002: Karpathy is back!

Ian Landsman (00:00) Welcome to Tocantown. Today is Thursday, May 21st. I am your host Ian Landsman. Aaron (00:06) And I am your other host, Aaron Francis. The OpenAI employees are about to get super rich. Carpathi is back. Elon loses on a technicality and Google has new products that are all named very similarly. We're gonna cover it all, but before we do that, we gotta give a shout out to Bento. Ian Landsman (00:25) Bento over at bento now.com the email marketing and CRM platform built for the AI era. Send your product marketing and transactional email with Bento. use Bento. Aaron uses Bento. Fantastic product. Check them out at bento now.com and thanks to Bento for supporting the show. Aaron (00:46) OpenAI is allowing employees to sell up to, wait for it, $30 million each in secondaries. So the per person cap was raised from 10 to 30 million. The valuation is 500 billion, some say 400. There's a two year vesting period and this is the first cash out for many employees. So what the hell is gonna be going on in San Francisco? when these jokers are walking around with 10 to $30 million. Ian Landsman (01:15) Yeah. Why don't we own a Lamborghini dealership in San Francisco? That's where the money is. Forget the internet. Aaron (01:25) That is where the money is. So what, mean, is this the new PayPal mafia? Are they gonna be seeding companies? Are they just gonna be buying single, you know, single family residences in California? Yeah. Ian Landsman (01:29) Yeah. Getting a little condo. I don't know. mean, that is interesting. Like all the money is going to flow out of this and start. I mean, this isn't even all your shares, right? Like this is you were capped at 30 million. You have more shares than this. So I'm more fascinated. Everybody's going to go off and spend a bunch of money, right? That's obvious and fine. But like, how do you keep these people motivated? How do you keep somebody who just got 30 million dollars motivated? Like, again, you're working in AI. It's the future. All the stuff. Aaron (01:44) No! I don't know. Ian Landsman (02:04) Man, I got 30 million bucks plus I have another 100 million sitting there for a little bit down the line. I feel like I want to go maybe take a trip, golf a little. Aaron (02:11) Yeah. I don't know. Yeah, I don't know. I feel like some, I think some people are addicted to the juice. So they're just gonna stay in the high-paced, like I'm in the center of the universe. That's gotta feel pretty good. But man, let it be known here. If I got 15 million, if I got 10, if I got 30, I'm out of the game. I'm doing fun stuff. I'm not gonna keep working for the man if I got 30 million. But I'm curious to see if this, ⁓ Ian Landsman (02:35) You're out. Adieu. Aaron (02:45) turns into angel investing or if it just turns into inflated housing in an already inflated area. That's kind of my curiosity is do these people start doing the PayPal, spread the money everywhere sort of thing. Ian Landsman (02:57) Yeah, it's so different now than then though, you know, that was like kind of a special time. There wasn't really that many individuals just like getting those big checks back then. Like there was a few, but like here you have like 700 people. All of them are getting 10, 20, 30 million plus lots of other startups right now. So many people have money at least on paper. Yeah. I mean, you would think the next generation of startups, this could be an interesting time to be raising money for a startup. have all these people now with money that they're going to want to invest in new, uh, Ventures right like they want to be involved in that kind of thing. mean like the Ben bites newsletter is like AI newsletter He's got a fun like people just have funds, you know, it's like way I've got some money. I'm just doing AI funds. You got some infrastructure. You got a dev tool come see me I'm raising money. I'm sending checks whichever direction Yeah, but Aaron (03:33) Mm-hmm. They do. Well, if you work at OpenAI and wanna let me wet my beak, give me a shout out, you can find me. Speaking of people that don't need money but are still working, Karpathy is back. He's back in the game. Okay, so give us the rundown. What did you think of this? Ian Landsman (03:52) Ha ha ha ha. ⁓ yes. I mean, I to say like, I'm not ⁓ totally up on the Carpathi background and all the drama and the stuff. I know he was like a co-founder of OpenAI. Obviously it's very interesting that he's jumping back in with Anthropic. Anthropic, believe, was entirely started by OpenAI employees to begin with, right? ⁓ So yeah, you have these two. I don't know. It's very interesting. What do you think about the difference between Anthropic and OpenAI? Like, do you think it's as significant as... Aaron (04:28) Like spiritually? Ian Landsman (04:29) spiritual difference, right? Is it really is that true? Is it real? Is it fake? I can't really get my hands on it. Okay, you think it's real? I think that's true for sure. Aaron (04:35) I think it's real. think the animosity runs deep. I ⁓ think the forward-facing dev rels on Twitter have fun with each other, but I think the hatred runs deep. ⁓ Ian Landsman (04:46) Mm-hmm. But do you think Anthropic is "less evil" than OpenAI? Aaron (04:55) I think they think they're less evil than open AI. I think Anthropic was started by the X-Risk, the effective Altruist, all the people that like really looked at ⁓ AI and thought if anybody builds it, everybody dies, so we should build it instead of someone else. Ian Landsman (04:57) Okay, yeah. So that was my impression, right? But the problem is that every time they're on TV, they're like, you know, we're literally destroying the world here. You're all going to be out of a job in a month. Like that's what they say every time. They literally say it every time. So like, well, that's weird. Like if you're saying you're trying to stop this from happening, but then all you ever publicly say is we are destroying the world. It feels like an odd. I'm not sure that I can just say, assume that they are so much better than open AI is. Aaron (05:22) Yeah, Dario's always depressed. I know, I know. I know, I haven't read the manifesto, so I haven't read the machines of loving grace. I think Dario is very sincere in everything that he is saying. don't think he's putting on a false front. I think he really does want to make machines that allow humans to prosper. I think his PR skills are below negative. Ian Landsman (05:44) Mm. Mm-hmm. Aaron (06:06) Whoever's the worst PR person in the world, he's worse than them. ⁓ But I don't know, Carpathi going there is interesting, because he's former OpenAI, former Tesla, then he retired to be a YouTuber, and now he's back in the game. ⁓ So I'll be curious. And I think he's explicitly working on recursive self-improvement at Anthropic. ⁓ Ian Landsman (06:29) Right. Yeah. Aaron (06:32) But yeah, just the choice of going to Anthropic versus OpenAI is extremely interesting to me. Say what you will about Dario and his public pronouncements of the end of the world. ⁓ It seems like people tend to trust him more, even if they don't like him, they tend to trust him more than Sam Altman. ⁓ I never really take anything I see on Twitter very, it's like seriously as a source of truth, but it does seem like a lot of people, like there was that, Ian Landsman (06:49) Mm-hmm. Yeah. Aaron (07:02) What was his name, Ronan Farrow? Did the big hit piece on Sam Altman that was like, this guy's evil. ⁓ So I can't speak to that, but it does seem like Anthropic ⁓ has the vibes of being less evil, ⁓ whether that is accurate or not. But this is a big move regardless. Ian Landsman (07:05) Mm-hmm. Well, I do think there's like the proof of that. Not that really that they're evil, but the perception of that is just in that like, at least on Twitter, devs seem to think codex is better model. But meanwhile, Anthropic seems to be growing much faster, right? Like the ultimately businesses are putting their money into Anthropic over OpenAI. And presumably at least part of that, I think part of it is they've embraced business more, but also part of it is probably a trust factor situation. ⁓ Aaron (07:33) Yes. Ian Landsman (07:49) Yeah. So I do think it's the other thing that was interesting about him going back there or going in there was that he's like pretty low on the totem pole, like not at the bottom, but he's not like he's way below the founders. He's got, he's reporting to some other person. So he's like, you know, three or four steps down. Maybe he wants it that way. Yeah. Give it to me. Oh, this is always good. Yeah. You get to pick your spot. Yeah. Yeah, that's true. Aaron (08:04) I'm gonna hit you with a sports reference. When you're Michael Jordan, it doesn't matter what your title is. When you're Wayne Gretzky, it doesn't matter what your title is. You want another one? When you're Mickey Mantle. Hey, three different sports, three different people. Elon Musk loses in court. It's over. Game over, my friend. What do we think? Ian Landsman (08:17) Whoa! Baseball one! Look at this guy. I love it. ⁓ Where'd we go? We lost him. yes, so Elon Musk lost in court. ⁓ you know, I thought it was obvious he was going to lose from the beginning. So, yeah, it was obvious. This was a spike. Every, if you read into it, it was very just like a spike thing. Nobody thought he could win anything at this. Aaron (08:39) Nonsense, it wasn't obvious he was gonna lose from the beginning. Get out of here with that. ⁓ Of course it was spite. No, no, this is the wrong take. This is supposed to be a one minute segment and you're gonna make me go longer because you have the wrong take. The case was not decided on the facts or the merits of the claim. So the case was decided on a legal technicality. Ian Landsman (08:51) No, it's the right take. This is a quickie. Yeah, there's a new quickie. Mm hmm. ⁓ Aaron (09:06) and ⁓ legal technicalities are the best kind of technicalities because they're legally enforceable. So the ruling was ⁓ after weeks of ⁓ trial and presumably many tens of millions of dollars of lawyer fees, the ruling took two hours and it came back with your claim is barred because you expired the statute of limitations. So no finding on the actual facts. ⁓ Ian Landsman (09:09) Yeah. ⁓ Mm-hmm. Aaron (09:35) If you wanna play the game, you gotta play by the rules. That's too bad. All right, next. Ian Landsman (09:42) Yes, what's up next? Google moves to the kind of industry standard AI plans. It appears of 200, 120 bucks as the norm, which I think is good. I do have a Gemini. I have a work Gemini plan. I have another Gemini plan. It's like it was all crazy. It was very complicated that even how to buy it was quite complicated. So it seems like they're cleaning it up. Seems good overall to have it be semi standardized. I don't know. What do you think? Aaron (10:08) I'm, I, and we can cover this in the next one. I have a hard time keeping up with anything Google is doing, to be honest with you. I believe the former $250 plan is now down to a hundred, but they also added a top tier of 200 and there's a $8 and there's a $20 and does it apply to Google AI Ultra Studio? Does it apply to Gemini only? So. Ian Landsman (10:14) Right. Right. Mm hmm. It is mixed in with also quantity of data in your account. So like you might have more data for like Gmail. It's not just AI. It's a whole Google stuff package. So you're getting a bunch of Google stuff in addition to the AI. Aaron (10:42) and presumably you're still getting YouTube premium if you subscribe to AI Ultra. Ian Landsman (10:45) Yeah, I didn't look, but I think maybe I have YouTube premium like seven other ways, but yeah, I still have to pay for it. Like, whatever. There's a one of those things. Anyway. Aaron (10:49) Of course, yes. So who can say things have changed? We have no opinions because we don't know what the hell they're doing. And speaking of Gemini 3.5 flash was released ⁓ just recently claiming to ⁓ be 4X faster speed and half the cost. I'm dubious. I press X to doubt on that. So it is a flash model. The knowledge cutoff is like January of 2025. Ian Landsman (10:57) Hahaha. Okay. is it that far back? Okay. Aaron (11:23) Yeah, it's super far back. So it's just like a, it's not a new pre-trained model. ⁓ I saw a couple, of course, all benchmarks are made up, ⁓ just like all statistics. I saw on the cursor bench that it's like super far down. It's like almost like number 20 on the cursor bench and it's incredibly expensive. ⁓ Ian Landsman (11:33) Bye. Mm. wow, that's crazy. Yeah, they up the price a lot. It's very weird the terminology they're doing where the old flash one was like flash light and very cheap and fast, but now this one's expensive and fast-ish, I guess. It's not totally clear. Aaron (11:53) Yes. Yes, so the fact that it's a flash model that's incredibly expensive doesn't make a ton of sense to me. So that part's a little confusing to me. But they've also announced Gemini Omni. What the hell is Gemini Omni, Ian? Ian Landsman (12:12) I'm not Omni is a different model because we don't have enough models. We got a new model. And this one, now this is the one that the deep mind guy is super into. This is his baby. Yeah. That this is the one that he thinks brings you to, to, uh, AGI. So this is like focused on give it anything in and it can spit anything out. Whereas most of the other like video models and things are texting video out. They're specialized trained. This is supposed to be more abstracted. Aaron (12:14) Ha ha ha ha! Demis? Demis? Yeah. Ian Landsman (12:41) I don't know how it works, you're asking the wrong guy. But anything in, anything out is the goal. I don't think it's actually even quite there yet. But it's a whole different family of models, ⁓ which I guess now we will have some level of access to. I don't know if they have pricing for that one yet, but it's not for coding specifically. But the idea is that this is the one model to rule them all, at least in there. Aaron (13:00) Yeah, so theoretically this is, ⁓ I believe this is a world model. So it's not just like a pure video generator. It's got some sort of like long contextual understanding. It's got physical understanding, gravity, kinetic motion, all of that stuff. So I believe one of the demos they showed was like a video that was generated of a marble rolling through a marble maze and the physics were at least directionally correct, if not perfect. ⁓ Ian Landsman (13:07) Right. Mm-hmm. Aaron (13:30) And it combines, I think it combines DeepMind, which is Demis or whatever. DeepMind's NanoBanana, VO, Genie, all of those models, which are apparently a different family of models altogether than Gemini models. Ian Landsman (13:35) Mm-hmm. bright. Right. It's not replacing those though, I don't think right now. Like it's not like NanoBanana went away and this is the new version. Like there is still NanoBanana, which is a whole separate thing. And then there's this thing, which is going to do all the same stuff as that thing, but differently, apparently. And there was like, we're not going get into it, but there was like 20 other Google announcements of various sorts. We did the quickies on the top four or five here, but ⁓ yeah. Aaron (13:51) No. So before we hit what's in your stack this week, I just wanna give a PSA to Google. Clean up your product offering. Nobody has any idea what is available, where or how to log in or pay for it. What in is in your stack this week? Anything new or removed? Ian Landsman (14:21) guys ⁓ Yes. So I've just did a little stack tweak, still with the cloud, still with the backup sort of a cross checking with the codex. ⁓ but I formalize that a bit more. So I've have been doing that thing where like I build something, whether it's a spec or plan or the actual produce code. And then I will write up a little thing to say, spin up solo MCP, which is your tool for managing all this stuff and, ⁓ have check whatever this document is or documents with codex. it goes off and does it. Sometimes I tell it to do two and whatever. But it's getting annoying writing all that. So now I have a skill, which I have very few skills. I'm not a skilled guy, a custom skilled guy. You have to earn your way into that tier. So now I have a custom skill that is just called like reviewing codecs. And then I can just say reviewing codecs and it'll figure out what I've been working on and say, okay, like the last thing we did was a spec. So I will review the spec or I will view the plan or I will review the code. And it Aaron (15:08) Agreed. Ian Landsman (15:26) just spins up to automatically. I would like your take on this. What I'm doing right now is actually two. One is architecture and one is more like ⁓ bugs and like the actual nitty gritty detail type stuff. So right now don't really have them both look at the same thing, challenging each other, but I have Claude that built it originally and it's reviewing everything. So I don't know, I kind of thought going broader would be more interesting. So that's where I would set it up for right now. What do you think about that? Aaron (15:37) Mm-hmm. I like the double review. would maybe, if I were doing it personally, would change it to like a architectural and product review because if you do architectural and bugs, somebody, the first one may say change the architecture and it obviates the bugs or change the architecture and it introduces new bugs, you know? So I don't know that I would stick with architectural and bugs. ⁓ Maybe do them sequentially, do the architectural and then do. Ian Landsman (16:01) Mm-hmm. Mm. Yeah. That's your dude. Yeah. Mmm, could do a two-step. Aaron (16:19) know, review and codex bugs, review and codex architecture. I know, I know. ⁓ But what's in my stack this week? I am still using Solo, which is my product, soloterm.com. And in there I use GPT ⁓ Extra High 5.5. I now have three subscriptions, three pro subscriptions to GPT. ⁓ One tool, Ian Landsman (16:22) Two steps though, I don't know, it's annoying. Okay. Token Maxer. I don't know if you be saying that publicly here. It's all right. OK. Aaron (16:50) It's okay, nobody will ever know who I am. ⁓ One tool that I have also added to my tool belt, it has gotten better, is Quiver AI. ⁓ So Quiver is a specific model that takes ⁓ pixel pictures, pixel images, and turns them into vectors. And there are some of these tools on the market, but many of them ⁓ are very blunt. And the SVGs that you get back are not like, ⁓ Ian Landsman (16:55) Mmm. ⁓ yes. Okay. Aaron (17:20) lines and Bezier curves and the actual SVG primitives, it's just a mess of points. It's like, it's brutal. But Quiver AI has done something. I don't know if they post-trained some open source model or what, but they've done something where the vectors that come back are super clean and could be handed off to a proper vector person, whether that's an Illustrator designer or some sort of motion graphics thing. Ian Landsman (17:23) Mm-hmm. Mm. Right. Just a bunch of yeah. Mm-hmm. Mm-hmm. Mm. Okay. Aaron (17:49) Quiver AI, ⁓ I will use it anytime I need to vectorize literally anything. Ian Landsman (17:56) I have to ask what the heck are you this for? Why are you vectorizing an image, like a screen, you're vectorizing your screen or you're just saying like a... Aaron (18:03) So you could go to, for example, could go to ⁓ Codex Desktop, have it create with its new images too, have it create a great looking logo, and then go to Quiver and say, let's make this into an asset for real. So instead of just like one static ping, you can get the vector that you can then monkey with programmatically, scale up, scale down, whatever. So that's typically icons, logos, artwork that like ⁓ goes on the website, stuff like that. Ian Landsman (18:10) Okay. Okay. Okay. Gotcha. Mm-hmm. So use it for icons and like that kind of thing. Okay. Okay, is this a SaaS thing or this is download thing? Aaron (18:37) They have, it's a service and you can hit it via API or through their in browser tool. Ian Landsman (18:39) Mm-hmm. All right, Quiver AI, they should sponsor the show. Check them out. Aaron (18:46) Yeah, they should reach out to us, Quiber AI. Thanks for the great product. What's next? Ian Landsman (18:53) Q &A, Q &A time. Aaron (18:55) Producer Dave, what do we got? Producer Dave (18:57) Yes, sir. Hello, everybody. How are we doing? All right. So got some good questions from the chat. Let's see. Let me go up and put them on the screen here. ⁓ Real quick one from this more of a comment than a question, but we'll allow it. Now, I'm sorry. That's the wrong one. Hold on. Hold on. This one here. Max said regarding the trial, it's just the wrong take from Eon for sure. Sorry. Sorry there. ⁓ Ian Landsman (18:57) There he is. What's going on? Jack, Elon, hey, hey. ⁓ is terrible. ⁓ No way. Right take. Aaron (19:20) Yep, agreed. Ian Landsman (19:23) Elon, come on. Producer Dave (19:24) Yeah, it's OK. That's OK. Mark in the chat, let's put this up on the screen. ⁓ So Anthropic and OpenAI have these flat rate plans. Should they offer higher plans? Why or why not? Ian Landsman (19:34) Mmm. Aaron (19:36) I would love a higher plan from OpenAI, because right now I'm paying them $600 a month, but I'm having to do it somewhat nefariously and somewhat cumbersomely, which is not a word. So I don't know if you're asking, should Google have higher ones? But in general, yes, there should be higher ones. Ian Landsman (19:49) Yeah. Yeah, I think they should all do higher because everybody's just doing it this ugly way. And it's actually bad for like their statistics and stuff. Like I'd rather just know that Aaron is on the $2,000 plan. He's using a ton of tokens or he's not either way. Like, and then, then when ultimately something changes with all this, like I can just deal with Aaron. know his usage reliably and all that. So if you're going to be subsidizing the tokens, I feel like, uh, it's actually better for you to not have it. People gaming the system and split up all over in million ways. So yeah. I'm kind of surprised they haven't done the higher ones. They may not want people to become dependent on the higher tiers though, so I don't know. Producer Dave (20:33) Okay, oh, all of a sudden we've got a flood of questions here. Let's put a couple of them up there. Question here. Yeah, we can work on this. Did either of y'all catch the standup yesterday? They had Adam from OpenCode on to talk about his burnout and AI psychosis. Ian Landsman (20:36) We should probably announce it in the beginning that we're going to do this. Yeah. Yeah. No, I did not. Aaron (20:47) I did not catch that. saw his tweet, so this is adam.dev. ⁓ I saw his tweet that said he burned out super hard and had to not look at a computer for a month, ⁓ which I'm of course very sensitive to. I've been deep in the token mines since Thanksgiving and I have not yet burned out ⁓ because, I don't know, I feel like there's... Ian Landsman (21:02) You Aaron (21:10) Like one, I'm having fun and two, I feel like I'm building towards something that's working. I have hope. Where there is no hope that people perish. And so I'm still like super pumped. So no, I didn't catch it, but I've seen his tweets. Ian Landsman (21:19) Yeah. Working on OpenCode also like a whole different level. like a, it's a whole different world from what you're doing. You're building something and people are starting to buy it. It's gaining momentum. Like OpenCode has tons of momentum, but it's like a BC thing. Where's the money? It's a million people using it or whatever. It's a tremendous amount of pressure. So yeah, I could see that being a little different in general, but then sometimes just burn out even without the tokens, even without anything. yeah, hopefully he's recovered. Aaron (21:43) Sometimes you just burn out. Producer Dave (21:47) All right, we got time for one or two more. What do you think? All right. Question for Aaron. But how do you manage to use three codexes from the same machine? Ian Landsman (21:49) Yeah, yeah, no, I think we've been on pace here. Let's do more. Aaron (21:50) Let's do it. ⁓ There is ⁓ an environment variable called like codex home or something ⁓ that the codex CLI respects. And so I just have three different aliases in my terminal. One is codex dash P for personal, dash W for work, and dash S for solo. And each of those aliases just pre-pens the codex underscore home environment variable. So they all have different, whatever it is, auth dot JSON, different approvals, different MCP servers. So that part's a little bit annoying. Like if I add an MCP to one, I have to add it to all three. But I don't have to log out or log back in. I just use that codex home, whatever it is, environment variable. Producer Dave (22:49) Okay. Did we talk previously about anthropic ping SpaceX for compute? Did we talk? Because that came in. Okay. Aaron (22:55) We have not talked about that. Ian Landsman (22:57) might have last week but we should talk about it again we didn't talk about the specific numbers that came out today I believe yes yeah yeah yeah I was like 1.2 billion per month or something like that Producer Dave (23:02) Yeah. One billion per month on compute. Aaron (23:07) 1.25 billion per month. Yeah, for like a long time. ⁓ So I'm a little bit confused ⁓ with, Ian, this is gonna shock you and you're gonna not be able to relate. I'm a little bit confused with what Elon is up to. ⁓ whoa, whoa. So Elon keeps winning, I'll say that. $1.25 billion a month is better than a poke in the eye. So that's a lot of money. Producer Dave (23:12) Yeah. Ian Landsman (23:12) Yeah, for at least a year. Mm. ⁓ no, he's such a linear clear thinker. What are you talking about? Mm-hmm. Sure. ⁓ Kind of that's not exactly how that works, but sure we can I'll give it you for now ⁓ Aaron (23:37) But, well, I'll poke you in the eye and then you tell me which one you prefer. ⁓ But then he goes and releases Grok code and I'm like, wait, I thought you were out. I thought you gave up on trying to win this game. So ⁓ I think it's great for Anthropic that they get access to this huge compute pool that he will on principle never give to OpenAI. That seems like a win. But I'm also very confused about the future of Grok and why he's doing Grok code if he's like, Ian Landsman (23:48) You're done. ⁓ Mm-hmm. Right, that does seem like a win. Aaron (24:06) buying cursor and subsidizing anthropic, then what the hell are you doing your own one for? That's my take. Ian Landsman (24:12) Yeah. Yeah. I mean, think that's, uh, he had said like a couple of weeks ago, like they basically like everybody left XAI pretty much like all the main people and slash he fired them, whatever the mix is. So I think that's like a whole rebuilding project. And I imagine it's going to take time to rebuild. So presumably and nobody's using it. Right. So they have all this extra compute. And yeah, I mean, why not sell it on topic? I mean, I think that people are taking this like he's making a billion dollars, but the problem is like, how much did you spend to build data center? have to maintain like there's a lot of costs involved here that are not. factored into that, but he's certainly doing better than if it just sat there empty, unused by XAI using 0 % of a whole data center, right? Like that's a total disaster. So yeah, I mean, I think this is probably good. sure they're happy at SpaceX to be able to clock like 12 billion in revenue for the year, kind of out of nowhere. So I think Elon's still winning with this, but yeah, he's doing a good job of papering over a lot of his ill-conceived ideas. So it's working out for him. Aaron (25:08) ⁓ nonsense. Nonsense. Ian Landsman (25:10) It's working out for him. Producer Dave (25:12) He could take just a small amount of that money and he could totally throw it our way, totally fine. That's right. Ian Landsman (25:16) He could, he could sponsor the show. You know what? You want to know what I I can be had. Come on, Elon, put your money where your mouth is. Let's go. Aaron (25:21) I could be had, no problem. I don't care. Whoever's giving me a billion dollars, I'm yours. All right, what do we got next? Ian Landsman (25:26) Right. Producer Dave (25:29) We'll do one last one here. I think it's a very important question. So this could be some interesting alpha for folks out there. What do you think? Could this be the next move for you, Aaron? You're going to short the almond market? Aaron (25:34) Mmm. Ian Landsman (25:34) ⁓ I like this one. So, we gotta read it. Are you shorting the Almond Market? Yeah, put it back up. So the audio people. Are you shorting the Almond Market due to the misinformation around AI data centers using too much water? Which is the funniest, I just love this thing is going around as if it makes any sense. People just think it's like an 1840s factory that's just like using tons of water on something or 1920s. Producer Dave (25:43) here, I'll put it back up. Sorry, give me a second. Yeah, yeah, sorry, sorry, There we go. Aaron (26:03) So the only reason I would short the almond market ⁓ is that I believe the average American is a rational being. And so I will not be shorting the almond market. All I have on my side are facts and figures and truth and science. And so therefore I will not be acting on it. Absolutely not. It's all perception, baby. There's nothing real about it. Ian Landsman (26:25) I would say that's the thing. I assume, I have not looked into this, but I assume this question is coming from a place where the almonds are already very expensive and people are buying up almonds thinking, and because they use a lot of water, that this water shortage because of data centers, which is not a real thing, is going to make the cost of almonds even higher. So hence we should short the almond market right now because once people realize that water is not an issue, that the price will drop. But I don't even know how you do that. Aaron (26:51) So I read it as once the public realizes that almonds use more water than data centers, everybody's gonna revolt and they're not. Everybody, well, the average person, yeah, so no. Ian Landsman (26:57) No, I don't think it's that. There's a whole big thing there with the data centers. That's quite interesting. I don't know if we want to do it in this segment right now, but data center misinformation and like why people are fighting against data centers is a fascinating topic. Maybe we'll add a topic for next week. Yeah. Yeah. Producer Dave (27:15) That could be, I was gonna say that could be a future week. Aaron (27:19) Yep, let's add it for next week. All right. Thanks, Producer Dave. Producer Dave (27:20) All right, I think that's it. Ian Landsman (27:22) All right. Thanks, Dave. Okay. I gotta put a total on this thing in our little outro dashboard here. This is killing me. right. Yeah, so kind of related to that a little bit here is Anthropic told investors that they expect to more than double revenue here. and reach double revenue to $10.9 billion next quarter and reach operating profit. So pretty fascinating. Aaron (27:50) Operating profit, the billions don't move me anymore. I'm numb to the how many billions of dollars whoever makes. The fact that Anthropic is like, yeah, we're gonna be operating net income profitable is insane to me. That is wild. First of all, this is the thing that people said would never happen. ⁓ Now, I guess the further question is, Ian Landsman (27:58) Mm-hmm. That's huge. It's wild. Yeah. Right. Aaron (28:17) are the people using the tokens profitable? And we keep seeing ⁓ notes from companies saying, ⁓ it's actually costing us money and we're not getting value out of it. That's a separate question. But the fact that Anthropic is moving into profitability, which to me feels honestly pretty quick, like they kind of came out of nowhere and they're already profitable. That blew my mind. So bravo, bravo, Dario. Ian Landsman (28:24) Yeah. Yeah. Yeah, very quick. Yeah. Yeah, I mean, and this is all I think that they've done a very good job even somehow, even with all the crazy stuff he says sometimes like getting locked into businesses, which is just obviously where all the money is. And once you adopt Claude in your business, like just getting out of that, it's just a whole big thing. You're not going to want to move. Like you're to have to have a very strong incentive to move beyond just we could save a little bit of money or this other model is a little bit better. Like that is not going to interest any corporation in Aaron (28:49) Yes. Ian Landsman (29:09) Let's tear it all apart and rebuild on top of open AI because open GPT 6.2 is 10 % better. Like nobody can care about that. Aaron (29:16) So you know why Claude ⁓ is so sticky inside of corporations? One word, procurement. They went through procurement and now it took six months to get through legal, through finance, through accounts payable, to get everybody onboarded. And just because 5.5 in my opinion is better, they're not gonna go through procurement again. And so whatever those forward deployed engineers figured out at Anthropic, Ian Landsman (29:22) Mmm, give it to me. There you go. Right. Mm-hmm. No way. Aaron (29:46) They figured out how to sell, which turns out is kind of still important. Ian Landsman (29:50) still important, the humans, you still gotta get to the humans, you still gotta interact with the humans, you still gotta convince the humans, and humans don't like to change. Devs always get this wrong, because devs like to change all the time, everything, and they think that's how humans are, but normal humans aren't like that at all. They don't wanna change ever. They're gonna be locked in there now. And yeah, I Anthropic, you know, I I still prefer Claude and the experience there, whether or not it's a little better or worse, who knows, but... ⁓ They're there. It's good. I like it because hey, we're locked in there. making money like I'm not gonna have to worry about like, we're out of business or things are looking rough and looking for new solutions. Great. You're already profitable. Let's roll. Let's build some cool stuff. ⁓ yeah. Aaron (30:31) Still gotta go through the humans, but give us time and some more tokens and we'll eliminate those pesky humans from the stack altogether. Did the profitability news come out before they hired Carpathi with the big check? I don't know. I don't know the exact timing on that. But yes, I'm sure Carpathi is worth a couple billy and probably, I don't know. Ian Landsman (30:37) Yeah. ⁓ I don't remember the timing, yeah. Do think he wanted to check though? I'm not saying, I'm not sure. I'm sure he got paid, right? I'm sure he's getting a ton of stock, right? But like, I don't know if that's why he came out. I feel like he's a guy who wants to be in the action, right? Like if he wants to be there to figure it out. Yeah. Right. Aaron (30:58) he certainly got paid. I don't think so. He's love of the game. He's a pure player, ⁓ which I also think the DeepMind guy is a pure player. He just loves the game. And I would have to guess Carpathi wants access to compute to do his stuff. Yeah, he just wants to play. Ian Landsman (31:14) Yeah, for sure. ⁓ stuff. That's like the open clock guy, you know, he's just like, they're sharing his screenshots. I spent a billion dollars in tokens or whatever. He's like, I just get to be in here and run it all night long, all day long. got a thousand servers just running cloth stuff. It's pretty fun. ⁓ Quotes in there. ⁓ Yes. Aaron (31:36) get to spend a million dollars a day on tokens. The action is the juice. Speaking of OpenClaw, we got a new OpenClaw competitor. We got ⁓ Gemini Spark, which I'll forgive you if you have no idea what the name Gemini Spark actually means. Gemini Spark is your new personal AI agent in the Gemini app that helps you navigate your digital life, taking action on your behalf and under your direction. So OpenAI bought Peter Steinberg, which is OpenClaw guy. Claude has their Claude desktop that can take actions on your computer and now has like remote access. so Google under the Gemini brand now has Gemini Spark, which is your personal digital assistant. So we are moving out of, or in addition to the agentic. Ian Landsman (32:09) Mm-hmm. Mm-hmm Aaron (32:30) ⁓ IDEs that every company is building the exact same version of. We're all now building our take on OpenClaw. So what do we think? Ian Landsman (32:32) Right? The exact same one. Yes. So this though is interesting to me, right? Because I think, I don't know if Google is ever going to really catch up to Anthropoc and OpenAI in terms of like dev, API, or even non-dev, like just pure API. I would hate to count Google out. I wouldn't totally count them out of that, but I think that they're not in the lead right now. ⁓ And the other guys have so much money dedicated towards it that I think, you know, they should be a fairish fight. Although Google has tons of money too. But they got a lot of money. Aaron (32:54) Never. Ever. Google has so much money and Google has so much income. That's totally different than money. Ian Landsman (33:07) They do have a lot of actual revenue. But Gemini Spark to me, was like the, the element of what Google should be able to do best, which is like, Google already knows everything about me. I use the calendar, even though I'm a Mac guy, but for services, I use the calendar. use Gmail, all the stuff I work in personal, all things like give me this agent that's in a nice safe thing that Google runs. Aaron (33:17) I agree. Yep. Yep. Ian Landsman (33:31) that I can say, you know, let you do stuff like have it continuously running or be up overnight doing stuff or checking that stuff at two in the morning. It has all those claw like things where it's always on and let it be always on and doing stuff for me and trolling through my email or whatever I want to do without me setting anything up and all that. And so I feel like this is going to be quite valuable and a smart move for them long-term if they do it. Aaron (33:53) If they can execute on it, they are better poised than anyone else to do it. They've got GCP, so they've got VMs. This Gemini Spark is gonna run in VMs. They've got VMs. They've got models, whether or not the models are leading edge is an open question. But they've also got my 18 years of Gmail history. Ian Landsman (34:04) Mm-hmm Aaron (34:20) They've got all my Google Drive, they've got my Google Calendar, they've got my wife's Google Drive, they've got my Google Photos. They know literally everything that I've ever bought, because every receipt comes to my Gmail inbox. And so if anybody is poised to build the personal digital assistant, I think Google can do it. And I would never count Google out, although they continue to stumble over their own feet. ⁓ But if they can pull this off, I think this could be a winner for them. Ian Landsman (34:41) Yeah. Me too. Yeah. And the beauty of this is like, people want their open clause on like the high end model. Cause it's a bunch of devs who are like trying to hack their dishwasher and they want the coding power, but this doesn't need that. Right? Like this is like an actual personal assistant that's looking at my email, like email, like the flash is going to be fine for like looking through my email, picking out some stuff, clearing out spam. Just, Hey, set this on my account. Like that's going to be totally fine for all normal person stuff. Aaron (35:03) Yep. Ian Landsman (35:16) review this even stuff like reviews presentation and give me some feedback summarize this thing like all that stuff's going to be fine with a not super cutting edge frontier model and it's just going to be baked into all your google stuff and you're going to pay him your 20 bucks it's all going to work and then we're to run ads around it right that that's the other google thing they're going to be like oh you asked about this well here's three great options uh and the advertiser is going to pay for that so maybe you're paying nothing for it that's very powerful Aaron (35:40) because turns out they have the biggest ad network in the world and also the second biggest distribute. They've got first biggest distribution on google.com, second on youtube.com. They own the entire world's data. So if they can't make this work, that's on them. Ian Landsman (35:49) Mm-hmm. Right. Yeah, I think they'll make it work. It'll probably be free. It'll be ad supported and it'll be good because you'll have better ads than you've ever had, which is beneficial when they're done well. Yeah, exactly. Aaron (35:58) I think they'll make it work. which I'm fine with. ⁓ Okay, whispers in the wind. We've heard of mythos. We've never set eyes on the face of mythos, because no man can look upon the face of mythos and live. However, certain groups, yeah, certain groups have been given access to mythos, including a group of developers that create a popular PHP open source framework called Symfony. Ian Landsman (36:10) Mmm. Want to. That's true. at this ancient Greek he's got here. Aaron (36:33) And Claude Mythos ⁓ found 19 real vulnerabilities in Symfony. And if you go to the Symfony Twitter account, it is 19 tweets in a row of CVE disclosures. ⁓ So say what we will about the marketing that is around Mythos. ⁓ What do we think about ⁓ the actual results that we're seeing out in the real world of Mythos? Ian Landsman (36:37) Mm-hmm That's cool. I desperately want this. Why won't they give it to me? I'm very sad. I thought the most interesting bit actually is that, so I've been working on this stuff for our own tools and like done some security stuff. And there's definitely a big problem with like, it finds stuff that aren't really bugs. There's a huge problem. And so you end up having to like sort through a bunch of extra stuff, which is a bunch of extra work for the humans. And it's stressful because like, well, it says there's a big bug, so I got to really reproduce it and I can't, but am I wrong? Is it wrong? So there's a whole thing there. Aaron (37:18) Yeah. Ian Landsman (37:28) but in the article about this, goes to say that it found 19 security vulnerabilities, which is not like, I'd like that. That's a good number. It's like, like right now, if you do this, you just ask the regular Claude or codex. It's going to come out with like 200. It'll literally give you hundreds. So it's like, okay, it found 19 and all 19 findings turned out to be real with no false positives. So that is awesome. That is like what I want. That's like, here's 19 problems. They're all problems. So you should fix them and we don't have to think about it. Aaron (37:49) Insane. Insane. Ian Landsman (37:57) We just know that it's giving us good, reliable bug reports that we can just fix and be more secure. So I'm very excited. I want them with those right now. I'm very sad they won't give it to me. Yeah. That's my take. Aaron (38:09) So what do ⁓ we think about the fact that they won't give it to you? How are we feeling about that? The fact that they're keeping this one close to the vest. Ian Landsman (38:17) I know, I get it. I mean, obviously just open up to a guy my size, right? Like now that means pretty much anybody can get it, which means they're going to start hacking people with it. But I feel solvable. Like, Hey, are these people doing bad stuff with this? It's tricky. Obviously it's tricky, but I don't know. We're going to, everybody's going to need this. Literally everybody. You won't be able to stay in business without it. That's just going to be the reality. Like you will just have to pay whatever you need to pay to access these kinds of tools, whether it's mythos or the ones in the future, because you're just going be getting You're already getting attacked all the time and now you're going to be getting attacked really all the time by something much smarter than the people who were attacking you before and you're just going to need the defense and it's going to be the only way to function. ⁓ Yeah, so I don't like it, but I wish I understand it for now anyway. Aaron (38:54) Mm-hmm. I am in fact quite okay with it, to be honest with you. It seems to me like they are, ⁓ I don't know if they're inflating their own egos on exactly how powerful it is, but let's hold that question aside and say that they believe it is that powerful. ⁓ And so if we just take that as a given, they believe it is that powerful. Ian Landsman (39:08) You okay? Yeah. Mm hmm. Right. Aaron (39:32) So far we've seen them go to FFmpeg, Nginx, Curl. Now we're seeing a little bit higher up the stack into frameworks like Symfony. ⁓ It seems to me from the outside, like they are working their way up the stack of important internet infrastructure and offering, I don't know if it's paid or as a service or like a free charity thing, like, hey, we're gonna scan your critical infrastructure because the internet relies on you. Ian Landsman (39:39) Mm-hmm. I think it's free. Yeah. Aaron (40:02) I'm actually okay with that. At some point it's gonna come out and people are gonna go freaking ham on the rest of the internet. But right now having curl, Nginx, FFmpeg, Git first access, ⁓ I'm okay with that. I wish I was a big guy, but I'm okay with that. Ian Landsman (40:02) Right? Yeah. Mm-hmm. Yeah. I know I think I'm definitely okay with that because also there's just going to be things if we would just run it on my product, right? It's going to find stuff that's in those products. And then I have no recourse with that anyway. It's like, well, what the hell am I going to do about curl, you know, security vulnerability? So working up the chain, I think makes sense. Probably going to be more important than ever. You know, our, our tool of choice for most projects is Laravel, Laravel PHP. And I think Aaron (40:28) Correct. Ian Landsman (40:45) building things on frameworks, especially for web based products is like going to be required basically because like they're, they're more likely long-term, right? The mythos too is going to go to them before it goes to me also. Right. So like we want to be able to just fold in those updates, have as much be in the framework as possible, rely on the framework official, part packages and things like that. So that we just know we can get those security fixes in a timely manner. ⁓ yeah. Aaron (40:57) Correct. Ian Landsman (41:13) And I think you have that plus like things like CloudFlare and similar tools that can react very quickly quicker than you can fix bugs. There's a side tangent here, but the CloudFlare guys had an interesting post this week related to all this, which was like basically moving bug fixes out quicker isn't really a great solution because you just break things in that process. And now you're in a worse spot than you were to begin with. So. Aaron (41:33) Mm-hmm. I know. Ian Landsman (41:38) Being able to at the network level possibly defend yourself against certain types of attacks without having to update your code at least so quickly. It's probably going to be a bigger thing too. Yay. Net stock. Go get your neck. Aaron (41:48) This is a classic damned if you do damned if you don't kind of situation. I do think this will continue to lead to the centralization and unification of the stacks. Like people have said like, what's the point of a framework if AI can just slop out 50,000 lines of code, I don't even need a framework. This is the point of a framework. This is the point. You build on Rails, you build on Laravel, you build on not next.js and Ian Landsman (41:51) Yeah, exactly. Yeah. Right? Yeah. Yeah. Aaron (42:17) you have all these security things baked in because the big dogs are getting access to this first. You put it on Cloudflare because they have untold numbers of security people. So I do think this is going to lead to the further centralization of ⁓ programming tech stacks. Ian Landsman (42:37) Yeah, which is its own danger in and of itself and not my favorite part of it. ⁓ but yeah, hard to imagine how most businesses that aren't gigantic can bear this load without depending almost entirely on those bigger companies. It's just not going to be feasible. So Aaron (42:39) I'm telling you. Who can say and that is left for an exercise to the reader. Thank you. That's our show this week. Find us at tokentown.com live streaming on X and YouTube every Thursday, 2 p.m. Eastern, 1 p.m. Central. And thank you to bento at bentonow.com. If you need an email service provider, check out Bento now, tell them that we sent you and we will see you next Thursday. Ian Landsman (42:58) Yes. See ya.