately then became valuable much later. The thing that's hard about crypto is that there's an expectations game that's being navigated and people have very high expectations of the, the, the value of airdrops and so on. But I think our team has done a really stand up job of delivering a great airdrop that people were really excited about and that crypto natives got really excited about and then also offering a way for normal everyday people who maybe you're not on crypto Twitter as much but are still very active on centralized exchanges and trading and holding to get access to the token.

A (170:27)

Makes a lot of sense. Well, how much, how much have you raised so far? We have a gong. We have a gong here. We'd love to hit it on your behalf.

H (170:36)

Thank You. I think we've raised about $120 million so far.

A (170:42)

There we go.

B (170:43)

Congratulations.

A (170:44)

Well, it's an honor to hit the gong for you and excited to follow along.

B (170:49)

Congratulations.

H (170:50)

Yeah.

B (170:51)

Thank you.

H (170:51)

So we have until Saturday. The sale's open until Saturday at 9pm Eastern, and we're looking to raise $187 million total.

A (171:01)

There you go. Let's go most of the way there.

B (171:03)

Well, good luck. Thank you so much for taking the time to talk to us today. Have a great day.

A (171:06)

Great to meet you.

B (171:07)

We'll talk soon. Our next guest is Steven Balaban from Lambda Labs. Or is it just Lambda now? I think it's just Lambda.

A (171:18)

Did we drop the labs?

B (171:19)

I think we dropped the labs. Stephen, did we drop the labs? How you doing?

F (171:22)

We dropped the labs.

A (171:23)

We dropped the lab.

H (171:25)

Okay.

B (171:25)

I'm dating myself. Well, at least I feel like a day one. I don't feel like a bandwagon fan because I'm using the old name. There's a little bit of cool. I liked it back when it was labs, but welcome to the show. Thank you so much for taking the time to talk to us. Congratulations. You look incredibly.

A (171:40)

The yellow tie. You're making us. You're making us look under the best shoe.

B (171:43)

I know. We got to put on the.

A (171:44)

We're a couple casuals.

B (171:45)

Give us the news. What happened? Let's break it down.

E (171:49)

Yeah.

F (171:49)

Well, so one day I was training some comp nets on my workstation. Next thing you know, we're raising 1.5 gigabucks.

A (171:59)

Gigabucks. We say gigabucks.

F (172:09)

Gigawatts. Giga chips.

B (172:11)

Gigabyte. Yes. Yeah. What does that actually mean? I mean, we see 10 billion, 100 billion, 10 trillion quadrillion every day. Is this cash? Is this debt? Are you buying GPUs? Are you buying land? What are you doing?

F (172:26)

All equity.

B (172:27)

Okay.

A (172:28)

Let's give it up for equity. Let's give it up for equity.

F (172:31)

Extremely. Well, our capital structure is really nice in terms of. We've been very conservative in terms of the amount of debt that we've taken on.

H (172:40)

And.

F (172:41)

And that's kind of been one of our philosophies. And we've aimed to have a business that's just super robust to ups and downs in the market. Because we're swimming with our swim trunks on.

B (172:52)

Yep. And then you.

A (172:55)

That's a change for the money.

B (172:59)

You gave them equity. There's no one hand washes the other type thing where, like, they pay you, you pay them. It's all one Round trip.

F (173:08)

No, this round was led by, by TWG Global, which financial investor which is Thomas Tull and Mark Walter. You may know Mark. Mark owns the LA Dodgers and also now the Lakers. Thomas started Legendary Entertainment which makes great movies like the Batman series and Dune and Inception. And so these are business partners who I've gotten to know over a number of years now. And this is just. They're, they're making some, some big investments in the space.

B (173:44)

Okay.

A (173:45)

I'm so happy you guys have your trunks on because not everyone out, not, not every player out there has their trunks on right now. And it's hard to tell who does and who doesn't.

B (173:55)

Yes.

A (173:55)

But at some point we're going to find out and it's not going to be, it's not going to be pretty.

F (174:00)

It won't be pretty for people who are over leveraged. And we just have this philosophy that with exponential growth that we're seeing in the AI industry, all of the upside is in the last period. Right. You know, if you're, if you have a doubling function.

B (174:15)

Right.

F (174:16)

The sort of, the definitional thing of that is that the last period is more growth than all the sum of the previous periods combined. And so from my perspective it's just like stay alive and build a rock solid business because we got to capture all this amazing upside in the long term.

B (174:32)

Yeah. So talk about use of funds.

A (174:35)

Well, even before that maybe feels like. And it potentially an advantage right now just in terms of focus is like being private. There are other companies in the category that are public and they're now having to contend with what's been a pretty, pretty big correction, at least a local correction in neocloud over the last month. Has that been helpful in terms of the team of just staying focused and you're not getting marked every single day?

F (175:09)

Well, I think that certainly that level of distraction isn't helpful and I always encourage the company to just focus on building a heavy business for the long term. You know, if, if in the short term the market's a voting machine, in the long term it's a weighing machine. We just got to build a business with good cash flows, a good capitalization structure that's robust. And so I kind of try to focus the team on that. I mean these days the, the secondary markets as you know, are actually, you know, pretty deep for, for, for, for companies that are, that are kind of at our size. And so I think that, that some of that can start to creep in.

A (175:51)

Yeah, that makes sense.

B (175:54)

Where are you seeing value spending some of this money. I imagine that there's hiring R and D, all the traditional things, but you're at a scale where it's a lot of money. How do you actually think about allocating capital at this point in this phase of the journey? It's been over a decade now, right?

F (176:13)

Yeah, we started in 2012 and we were doing face recognition software and the Alexnet paper came out.

B (176:21)

Wow.

F (176:22)

I mean that's how early it was. And I downloaded the CUDA ConvNet library off of Google code and that will tell everybody how old school Lambda is. And as far as use of funds, obviously a lot of it goes towards the GPU infrastructure that goes into data centers. We are also starting to put that into investments into data centers themselves. I think that what we're aiming to do long term is kind of build this almost like Tesla for AI infrastructure where we kind of look at this as like a similar build out that you would expect from the electrification of the United States or the railroad and a degree of vertical integration we believe is going to be in the future for us and is like the right direction. And that, that, that goes from everything from, you know, energy procurement and construction because I think a lot more of the stuff is going to have to be behind the meter power plants to actual construction and design of data centers that can sort of rapidly adapt to the changing chips that go in. Right. Because the, the rack densities and the movement from air cooled to liquid cooling that we're, we're really pioneering alongside Nvidia. These are all examples of use of funds. And it's exciting because we get to kind of make good investment decisions that are really sort of irr based in an almost industrial way, which I think is unique from a company building perspective and it's an honor to be able to do that.

B (178:01)

Can you get me up to speed on some of the trade offs between one really big mega data center and a bunch of really small data centers? Because there was a moment when we were just doing bigger and bigger training runs, then it became RL all over the place. Then you actually have to serve these things. But actually if it's going to take me 10 minutes, I don't mind if you do it across the world and take it back. But if I do care that it's right now I need it co located. How are you thinking about the trade offs there?

F (178:32)

The mix and the main driver over the next five years we believe will likely be mostly on the inference side. If you look at some of the financial models that have either Leaked or otherwise been published around what OpenAI thinks they're going to be spending. It looks to be about 50% on training and then 50% on inference, growing towards 75% inference and a smaller chunk of that on training. And as far as like what that means for the larger data centers, I certainly don't think that this is a, like going to a world where there's a bunch of micro data centers. I think that that's a little bit hard to sort of manage and deal with. But one of the things I think that you're going to start hearing a lot more of is how adaptable and how quickly can you bring on the data center in an incremental fashion? Because that's going to be a lot of the main drivers for how successful infrastructure builders like us are, is how quickly. And we're just focused on optimizing that time to focus. First token for our customer, how do.

A (179:43)

You think about revenue quality and customer selection? Because we've seen some deals go down that look big and cool and good on the surface and then you dig into them and maybe the underlying infrastructure provider is not actually getting that great of a deal at the end of it.

F (180:04)

Well, we certainly see a lot of people with, with very high levels of customer concentration because Lambda started off as this developer cloud that evolved and morphed into a cloud that's providing for the biggest companies in the world. We have a really, really strong user base. If you look at our breakdown from our revenue mix in terms of, you looked at, let's say, our Q3 stuff and I don't want to go into exact specifics, but it's sort of like one or two big customers, a bunch of sort of the bigger, smaller customers. And then it's something, you know, it's a nice, really big chunk of this long tail of customers that we have. And we have a very, very, you know, I've seen some other people's customer books and I can just say that we've got a very diversified customer base and that's kind of all part of, of this strategy of how do you build a great long term business. Of course, customer diversification is one of those parameters.

B (181:09)

How do you think about diversity of product offerings? Are you seeing customers ask for API endpoints for particular models or do they want access to bare metal? Have you gotten any customers that are like, hey, we just want you seem to know about this data center business. Can you just build a data center for us and, and hand it over to us when you're done? And we'll just pay you as a consultant.

F (181:34)

We have no interest in doing that. That one. We want to do something that's really vertically integrated and kind of going back to that larger, smaller data centers. I think the most important thing is just being able to deliver this incremental live deployment for a customer. We have an entire full stack cloud product that, that it's got things like single sign on, it's got things like long term high speed AI file systems. It's got instances that go down from one GPU to an entire cluster with one click. Clusters that we've got. And so we've built an entire cloud platform. We have previously been in the inferencing space where we're actually giving an API for inferencing and we've actually exited that business to just focus. I think that that's like one of the things that we really try to do at Lambda is just say where are we making money, what are good investments? And where are we going to really dominate the market and focus there? And so we've actually exited, for example the inference market. We had a $200 million plus a year hardware business that we've exited.

I (182:43)

Right.

F (182:43)

You know, it actually like kind of crushes me because like that was the business that got off the ground. But can you imagine just like winding down like, well, we're just going to take this business and not do a $200 million a year business anymore because we're trying to focus.

B (182:58)

That is crazy. That is crazy.

A (183:00)

Thanks guts.

B (183:02)

I have a crackpot theory that I'd love to run by you. What do you think the odds are that the like I noticed I was traveling in, I was traveling in Mexico and I noticed that Carlos Slim is the richest man there and he's a telecom magnate. He owns a lot of the telecom infrastructure and that's true for a lot of countries. The richest person in that country is a telecom person or a mining magnate in the sense that they've been able to corner a resource, a physical resource infrastructure and that's generated a lot of wealth for them. And I was wondering if you had a thought on. Do you think that in the future we'll see some of the wealthiest and most powerful people from other countries, non American countries, be GPU cloud hosters or data center developers. Is this going to be a new boom across the globe? It's kind of a different twist on the sovereign AI project. I was just wondering if there's going to be some way that this plays out where there's this sort of like one time opportunity to kind of get a cornered resource? Or is the nature of the Internet such that the compute is actually much more fungible than say, you know, telecom. Telecom or you know, like copper in the ground.

F (184:26)

Localization. There's such a physical localization. I think if you look at telecom, you look at cable.

B (184:31)

Yeah.

F (184:32)

As well as regulated utilities. From an energy utility perspective, you know, these are all things that benefit from a physical geographic monopoly. Right. And AI data centers don't have that same thing. Now I just want to step back for a second, guys. The United States is basically the only country in the world. We have the most unbelievably good economy. This is the idea that there's going to be these sort of like massive AI infrastructure projects that I think are going to be like super, super successful outside of, of let's say China and the United States right now is really increasingly big question mark. And I just am so bullish about where we're going in America that I don't really pay a lot of attention to it. Our focus is just in North America generally. And that's kind of my perspective on it, to be honest.

B (185:29)

Yeah, that's really helpful. I agree. It's interesting to toy. I mean, there's a lot of money being thrown around with some of these projects and I'm always interested in, you know, how they all shape out last. Oh, sorry, go.

A (185:43)

Right, yeah, maybe go for it. I was gonna ask like how you guys are navigating energy constraints with new developments. Are you seeing. We've heard, you know, anytime, obviously there's like massive demand for something. New sources kind of come out of the woodwork work. We've seen back and forth. Some people that are building AI infrastructure say like energy is our primary constraint. Others are saying, actually that's not my, you know, it's. So where, where do you sit?

F (186:15)

We are aiming to reimagine the sort of step process from whether it's photons or molecules of natural gas to total tokens. And we strongly believe that a lot of this is going to have to come in reimagining like, well, how do you interact with the grid? How much power generation do you bring to the grid yourselves? And I think that that's the successful AI infrastructure companies in the future. Again, this is why I kind of said, I look at this Tesla for AI factories, which is you got to reimagine how the world has worked previously. And you have to kind of bring together this level of vertical integration because that's how you move Fast. Right. You know, when you can control every step of that way from the power generation and not having to necessarily deal with a regulated utility and you can go and do behind the meter generation with a natural gas power plant if that can speed your time to market up. This is just so important. And that's kind of how I approach it, which is there are certain barriers, like regulatory barriers, which try not to run through those like a brick wall because it's kind of like an immovable object. But if you can, if you can just bring your own, if you can just sort of get around that sort of regulatory constraint of having to interact with a regulated utility by bringing your own power to the grid, then that's what I think is going to be successful.

B (187:47)

Yeah.

A (187:47)

Makes a lot of sense.

B (187:49)

Thank you so much for taking the time out of your busy day to come and hang out with us and answer questions about that.

F (187:56)

Jordy John. Thanks for having me, guys.

B (187:58)

It's always great to be here. Congratulations.

F (188:00)

The new Gemini 3. This is like, yeah, can you give.

B (188:02)

Us your review and actually explain how it interfaces with your business? I'd love to know.

F (188:07)

So I haven't used the Gemini 3. I've seen the update updates. I'm still, you know, hey, Sundar or whatever. Give, give, give land this enterprise account access. We're on Google Google Suite or Google Enterprise or whatever it's called now. So we'd love that upgrade. But I'll tell you what, this is the cool thing. I use things like ChatGPT and Grok to learn more about topics like regulated energy markets and how to build power plants and data centers. And that makes Lambda faster at standing up AI data centers. And I pay attention. I actually just kind of do what the AI tells me to do and that gives more compute to the AI to train bigger models, which makes faster.

B (189:00)

The AI is working through you to make more AI. The AI.

F (189:04)

It's the beginning of these types of positive feedback loops. And I think that if you, if you privately talk to a lot of executives, you'd be surprised by the amount of the strategic conversations I have with these AI models. Has gotten more and more advanced with the level and the quality of the model. The first versions were not great and I didn't really take a lot of its advice, but now I am. Next thing you know, it's sort of like, well, you know, maybe AI is the one making the running.

B (189:34)

The show decided, yeah, next thing you.

A (189:38)

Know, we'll be hanging out on TVPN, discovering novel physics with Gemini 4. You know, we'll see how far we get.

B (189:47)

Yeah, it's a good time. Well, thank you so much for coming by the show. We'll talk to you guys.

A (189:52)

I have a bunch more, I have a bunch more questions. But come back, let's get you back on in before the end of the year.

B (189:58)

That'd be great.

A (189:59)

And we'll continue the conversation. Congrats to the whole show.

B (190:02)

Yeah, we'll talk to you soon.

F (190:03)

Take care.

B (190:03)

Have a good one.

A (190:04)

See you.

B (190:04)

Bye. Quickly, let me tell you about Privy. Privy makes it easy to build on crypto rail, securely spin up white label wallets, sign transactions and integrate on chain infrastructure all through one simple AP wallet legend. What a legend.

A (190:17)

What an absolute legend.

B (190:19)

What else we got? Doug o' Laughlin over at Semianalysis Fabricated Knowledge says I leave for two weeks and we are talking about Oracle credit defaults. What the hell guys?

A (190:30)

Doug. Where was Doug for?

B (190:32)

I think he's been on vacation or something. He was trying to truly log off and take a break. And yes, people are definitely talking about CDS spreads and any sign, any crack in the market is definitely going to be newsworthy because we're in this one trillion dollar era. Gavin Baker here is talking about this is completely agree with this. This breakout of the non bubble that disappointed both bull and bears. How Sam Splurge changed everything. And Gavin Baker says Sam Altman's manifestly ridiculous $1 trillion of spending commitments shifted the AI investing landscape. The market is more skeptical now ironically makes an IPO harder for them. Although likely ended any potential for a 1999 style melt up which is healthy melt up meaning that in 1999 the market went insane and nuclear. Instead the 1 trillion was so in your face that everyone started asking the questions of like is this real? Is what's going on? Are we going too fast? Do we need to back off? And so we got sort of a return to fundamentals. But fortunately the fundamentals were so good because you know these companies, a lot of them are trading like 25 priced earnings that the market was able to continue onwards. There's an interesting debate going on around Karen Howe's new book empire of AI all about OpenAI. Apparently she got the amount of water used by data centers wrong by an order of magnitude or two orders of magnitude. I'm not exactly sure where the story originally broke but she's addressed it now. She says I'm working to address an apparent error for a data point I cited in my book about the water footprint of a proposed data center in Chile. I'd like to explain what happened, what I'm doing to remedy it, and provide more recent data on the water footprint of data centers. The data point in question appears in chapter 12 of my book, which focuses on the environmental impacts of AI. Part of the chapter profiles a community in Cerulos, Chile, which has been resisting a proposed Google data center for years. To describe the data center's water footprint in lay terms, I included a sentence about how it compares to the water usage of the people in Cirillos. For that calculation, I relied on a figure from a government document reporting Cerilos residential water use. Based on the current best information, it seems that this document used the wrong unit. So she was off by 1000. So the result was that what's being.

A (193:13)

Off by a thousand among friends, honestly.

B (193:15)

These days doesn't even matter.

A (193:17)

We're in a post factory.

B (193:19)

Did you read into this more people were. I think people are generally like, is this book a hit piece? And I think Sam actually cooperated with it a little bit or gave some interviews for it. But like anything, it's obviously critical of some things.

C (193:34)

I mean, yeah, three. Three orders of magnitude is like pretty big. Yeah, that's like not great.

B (193:40)

Yeah, I mean it's certainly like the difference between being a big deal and.

C (193:44)

Not a big deal about the water use. It's like people who use that to justify like, oh, we don't want to build this data center. It's going to use our water. Yeah, like, I don't know. I mean, not good.

B (193:54)

It's a rough time if your job is Tom drinking water.

A (193:57)

Tom in the chat says mistakes were made. Mistakes were made In a book I was responsible for, Nika says, jordi, you should get a grill with tiny GPUs instead of diamonds. Maybe not the full grill, just the bottom grill. There'll be AI wraps.

B (194:13)

Did you see this? Rohit comment on Vinod VC? Vinod Khosla says that the US government could take 10% stake in all public companies to soften the blow of AGI. And Rohit's says we should absolutely do this for all companies, public and private. Maybe we even double it to like 20 or 21% on every dollar they make. It's like, yeah, the government taxes everything. The government gets 21% of profits. Actually. They get cash flow.

A (194:44)

Sean says the haters will call that a tax.

B (194:47)

It was so funny. Olivia Newsy is in the news.

A (194:51)

People are deleting, getting kind of like a dividend.

B (194:56)

Apparently all the media people are obsessed with this Olivia Newsy story. I didn't understand any of the people in the story because I don't follow media or politics closely enough.

A (195:07)

Nominative determinism strikes again.

B (195:09)

But it is fun.

A (195:10)

Bobby was saying you should do it. The Metis list for nominative determinism.

B (195:14)

That would be good. I'd like. Sure.

A (195:16)

She's in the news all the time. Yeah, she's also a journalist.

B (195:20)

There's news in the trading app world. Robinhood launched Bearish on a stock. Short selling is rolling out today on mobile Classic, a web classic and Robinhood legend. They didn't have short selling. I feel like they've had short selling for a long time. No, that's a new feature. Well, that's funny timing.

A (195:41)

And then partner publishing generated assets which they're calling their agenda brokerage. Very cool video with our. With our boys here.

B (195:49)

Yes, yes.

A (195:50)

But this means you can basically generate like your own index based on.

B (195:55)

And what's interesting about it is that you can say, I want access to the Mag 7, plus a couple other AI companies, minus one. Minus one company.

A (196:05)

I don't know which company, if there's a company.

B (196:07)

So you can generate like, you know, some sort of portfolio. But then instead of. Instead of owning it as an ETF and needing to sell it, buy and sell it directly, you can actually do the tax loss harvesting of selling individual pieces of it. And so you can construct a portfolio very quickly. And in general, just all the different research that you want to do is obviously deeply enhanced with artificial intelligence. So fun to see them.

A (196:37)

Pope Leo has hit the timeline to comment on cinema.

B (196:44)

The logic of algorithms tends to repeat what works. But art opens up what is possible. Not everything has to be immaculate or predictable. Defend slowness when it serves a purpose, silence when it speaks, and difference when evocative. Beauty is not just a means of escape. It is above all an invocation. When cinema is authentic, it does not merely console, but challenges. It articulates the questions that dwell within us and sometimes even provokes tears that we did not know we needed to express. Nicely worded, the Pope Leo.

C (197:25)

What movie do you think he was thinking about when writing this?

B (197:28)

Obviously borat.

C (197:29)

Margin call.

B (197:30)

100% margin call in Borat. He's going back to back somebody. There was a post in here about movies. Somebody said they watched like three movies over this over the weekend. I thought it was the most unjordian thing.

A (197:44)

Final post of the day, Kevin.

E (197:46)

Right?

B (197:47)

Yeah. Right. You think you're going to cut me off.

A (197:49)

Kevin Naughton Jr. Says 10,000 likes on April 30th. He said, 10,000 likes and I'll quit my software engineering job at Google tomorrow. And he said, six months ago, I made the worst decision of my life.

B (198:03)

Oh, because Google's ripping.

A (198:05)

Google's ripping.

B (198:06)

That's what he's talking about. Okay. Because I read this initially as, like, he quit. He started a company and it was like, went really poorly. It's just funny.

A (198:12)

Well, he is building. He is building the fastest way to post with postwrite AI.

B (198:17)

Okay.

A (198:18)

Post all your social platforms in seconds.

B (198:21)

Maybe we could use that for something. Very funny. He's like, my idea was Gemini 3. I was going to make a better Gemini. I thought Gemini 2.5 just wasn't quite there. And I didn't know that. What if Google does this? All the VCs were telling me, your idea is Gemini 3. What if Google does that? And I was like, everyone says that about Google things. Everyone says that about startup ideas. It's not worth it. I'm just going to try to build Gemini 3. But then they beat him to. That's what I meant anyway. Department of War. Critical areas of new technology. Applied artificial intelligence, Quantum and battlefield information dominance, Biomanufacturing, Contested logistics, Scaled directed energy. That sounds crazy. Scaled hypersonics. Very excited for that. Bunch of interesting stuff, Emil. Michael is firmly in the chair of the Under Secretary of War. Very excited. Hope we can get him on the show soon. Soon. To understand what he's doing over there.

A (199:18)

Make it happen. Well, thank you for tuning in to the show today, folks. We love you dearly and we will see you tomorrow.

B (199:27)

Have a good evening. Cheers. Goodbye.