Latest / Artificial Developer Intelligence / OpenAI Beats Musk, Gemini 3.5 Flash & AI Burnout Mitigation

Transcript

OpenAI Beats Musk, Gemini 3.5 Flash & AI Burnout Mitigation · Artificial Developer Intelligence

• Dan (00:00) and Microsoft announced this week that they're actually canceling Claude Code subscriptions for a lot of their folks because their costs had actually exceeded the cost of human developers. Shimin (00:11) Right, but what do you expect if you're running LLM leaderboards? Hello and welcome back to Artificial Developer Intelligence, a weekly conversation show about AI and software development. We bring you the most insightful posts and our hot takes each week by going through hundreds of links and dozens of newsletters. This way you can stay on top of things and still have a life. Because we do not have lives. My name is Shimin Zhang, and with me today is my co host, Dan, his Butt pillow collection is only second to his tiny computer collection. Lasky. Dan (00:46) I wish that were true. My chair's actually kinda uncomfortable if I'm being honest. Shimin (00:47) Yeah. It seem i yeah, it seems like a real genius move there. I wish I had a butt pillow. how was your long weekend, Dan? Dan (00:56) So all right. yeah, very, very rainy in the Midwest. So just getting hammered all weekend. So it is what it is. Stayed inside, did a lot of compute. I also wanna say for the record, I think my middle name this week should be ⁓ Dan. I missed I miss being called a study buddy. Shimin (01:05) Nice. ⁓ Yes, we can move back to study, buddy. Dan (01:16) It's I don't know what happened to study buddies but I thought it was good. So anyway. Shimin (01:20) ⁓ yeah, my my weekend was good. I made ⁓ lobster rolls. They were delicious. ⁓ Very expensive lobster rolls, but they were delicious. Alright. On to our show. Ugh Maybe. We'll find out. Dan (01:28) All right. Yes. Was it worth it? Yeah. Shimin (01:35) On this week's show, we're going to start, as always, with ⁓ the news, the news threatmail, where we're gonna talk about the resolution of the Sam Altman ⁓ OpenAI Elon Musk case, as well as the Gemini three point five Flash, the latest frontier model from Google. Dan (01:52) Then we're gonna have a rare dip through the hardware hut where we ⁓ answer the question of is my the Royal My ⁓ forty eight thousand dollar GPU server worth it? Shimin (02:00) Mm-hmm. Yeah. ⁓ then on to technique corner where ⁓ AI burnout has been in the news a lot this week. So we're gonna talk about some mitigation strategies. Dan (02:12) Post processing we'll talk about human bottlenecks, which I'm sure we're all feeling. Shimin (02:17) then Dan is gonna rant about this thing called a slop grenade. I still don't know what it is, but we'll find out. Dan (02:24) Yes, we will. And then finally on two minutes, we'll go through a couple of articles about ⁓ how the bubble's doing. Shimin (02:30) Yeah. All right. Okay, so this week we got the verdict in the open AI ⁓ Elon Musk trial. OpenAI won. I think on the technicality of statue of limitations. So Dan you brought us this ⁓ New Yorker article, very long form New Yorker article. Dan (02:46) Yeah, it is. And I also think I'm pretty sure it's the first time we've ever covered a New Yorker post on the podcast as well. And that was partially the reason why I was interested in it because they have a very sort of distinct style of writing. But also, ⁓ it's kind of wild to have something like this wind up on the New Yorker, right? That's pretty mainstream. And here we are. Like this is mainstream news these days. Shimin (02:55) I think so. Mm-hmm. Dan (03:11) But yeah, so extremely long article with a lot of meandering. But the part that I found interesting, aside from just sort of like the facts of the case, which I feel like have probably been discussed ad nauseum, but you ⁓ did summarize it pretty nicely there. ⁓ with yes, Sam Altman won is that really what they're looking at is We had one billionaire fighting another billionaire about who's gonna control essentially a for profit AI company, right? That was being formed. And one of the things that at least I took away from the trial was that it very much seemed like Elon was kind of angry because he didn't think of it first. You know. and that begs the question is like, should these individuals be responsible for AI safety, right? Because it's really just coming down to the, you know, truly like personalities over like process or anything else there. And then that also kind of begs the second question, which is that Shimin (04:00) Right. Dan (04:10) Because of the unique situation in this open AI nonprofit to for profit split, is there a structural ⁓ sort of like quest for profits. And this sort of gets into like Rahul's book a little bit, right? Where he was talking about the deep mind folks and how like the the quest for profit pushed them in certain directions that ⁓ is sort of inevitable because of just the way this is all being commercialized. Or is there something that could have been avoided in in this structure where we could have maybe actually stuck with something for the betterment of humanity, you know? Shimin (04:36) Mm-hmm. Dan (04:45) ⁓ so I just thought it was interesting to look at that trial through that lens in addition to the, you know, sort of more detailed one. Shimin (04:51) Right. ⁓ you know, the article talks about how a lot of the workers at OpenAI as well as some of the participants of the trial were very downbeat or kind of ambiguous about how much they were making out of this whole deal. I think one employee of OpenAI couldn't answer the question of whether the stocks he sold was more or less than twenty million dollars, right? They're all trying to play out this like very high minded I'm in it for the values and the benefit of humanity and ⁓ we're here to prevent artificial ⁓ intelligence from taking over the world and killing all humans. but the structure that they're operating in ⁓ forces them to really have to be a kind of a for profit organization first and foremost. And I think A similar article came out this week from Anthropic that basically stated, you know, they wanted to be a frontier lab and work on safety, but because of market forces, they also have to now be a frontier lab that also makes money and how that puts them in a bind. Now, this all could just be marketing speak. ⁓ and I think action does speak louder. for example, in this article. Elon Musk's attorney, Stephen Molo, argued we could all die. We could all die as a result of artificial intelligence. And then the judge observed that Musk might have endorsed this argument more sincerely if he didn't fund XAI as a competitor. Dan (06:10) Ha ha Yeah. That was that was a great pull quote. ⁓ and I guess the other one that we should talk about if we're gonna really do the article justice is ⁓ what is the butt pillow thing and and what did it mean? so yeah, apparently the actual courtroom had very hard Shimin (06:19) Yeah, I really like that. Dan (06:35) seating surfaces. So open AI's lawyers and executives all brought their own seat cushions. so it's sort of a metaphor that he uses throughout the article and it's I would argue towards a sort of collective complacency. But it was pretty funny regardless to how butt cushions keep coming up, butt pillow. Shimin (06:54) Yeah. And Gideon the reporter also got one, 'cause it just seemed really comfortable. And now everybody wants one. Maybe maybe that's the one thing we'll get out of this whole deal is everybody will get their own collection of butt pillows. the the last thing I wanna mention is some of the other character pieces that came out of this trial. ⁓ f one quote is Sam Altman didn't exactly come off well. Though he did get into some dryly amusing lines about having to endure Musk's desire to show everyone memes on his phone. So the richest man in the world is just like you and me. We just like to bombard our friends with memes over their objections. Dan (07:33) Payment. Meming ain't easy. Shimin (07:35) Yeah. Okay, now that that is all settled, let's move on to some more model news. last week Google I.O. happened, and as a part of Google I.O. Alphabet ⁓ introduced Gemini 3.5 Flash, their latest flagship model. they're only releasing the Flash model as of today. ⁓ they're going to release the ⁓ extended regular baseline model. probably in their next release cycle. And they are for a frontier model, it is interesting that they're immediately releasing it as a part of the AI overview, as a part of Google search. So it looks like Flash is fast and cheap enough to run that they can just give it out for free for billions of web searches. ⁓ Dan (08:20) I've always wondered this, but it wouldn't surprise me if Flash effectively means it runs on a TPU. Which is probably why it's cheaper. But that's just like my hypothesis, man. Shimin (08:30) have ⁓ I had a chance to ⁓ play with Flash yet? Other than ⁓ looking at the AI overview when you Google things. Dan (08:31) Ha ha ha. I have not other than the the overview, although admittedly I don't even use that too much 'cause I tend to use Kagi a lot for search more than Google these days. But I I actually let my Gemini subscription lapse because I haven't been using it enough. so maybe I'll have to check it out again. Shimin (08:57) Right. ⁓ the funny I guess the funny thing that happened as a part of the Flash three point five release was for a couple of hours, at least dozens of minutes out there, last week you couldn't search the word disregard. Dan (09:10) I did see that. Yeah. Or a couple of other ones that were basically like cancel or abort or something like that. It was like it would just take it as a literal command. It's kind of funny. Shimin (09:16) Yeah. Yeah, it's pretty funny. ⁓ growing pains. I Dan (09:25) Yeah. But really I wouldn't put that on the model. That's that's like strictly some harness engineering failure right there, you know. Shimin (09:32) Yeah, absolutely. They're d they probably should have included a few more ⁓ regex terms in there. I did give a run it through my usual what should I do with two and a half acres in Washington ⁓ benchmark, my very unscientific one. was using Deep Search. The model is quite good. I will say it did more analysis than the Gemini. 3.1, which was more of a let's just summarize everything. 3.5 did a little more analysis and it used a lot more LaTeX equations. So it it feels more mathy to me than than the 3.1. But as when it comes to the results themselves, it gave me less diverse results. So instead of giving like the the usual 20, 25 different options, it gave me like half a dozen. they were well summarized and well analyzed half a dozen, but was still only half a dozen. It was very fast and I was really appreciative of that. What else? and it did it did well when I ⁓ asked it to elaborate on some of its answers and choice and have follow up questions. So overall, definitely a more thinking version of three one feels good. Still quite ⁓ distance between the three five, at least flash, compared to say Opus or GPT five five as to be expected. This is a flash version. and of course it drew a pelican. And I like Simon Willison's description that this pelican or somewhere from Hacker News description that this pelican looks like it in Miami for a crypto conference. It does. It does. Dan (11:02) Yeah. Or like a dark wave music festival. It's got that like slashed sunset from the eighties. And yeah, it's pretty great. Synthwave, I guess. Whatever. Yeah. Definitely has one of those Bluetooth speakers on the bike blasting synthwave at the top of its Shimin (11:09) Yeah. Yeah, definitely yeah. Very cool. Potentially those drugs. Like yeah. So Yes. ⁓ I can see it. other than that, it it looks pretty good. The the body, the frame seems a little twisted, but a lot of interesting details, so give it a shot. Dan (11:34) yeah, that is true. For those of you listening at home, the sort of top piece of the frame, particularly I think, because of the miter they've used on the S V G, like or the border intersects in a really funky way that almost makes like an A. Shimin (11:47) Yeah, it crosses. ⁓ so you know, to put a button on it, Google is not out of the AI race by a long shot. this is a frontier model that they have the spare compute to give to everyone. So TP Dan (12:02) And as someone that's been running dwarf star four a lot this week, on my local, I can tell you there's definitely something to be said for generation speed. 'Cause I don't got it. Shimin (12:12) ⁓ Wasn't there a how slow ten tokens a second actually is post that was making around around the around the interwebs? what is like a usable token speed for you? Dan (12:27) You know, 10 is usable, but it's it definitely causes me to think more about how or what I'm about to ask versus just sort of yoloing with some of these things. So in a way it's actually kind of like a good restriction on me because it forces me to prompt a little bit better and like think about what the output of the prompt is likely to be. But it's fine for like shorter stuff, it's fine to be you can actually be conversational with it, you know. But if you're gonna be like, okay, now give this to a coding agent and like crank out a lot of code, be prepared for it to take a good four hours, something it would take Claude like five minutes, you know. At least on my hardware. But Shimin (12:49) Right. Right. Yeah, I think we're gonna we're gonna come back to this later, but knowing when to use AI for a particular task or how to shape that task could become a meta AI skill in the coming month. Alright. On to hardware hut. Dan (13:20) Yeah, speaking of running, well, this isn't really about running models locally, but so how do I even say this? Ro Rosemine, I guess, has a a blog, rosemine.ai. And ⁓ they apparently left their Fang job ⁓ in twenty twenty four to become a a researcher. I assume the sort of subtext is like ML or AI research. Shimin (13:42) Mm-hmm. Dan (13:42) ⁓ so in order to do that, they needed GPUs. And they built a custom six by six thousand Ada GPU server. and so they went through a whole bunch of really interesting stuff. It's worth reading. ⁓ I won't go into vast detail about like why they chose them, but it wound up coming down to price to throughput ratios of the Ada versus an H100 versus an A100. and so that overall came out on top. the other kind of funny thing they talk about is how do you run? Because when you think about it, like these are pretty beastly GPUs, right? Like, I mean, one of the things I really actually appreciate about my framework desktop is like I can run it off a in fact, I have several things running off that same wall socket and it's fine, even under full load. But if you're running These beastly GPUs, you can't actually run them off of either a single power supply, or if you do run it off a single power supply, it needs to have pretty b I think it needs to be a 40 amp line or something like that. It's pretty beastly power constraints. So they apparently hired like a professional builder or something to sort out the power supply thing for them because they're worried, significantly worried about setting their ⁓ Shimin (14:42) Wow. Dan (14:53) House on fire. but the real reason why I thought this was interesting is like is there's sort of like this question I think in a lot of people's minds like, why would you use your own GPUs when there's so many like good cloud things, right? Where you can just rent them if you're doing, you know, experimental projects like this, or just talk to any of the frontier providers if you're just doing inference stuff. So why would you want this in your house sucking up power? And Shimin (15:06) All right. Dan (15:18) The answer from this really is that at the end of the day, at least for the type of experiments that they are running, which were also kind of interesting and maybe worth talking about, they came out ahead doing this. So not only did they think they said it was something like 70, 76% utilization, I guess, for the the vast yeah, for the vast majority of their runs. it became Shimin (15:27) Mm. Yeah, seventy five to eighty five. Yeah. Dan (15:39) It they hit break even. and then everything past the break-even point, they're basically just paying cost of electricity, which is admittedly pretty high. They're saying $125 a month to run the thing all the time. so yeah, I mean it's like I think for very specific use cases, it it can definitely be worth it to invest that. ⁓ But one other thing they didn't talk about, but you see in some of the graphs on the post are they they had a utilization graph where they're showing like how much they're having it crank. So There's these dips in the graph. So it's pretty much like looks almost like a kind of lossy router graph or something like that, where I where you're hitting p periods of packet loss. And it each one of those dip periods of dips, something broke. There's only one where it was human, read a lot of papers. The rest of them were server broke, only three GPUs available. ⁓ another big dip, riser failed. Another big dip, riser failed, which I guess is like the PCI risers to connect the Shimin (16:16) Mm-hmm. Yeah. Dan (16:36) the cards to the motherboard. so there's definitely some some caution there. Or, you know, be the type of person that's just like willing to hack on stuff like that and you don't care. ⁓ Once they hit that break even point, it's essentially quite cost. Shimin (16:51) Right, that's assuming like a almost like eighty five percent utilization rate, which which like is very high for the average developer who isn't doing these reinforcement learning experiments, like the author of this blog post is. I think a lot of folks are you know, for for your average developer, like you know, for me in particular, the trade off is like do I Dan (16:52) It's really Utilization. Yeah. Shimin (17:13) buy a beefy machine. Like I'm considering buying the latest version of ⁓ either ⁓ Mac Pro or the new MacBook with a hundred and twenty eight gigs of unified memory so I can run some of the latest ⁓ open weight models locally or should I just keep on, you know, throwing money at Anthropic for a better model that doesn't require such a high upfront cost. Like it's it's a long time to pay back five grand in Dan (17:34) Yeah. Yeah. Shimin (17:40) ⁓ local model inference costs. Dan (17:41) Yeah, even at the hundred dollar tier, you're still that's a couple of years, right? Shimin (17:46) Even at two hundred dollar tier, it's two years and like who knows what model landscape will look like in two years. So unless you're actively toying with weights or training it, I I find it a little harder justify. Dan (17:59) Well, the other thing to consider and part of the reason why I bought this in the first place is like like data privacy matters and it's like you know, certainly I have varying levels of trust for the different frontier labs, but like one of the projects that I'm interested in doing is like I keep a pretty consistent journal, right? Like I write all the time. Would I trust any of the any of the frontier providers with that? No. You know. And Shimin (18:06) Yes. Mm-hmm. we're gonna come back to this. Right. Yeah. Dan (18:25) yet I still wanna like mess around with it. So like that to me that's a pretty good use case for this kind of stuff, but or like local models. But I also just on a personal note, why would you pick a Mac versus something like a digits or a a three ninety five Max machine? Shimin (18:41) 'cause I wanted to I needed to upgrade my laptop really. I'm not gonna get I'm not gonna get another desktop. My laptop is on M1, you know, from like two thousand and twenty one. So it's not quite a end of life, but yeah. Yeah. So I might as well. If I can run DS four, it makes sense. ⁓ Dan (18:44) Anyway, okay. Yeah. It's time. Yeah, that makes sense. Yeah. Although if I can deviate for a moment, the rumor is touch screen Macs are coming out with O L E D in the next the next version, but they've been saying that for a while, so Shimin (19:06) Yeah. I do not need a test touch screen Mac. My screen is dirty enough as it is. ⁓ it's just gonna have like hot sauce and fried chicken fingers on it. No thank you. so I think the answer to the question was my forty eight K GPU server worth it, and is it gonna be worth it for you? ⁓ as always, it depends. ⁓ I should let you say ya. Dan (19:11) Ha ha ha ha. Yeah. That's what I was g no, I was gonna say it. Yeah, it's all right, you said it. It depends. Shimin (19:29) Yes, it depends. Alright, a true a true senior engineer moment. Let us go to the technique corner where I actually have two different posts on the same topic about AI assisted engineering and the burnout that a lot of devs are suffering right now. So this first one is from Evo Martians, which is a I want to say high end premiere software consultant company and this is from a their blog. first the problem of course is that a lot of devs are suffering some burnout. and it is because of a lot of things that we have already covered, you know, the context switching, the feeling like you're not necessarily being ⁓ productive because you are no longer heads down in the code base all the time. you are forced to create more work. ⁓ because how fast it can feel like the coding is getting done. And some devs are even feeling like they need a career change. Now well first off, Dan, do you feel this way? Dan (20:30) Sometimes it's my honest answer. I mean, there's definitely been like I I don't know. Like a lot of this is week to me week to me. And I that's part of why I do the podcast, because I think it's gonna be very funny slash insightful at some point to go back and look at this and be like, Okay, on week one I was like, this is great. On week two, I was like, This is terrible. On week three, I'm like, Yeah, it's okay. You know? And I feel like the same thing goes with this. So I wouldn't say I'm like in a state of burnout around it, but like I do feel pressure and sometimes like mismatched incentives, right? And I think things like leaderboards and stuff like that can kind of do that, where it almost feels like the incentive is to push as many tokens as you can. versus like my incentive really is to like build stuff that brings value to people, you know. That's that's why I do what I do ultimately. And do I and I like I've mentioned this before, I'm on the middle of the spectrum, right, in terms of like enjoying the craft and versus shipping value. And to me that's a sliding scale. And I think maybe that's sort of a I won't say unique opinion, but definitely I see more one side or the other in in blog posts, possibly because makes a better headline, but Shimin (21:41) No, it d being a 10X engineer does make a better headline, right? Like that's why that's the survivorship bias there. Yeah. I've been feeling this way. I've been I've been feeling a little bit burnt out. Like there are days when I'm working on, you know, seven concurrent Claude code sessions after work. And like a lot of them are long running, a of them are doing convergence work or iterative work, but you still have to tab. Like you just have to tap through all of them and see what's going on. And it it feels like yeah, I have a lot of things happening. but the context switch gets heavy ⁓ at night. So well the according to the evil Martians, you should fix or ⁓ mitigate your burnout symptoms by the following. First, an acknowledge your wins 'cause you're still in control. Dan (22:13) Mm-hmm. Shimin (22:26) It shouldn't feel like we're just overseeing our AI. Like we still do have a hand in shaping the output distribution. second is rethink your AI workflow. work with them in a way that reduces your contact switching costs. So maybe a little more planning, and a little less review and then keeping your parallel work streams under a limit. I think we spoke about like three, four agent, that's usually the top and don't try and do more. next is ⁓ keep exercising your craft. So don't always just let AI do other things. You know, still put hands on keyboard every now and then. ⁓ because you you are gonna need that discernment and that experience to come up with good judgment. ⁓ and then have Better and more disciplined work life balance. yeah, that's always nice to have. And and lastly, find new areas of interest ⁓ outside of your work. all good Dan (23:20) Which are also not just more vibe coding. Excuse me, agentic engineering. Shimin (23:23) Hmm. Yeah. ⁓ any of these ⁓ speak to you? Dan (23:29) I I I guess. I mean, I think acknowledge your wins is a big one, but I think pure acknowledgement doesn't replace the sort of like dopamine hit that people were getting from shipping directly. and I think that still comes down to like sort of the manager thing we've talked about before, right? Is like now ⁓ almost every engineer is grappling with like what it means to be a delegation. manager, right? but the other thing I feel like is missing from this is that they they don't talk about value mismatch 'cause I think that's a lar usually a a large chunk of what causes burnout, in my opinion. Right. Like your your your company or your job or like the industry writ large is valuing a certain side of things and you might have like and that thing is not handcrafting the code. Right. And you either still have that value for handcrafting the code or are mourning it in some capacity. And I think that's a recipe for burnout where like Shimin (24:01) It elaborate. All right. Yeah. Dan (24:25) It's it's hard to until you can kind of like get through that one way or another, like you're gonna have a bad time. Regardless of the other things. and you know, it's like, you know, the old ways of burnout are not dissimilar, right? If you work in a place that like only cares about Jira tickets and nothing else, like that's can be soul crushing for certain types of engineers that care more about like value, you know, to customers. Shimin (24:32) Yeah, that's fair. Yeah. Or even worse, having unrealistic deadlines that you know from the start is never gonna be met. And there's no way of actually getting there. So the only answer is to work harder at a deadline that was never gonna get met. The death march, as I call them. yeah, we actually have ⁓ yet another AI fatigue and burnout article this week. I I actually came across like four this week in my readings. So Dan (25:01) Yeah. Well. Shimin (25:13) It's p it's been a burnout heavy couple of weeks, I suppose. this one's actually from February though. It's from Siddhant Khara Excuse I'm sorry if I'm butturing your name. and he's been one of the ⁓ early earliest adapters of AI. he has a agentec engineering guide, where he ⁓ compiles some of the lessons fff when it comes to using AI ⁓ in software development. And ⁓ he talked about Kind of a similar issue, right? Like there is a change in our development workflow. We go from writing code to spending most of our time ⁓ reviewing code. But even more so, he talks about this fundamental mismatch between the way that developers like to think. We like to think in deterministic, repeatable inputs in equals outputs out. And having AI models that constantly change, especially as things get upgraded, the system prompts change. You are forced to deal with a lot of uncertainty. ⁓ and that causes a lot of anxiety, but also ⁓ just a lot of churn. Like every time a new model come up comes out, like maybe your handcrafted prompt is no longer going to work, and you have to rethink everything. Sometimes the the labs tell you you have to rethink things, like OpenAI did with GPT-5.5. So I definitely agree with a lot of these issues. and you know, he also warns about the AI the thinking atrophy that we also talked about a lot. You you're afraid that you can no longer work on really the hard problem because you haven't spent the time thinking about the hard problem thing a little bit. You're mostly in review land. So if you have to whiteboard something, you might kind of get freaked out. And what he Dan (26:37) Mm-hmm. Well, you know, if you run a really slow LLM It helps with that a little bit. Shimin (26:55) That's exactly what I was gonna say. This comes back. And the suggestions ⁓ he gave that worked in his experience include time boxing AI sessions, so not let an agent just spin ⁓ on and on. something that I'm very bad at, so I will do a better job of time boxing my time with AI. separating AI time from thinking time, box time without the distraction and the context switch. So you can still do deep think. ⁓ super valuable. accepting 70% from AI and hand code the rest. This one's interesting. 'cause Not reviewing everything AI produces is also a part of the suggestion. So there's a natural conflict there. But mostly you know, being strategic about the hype cycle, don't always go try out the first tool because you're always gonna get FOMO from Twitter and Hacker News, right? ⁓ take a little time, wait for things to shake out, and focus on the fundamental problem things, such as you know, like learn about Coding harnesses and the problems that coding harnesses solve and the maybe the major techniques that coding harnesses use to improve performance, but not necessarily try every single coding harness out there. I think that's super valuable. Dan (28:04) Unless you're crazy like us. Shimin (28:05) even we don't try all of Dan (28:06) That's true. I try to try all of but yeah, there is. Shimin (28:09) Yeah. I try to give every single major one like honest shake, but it's n it's impossible to try all of them. but we also don't have lives. So yes, for that matter. I think we are going to I I think if if nothing else, you know having all this AI burnout chat chatter ⁓ means that Dan (28:17) Or infinite budgets for the podcast. Shimin (28:27) The industry is probably on the cusp of rethinking our relationship with AI in some way. Like we're realizing that just tackling this tool on our old workflow might not just get the job done and we need to we need to set some new boundaries and new patterns. Dan (28:42) And and Microsoft announced this week that they're actually canceling Claude Code subscriptions for a lot of their folks because their costs had actually exceeded the cost of human developers. Shimin (28:53) Right, but what do you expect if you're running LLM leaderboards? Dan (28:57) True. And but also like they are in a fun position where they have, you know, their own GitHub products and then they were also buying Claude code licenses for people. So I could see how costs would exceed there too. Yeah, for sure. But Shimin (29:10) Yeah, so we're definitely seeing some ⁓ pushback. I I think I'm tryna trying to put myself in the shoes of like a CTO or CEO, right? Like you were given a demo where this magical box can do this amaz seemingly amazing tasks. Because you don't actually have the necessary judgment expertise to to know what's good and what's bad for what is being generated as a part of these demos. And then And and clearly this is where the industry is going, so you then force not force, encourage it upon the entire entire employee base. And now you're realizing that, hey, maybe we need to encourage smarter and not just measure things based on token count. So makes sense. The pendulum's starting to swing back. Dan (29:50) Yeah. Which I feel like that's a yeah, and that's I feel like that's a pretty normal reaction to any new technology, right? It's like for better or for worse, folks are like super all in on it. And you know, like look at blockchain, for example, right? There was like that period of time where people were replacing every possible usage of a database with blockchain for for like six months and then that sort of calmed down and now we've like sort of have steady state like the Shimin (30:10) Ha yes, I remember that. Dan (30:17) three or four use cases that sort of make sense, you know. And I feel like we'll probably see something like that with this too. Shimin (30:20) Yep. ⁓ Maybe in four months we are all gonna look back at this time deeply nostalgically as like, remember when the company like gave us all the tokens we wanted and want us to spend more tokens? Now our token usage is is like a part of our ⁓ evaluation. If you use too many tokens, it's coming off your paycheck. If only. Yeah. All right. and I actually have a related post processing post, to our AI burnout question. ⁓ Dan (30:31) Ha ha ha. Yeah. Shimin (30:49) This week. It's titled Human Bottlenecks from Boretti.me. It's it's really interesting. It got me it got me thinking, and this is relevant to our personal journal news from earlier. The basic question is there was a tweet going around saying your laptop has ⁓ it has a a hundred million USD startup in it. You just need to figure out the right sequence of words to get it out. Right? So and and this is true in my experience. Like we know this general intelligence is so powerful that we can probably do something really great from it. but we're actually not gonna see that is the author's opinion. And the reason we're not gonna actually see everybody going out there creating a hundred million dollar AI startup is because of two main reasons. One, ⁓ the serious context of use. So the number of people who build who use their AI to help them build a flashcard app when they do not use flashcards in their daily lives or even use a the Anki app, you know, is extremely high, right? So if you don't actually have an important problem for the AI to solve. Dan (31:49) Ha ha ha. Shimin (31:57) Then what good is having this powerful tool? It's similar to the old the internet gives you all of human knowledge that's just one click away. But most people spend that time looking at cat and dog videos instead. And here is really a really ⁓ it's coming for it's coming for us, Dan Finally, here's the tools for thought slash note taking people. God save us. It's always the same thing. Your folder with notes, pardon me, your second brain, plus an AI agent that writes, edits, synthesizes information, answers queries. You could build this in an afternoon, and it won't move the needle in your life. For the same reason that building the second brain in the first place didn't make a difference. Dan (32:29) Ha ha ha. See, I actually don't agree with that one. I am the no taking people and I think having the second brain made a big difference in my life, but but I also probably didn't take the same approach that everyone takes. Shimin (32:54) Is is and I'm quoting here, is the deliverable you taking a screenshot of your obsidian graph and tweet about it to show to show off how much it looks like a incomprehensible ball of twine? Dan (33:06) No. I also would never let the agent write things in it. For better or for worse. I think it's interesting for synthesis. And I also ran very briefly. Sorry, this is like such a tangent. But but maybe we have some overlap in listeners with the obsidian nerds. And I did run an interesting plugin that basically builds a vector Shimin (33:06) Good. Yeah, I take a different approach. as as as we spoke about. Dan (33:27) Database of all your markdowns. And then it will surface links for you in like the side panel based on the semantic similarity to what you have up in your current tab. And I found that to be utterly useless. Like I thought it was gonna be super cool and it was gonna bring up all this stuff, and it's like instead is just like Shimin (33:42) Ha ha. Dan (33:49) We w just had provided basically no value to me whatsoever. So, yeah, there's that. Shimin (33:54) I ⁓ again, also also a big note taker. also a big fan of Obsidian. I have never used that stupid graph feature. A, I don't tag my shit like that. And B, what the hell is that graph gonna tell me that I can't know just using the search function? I never understood it. ⁓ so I'll be the first to say. that being said, second brain is definitely useful. I I also will push back very hard on this. Is Dan (34:05) Yeah. Shimin (34:19) the ability to look at trend lines and s ⁓ spines of ideas in my notes at a cost that is almost zero, is extremely powerful. And I think, you know, time will tell, but personally I it's been it's been hugely helpful for me, especially as the models get better and they're able to kind of see things from different perspectives. And we're chatting off the mic about this, but I've been doing some personal coaching stuff with my Pi agent and it it really gave me insights that I didn't know about myself. whether those are real insights I don't know. It it feel it feels like real insights. ⁓ so so it might be ⁓ psychosis adjacent, but only time will tell. And the second limiting factor to why aging would not make you a super bionic ⁓ augmented human is that the limitation has never been the tools. the limitation has always been your yourself, your Dan (34:53) Only time will tell. Yeah. Shimin (35:16) energy level, your intelligence, your motivation, your executive function, maybe just time, right? and just because you have an age an agentic model doesn't mean you would automatically know the right questions to ask. You need to have a certain level of both motivation and knowledge to actually get the most out of it. And I'm hoping that this is true. Cause that makes us relatively competitive. I am skeptical that there because AI is such a general piece of intelligence, I'm I kinda think that it's able to coax that kind of intelligence out of most people, at least to some extent. So ⁓ or insight or just like a good workflow can patch up a lot of variants in the underlying Dan (35:42) Uh-huh. Shimin (36:02) population, I wanna say. And I'm and I'm hoping that this is true. Because if it is true, then it truly allows f people to gather knowledge much more quickly so they can make better judgments and ask better prompts. And I'm I'm betting on and I'm hoping that this this is actu this internal limiting factor ⁓ hypothesis from a blog post is also wrong. Dan (36:24) I was walking my dogs the other day. This is related, I promise. and I I stopped briefly because there was like a little crowd of people and there was this little girl who wanted to see the dogs and ⁓ I have two cute little dogs or whatever. And it happens a lot. And she just wanted to pat them, whatever. And then as I'm walking away, she shows me this phone that she had in her hand the whole time and goes, Shimin (36:35) Mm-hmm. Dan (36:46) It's ChatGPT and then just starts giggling at me. And I was like, I just had this like kind of like, you know, what is that galaxy brain moment? Where I was like, what the hell is this like? I don't know. She's probably like six or seven years old girl talking to Chat GPT about something. They were clearly having like a really, you know, long conversation. And I'm like, I wonder. Shimin (36:59) Right. Dan (37:08) It's one thing to look at it through sort of the lens that we have, but it's gonna be very interesting to see what her lens is when she's, you know, our age. Shimin (37:16) Yeah, I think about that all the time. the first AI native generation of folks. It's kinda like how I know I'm too late for like TikTok and WhatsApp and not WhatsApp. It's the other one. Snapchat and like you know, COVID wasn't a fundamental part of my upbringing. ⁓ Dan (37:32) But I I the will I will say the one thing that really surprised me was when when they had the quote unquote digital native, right? Digital native didn't mean competency in software development. I would argue that may have actually even slightly gone down during that time period. And the the reason for it, I think in at least in my mind, was that like people grew up with like social networking as a refrigerator, right? Shimin (37:47) Yep. Mm-hmm. Dan (37:59) You know what I mean? It was just always there their whole lives. Whereas it's like I think you and I were in the that sort of cusp era where we grew up without it but then like had it, you know, later on. And Shimin (38:09) Yep. Yeah. Dan (38:11) And so like I don't know, something about that I think stoked this curiosity in me or just like a platform level curiosity that I had that I think that people like where if you know, like Facebook, for example, was always there, you might not sit there and think about how does Facebook work at a you know crazy scale or something like that. You'd just be like, it's there and it's you know, I mean, I don't spend a lot of time thinking about how my refrigerator works. That's always been there in my entire life, you know. Shimin (38:26) Right. Refrigeration is amazing. so if we have two different pole on the one end of the pole are people growing up with AI will become that much more knowledgeable, will become that much more effective because they all have a personal tutor in their pockets from like the age of three. And on the other pole is this hypothesis that people are gonna not know. How to do critical thinking because they can just ask the AI to do the critical thinking for them. The answer is probably some combination of the two. Dan (39:08) Hopefully. The part that scares me personally is that I feel the other one happening to myself and that's why I'm just like, oof, I don't like it. When I I catch myself advocating either intelligence or decision making to an LLM, it scares me a lot. Like, don't get me wrong, I'm pro LMs in general, but like I'm more cautious about that and what that means for me either professionally or personally, you know. Shimin (39:36) That's interesting. Okay, so so here's a thought. What if ⁓ you think about how ⁓ forty percent of America used to all be farmers and doing manual labor every day? Right, and then we have machinery and people basically stop doing manual labor. but now we go to the gym and we socialize with our friends to make sure our bodies Yes. In like a rigorous structured way. Dan (39:55) Do manual labor again, yeah. We pay money to do manual labor, yeah. Shimin (40:01) So maybe one day we're gonna go back to some version of Let's All like the the Enlightenment cafes where people just like sit together without their phones, have stimulating drinks, and just have like judgment discussions to make sure their judgments are still sharp. I love this future Dan (40:18) I think your privilege is showing a little bit there too. There's plenty of people that still do very manual labor for a living. Shimin (40:20) ⁓ Great, that's that is fair. That is fair. Well, I guess that was my soapbox. that was my random Shimin's soapbox that we just snuck in there. So let's move on to my favorite segment. Tell us, what is slop grenade Dan (40:33) It's a good one. a plan a plans, a planned totebox. Yeah. So I came. and this is making the rounds this week, so this is combination rant slash article. It's not really an article, but ⁓ a website, no slopgrenade.com. and w one of the folks that spoke about it ⁓ was saying that this is like pro perhaps the new generation of Nohello dot com, which was the if you're not familiar with that, it was the like the the age old thing where someone sends you a a DM and is like, Hi Shimin Shimin (41:06) Mm. Right. Yes. Hi. Well actually I'm gonna leave you on read Sorry. Dan (41:08) And then they don't say anything. ⁓ Yeah. Right. And then you have no idea what they want, right? So then you eventually say hi. And I'm like, yes. so I have a question about that service that you wrote three and a half years ago. and why did you choose the colour? Shimin (41:20) Right. Well, you know, interesting you asked that. ⁓ orange is a royal color of the Netherlands. Dan (41:21) Which you we could have just led with. Yeah. So the modern version of that is I mean and Shimin are having a conversation and he asks me a question. I don't know. Shimin (41:34) Dan, why did you cheer that microservice architecture four months ago? Dan (41:37) Right. And instead of answering that question, what I do is copy his question into the code base with Claude running and then paste the output of Claude directly to Shimin ⁓ so what they are terming that is a slop grenade. You've been hit by a slop grenade. Which I just find the whole concept of that to be funny. And yes, I have unfortunately seen this in practice. so what they gently suggest is at least take the time to summarize the output. Shimin (41:51) Mm. Yeah. Yeah. Dan (42:08) And and I think what I've been finding myself doing is like I definitely do that sometimes. Like I will ask Claude for at least a pointer or something on, you know, like what the question is if I don't have the full context. But I usually will also reference, like, you know, I sent this to an LLM. ⁓ this is what I think based on the output that I got from that, you know, more i in in the subtext there, at least in my opinion, is like. Shimin (42:23) Mm-hmm. Dan (42:32) Plus my professional opinion, right? You know, which I don't state, but that's, you know, like the it's it's that judgment piece you talked about. So haven't had any cause to to drop this on anyone yet, but ⁓ you know, now we have an emoji on Slack for no slop grenade as a result of all the discussion. So maybe that'll get used. We'll see. Shimin (42:34) Yeah, yeah, yeah. that's lovely. Okay, so what I do a lot of times is two form. One is I I have I I do use AI for some of these kind of questions. But one, of course, you need to make it very clear this is AI slop grenade. And two, I would actually write my one line opinion up top with like maybe a two line reasoning back up and then paste the whole thing in case they want the full bell and kaboo. Yeah. Yeah. Dan (43:20) like what yeah, to understand your thinking based on that thinking. Yeah, that's not bad. If you can do it in a way that doesn't like melt the entire chat into the ground, that's probably reasonable. Shimin (43:29) I okay, so like I almost think like these MS teams or Slack or WebEx or whatnot should have like a toggle feature for AI generated account. Yeah, yeah, yeah. Just like you're thinking, yes. Show my AI research. Yeah. Dan (43:38) Like show thinking almost, yeah. Like show prompt and and result, yeah. Yeah. That's fair. Shimin (43:46) That would be a UX improvement. Dan (43:48) And I think I think that is maybe one of the things that people are reacting strongly to on this is like sometimes you just don't care what the prompt was, right? Or what the output was. You really just want the human or L and answer and infused. But there is also like an element of this in like, let me Google that for you. Which is just like well, the microservice architecture. How why didn't you clone the repo and point clawed at it? Like why'd you ask me in the first place, you know? So ⁓ Shimin (44:16) yeah. Yeah, okay. I d that has not occurred to me. That's a good point. Yeah. Dan (44:19) ⁓ I don't know. Haven't really crossed that bridge either. But we we're living in this wild new world where maybe ⁓ let me claude that for you dot com will be the next one besides no slop grenades, but Shimin (44:31) Yeah, just show some courtesy to your friends and coworkers. Like the rules of etiquette has not changed. So like you wouldn't yeah, just throw the book at someone when they ask you what a word means. Like that's not how this works, you know? Dan (44:38) Right, yeah, exactly. Just be be be respectful. Yeah. Let me paste you the entire dictionary definition for it. Yeah. Shimin (44:47) Yeah, just like some guy like just throw a cubicle three cubicles over just lobs lobs a dictionary at you. ⁓ I get it the reference now. Yes, it is exactly like that. ⁓ good rant Yeah, yeah, thank you. Dan (45:00) yeah. What I'm here for. Shimin (45:03) Okay. On to our last segment. Two minutes. to midnight Dan (45:06) Yeah. So let's look at how the AI bubble is doing this week through our ever present lens, the atomic clock. Or yeah. Shimin (45:14) Yeah. Armageddon clock or atomic clock. ⁓ the we we were at six minutes. We've been kind of going a little bit ⁓ farther away from midnight this last couple of weeks. Yeah. And Dan (45:17) My getting clock. Yeah. Backwards. Yeah, it's wild. All the things I had in for 2026, that was not one of them, but here we are. So Shimin (45:31) this week my first article is a ⁓ Morningstar. I I'm not sure if this is a blog or it's more of an investment note No, I think it's more of a blog. by Zachary Evans. Where this is a deeper dive on the ⁓ SpaceX IPO that we spoke about a couple of weeks ago, where because ⁓ openAI anthropic and space access planned IPO will be so large indexes like the SP, NASDAQ, and Morningstar are are almost forced to include them from day one. And this article has some really awesome pie charts about how The largest public companies impact the I think this is a Morningstar 100 index. Before the IPOs, the tech companies, Nvidia, Apple, Alphabet, Microsoft, Amazon, Broadcom, Meta, and Tesla make up more than 50% of the Morningstar 100 index, which is kind of an extraordinary number. Yeah. that's how high our existing concentration is in in AI. And when SpaceX Dan (46:31) Wild. Shimin (46:39) goes IPO, it's gonna be four almost four and a half percent of the entire Morningstar 100 index by itself. OpenAI will be another 3%, and Athropic will be another 2.5%. So altogether, 10% give or take, of the Morningstar 100 index is gonna be just these three AI companies. Bananas. and then it talked about some of the ramifications of this. Usually when an index like SP and NASDAQ ⁓ sees a a new company's IPO, they usually will wait for twelve month of trading for the volatility of the pricing to go away before they add it to the index. But SpaceX has a sweetheart deal, maybe not sweetheart deal, rule change with Nasdaq that ⁓ reduces that twelve month period to something like fifteen trading days. Dan (47:16) Mm, mm. Shimin (47:28) and SP is trying to do the same. It is also in talks of considering a fast entry rule change to the SP 500 to allow for these new IPOs. And of course, the rev ramification of that is everybody who buys SP and NASDAQ index funds, all of our life savings, 10%, give or take, is gonna be funneled. Dan (47:46) It's gonna eat some of that volatility. Yeah. Shimin (47:53) into these IPOs. That is trillions of dollars in fund. Like money money has lost its meaning s sometime around when we started this podcast. But yeah. Are we Dan (48:04) But the the flip side of this that I don't know that I mean we've talked about it in the previous context of the SpaceX thing is that it's now as they go public, they have to start talking about their financing, right? Which means they're also gonna start yeah, I mean the yeah, in the initial filing, but like that means they're also gonna have to start like really showing where this money is going. And I think we're gonna see the same thing happening with Shimin (48:15) They already have, yes. Dan (48:26) like open AI and anthropic too, if assuming they follow closely, as everyone is kind of assuming. And the part that I find interesting about that is the at least the software market, in my experience, has shifted drastically from the sort of like growth mindset companies to, you know, everyone was in like even Meta was doing their like efficiency push, right? Like with that, their like age of efficiency or whatever. forget what they called it, but like and so that isn't necessarily the investor focus anymore. So I'm wondering whether or not they'll be able to get away with it. Shimin (48:59) Well, I have one word for you. Tesla. Dan (49:02) Yes. Shimin (49:05) ⁓ I I'm not sure if these companies will play by the same rules as your average average companies. Yeah. Dan (49:10) Right. No, I'm not either. And so that's basically what I'm trying to say is like will will that still apply or no? I don't know. It'll be very fascinating to see. And also very fascinating to see, like, with real numbers coming, you know, presumably coming out for some of the fr other frontier companies, it'll be fascinating for me to see who's right. Like, is it where's your ed's at has been right all along? Is it, you know like or what accounting tricks are they gonna pull to hide some of these things? Or yeah, it's it's Gonna be an interesting couple months here. Shimin (49:38) We should have Ed on the podcast. We should reach out to Ed. I think I think he would be a great person to have on the on the clock. and speaking of how profitable companies are. Dan (49:46) Yeah, so this fresh off of hacker news, but I thought it was lovely and super on point. is AI profitable dot com? And it's of course a resounding no, which isn't super exciting. But the thing that's really cool about this is it gives you a breakdown of like capital expenditure versus profit by billions. sorted by ⁓ well you can change the sorts but you can change it from like you know capital expenditure revenue spend least revenue blah blah blah so it's actually kind of wild to see that Amazon at least according to their data has the highest capex spend on AI so far at negative two hundred and ninety one billion alphabet coming in second at two sixty two Microsoft in third at two thirty five ⁓ and you know, your old friends Oracle and Meta are on there, but like, yeah, there's a pretty big number jump before you get down there. I don't know. ⁓ you you you you liked Oracle a lot for Shimin (50:39) Why why are they my old friends? I I I I do I do like Oracle a lot. ⁓ when it comes to I believe Oracle is the canary in a coal mine, so yes. Yes. I'm not here like, you know, being a world's biggest fan of Java or something. Yes. Dan (50:50) We'll pop the bubble. Yeah, exactly. That's what I mean. Like in that sense, not a I'm still mad about the JavaScript trademark. Yeah. But it's just funny to see, like, you know, we go down quite a bit until we get to the frontier labs, right? Like opening eyes at minus twenty-seven and thropics at minus twenty-six. So it's they almost feel like a drop in the bucket compared to the infra players. I guess shouldn't be that surprising. ⁓ and of course the only one in the green is ⁓ is the green company. Shimin (51:19) Of course, of course. By by a positive two hundred and fifty three bil Dan (51:22) Nvidia by a significant margin. Yeah. Two hundred and fifty three billion. Yeah. It's kind of funny. Shimin (51:28) ⁓ And of course it's it's not like even ⁓ Nvidia at plus two hundred fifty three billion dollars is l significantly less than like what Amazon has is losing, right? Like that's two ninety two ninety one. So net is probably investment. ⁓ which is fine. Dan (51:40) Yeah. Yeah, I mean you could also look at like both Amazon and Alphabet have their own inference platforms too, or like compute platforms that they've spent time developing the silicone fower. So that number isn't apples to apples, right? ⁓ Shimin (51:54) Yeah. Yeah. Right. And they are gonna get ongoing capacity from this investment. Dan (52:04) Yeah. Yeah. Shimin (52:05) Yeah. I was shocked at how well Oracle and OpenAI are doing compared to everybody else. You know, as a team Oracle and guy. And I know you're an open AI guy. I I'm here looking at this going like, Oracle only lost thirty nine billion dollars and OpenAI only lost twenty seven like what trump change. I think we do have Dan (52:16) Ha ha Well, wow, great. Yeah. Man, I spent that fishing this weekend. You can't even I don't fish, I just yeah. Shimin (52:31) ⁓ I I think the difference is again, ⁓ Oracle has a lot of debt to fund that negative thirty nine billion. Whereas say what you will about Amazon, Alpha Bit, Microsoft and Meta, they are cash flow generating machines. So I do think that that is a difference. Is like they are losing so much money because they can. And I think other than Meta, the other three have not been getting debt financing. So they're they're essentially burning their revenue in this, which is much more sustainable. No. Okay. Dan (53:02) Yeah. Is it the right bet? Who knows? But ⁓ definitely means they'll probably still be around as companies. Shimin (53:08) Yeah. So all that said, how do we feel about our Clock this week. Dan (53:13) That's a good question. I mean, I don't think we have enough data until these IPO start coming out to start really making big moves here. I think that we're still seeing cash flowing in, right? Shimin (53:24) Yes. I want to say that the fact that NASDAQ changed its rules is a sign that you know the the three IPOs we're looking forward to this year will have more financing locked in. 'Cause like SpaceX already has billions of dollars pretty much locked in because of its Nasdaq rule change. Well, SpaceX OpenAI and Anthropic, assuming they IPO this year. it's retirees and teachers unions out there are already giving them some of their funding with the blessing of Nasdaq. ⁓ S and P, which is by far the biggest of the indexes, are still talking about the rule change. But Dan (54:04) Mm-hmm. Shimin (54:05) I'm I'm leaning towards like going back maybe fifteen twenty seconds. Dan (54:09) What away? Like 615? Yeah. Yeah, I think that's fair. Cause I think that if if that happens Shimin (54:10) Away. Yep. Yep. Dan (54:16) It's probably gonna it's gonna mitigate the cash crunch, right? That we've been talking about on and off here on this segment. And then or at least to some degree. And then it's also going to potentially cause like the transparency may or may not offset that. We don't know. So that's fair. Okay, I'm with you. Shimin (54:34) Yeah. I mean as as sad as I am, ⁓ I feel like Elon has his pinky in my pocket right now. Ooh, that sounded that sounded wrong. Okay. I'll note. Dan (54:41) Ha ha ha. If you've been spending more time talking to Grok, is that what's happening here? Shimin (54:46) p yes. Gotta ask Grok to generate a picture No, I'm not going to. That is disturbing. Dan (54:50) ⁓ Grok got skills. We missed that in the news too. Grok has finally gotten skills. Shimin (54:55) ⁓ they yes, they have their own coding harness being released. Dan (54:57) So weird thing to be catching up on, but but here we are, so Shimin (55:00) Well, that's that's a good thing, right? Like folks are definitely gonna use Grok to create only good skills, not Mecha Hitler skills. Dan (55:09) Do you even need to create those? I feel like that's just ⁓ normal prompting. The Grok experience. Shimin (55:11) Yeah. It's inherent. It's part of the ⁓ RLHF process. Yes. Okay, so are we good with six fifteen? Not a ton of information, but I feel like they are stealing everyone's 401k ⁓ stealing loosely in quotes. This is not financial advice, people. we haven't said that in a while. do not make investment decisions based on what we talk about here. And ⁓ at six fifteen Dan (55:19) Yeah, I think so. Yeah. Shimin (55:35) We have the clock set and that will be ⁓ wrap on the show. ⁓ thank you listeners for joining us on our study session this week. If you like the show, if you learn something new, please share the show with a friend. You can also leave us a review on Apple Podcast or Spotify. It helps people to discover the show and we really appreciate it. If you got a segment idea, a question for us or a topic you want us to cover, shoot us an email at humans at adipod.ai. We'd love to hear from you. You can find the full show notes, transcripts, and everything else mentioned today at www.adipod.ai. Thank you again for listening and we'll catch you next week. Bye. Dan (56:10) And if you missed Rahul this week, he'll be back in a couple of weeks. So just bear with us. Shimin (56:14) He is being kidnapped by the Mecca Hitler there. Dan (56:16) Mm. It's true.