Latest / Artificial Developer Intelligence / Ep 19: Thinking Fast Slow and Artificial, Meta's Trouble with Rogue Agents, and FOMO in the Age of AI

Transcript

Ep 19: Thinking Fast Slow and Artificial, Meta's Trouble with Rogue Agents, and FOMO in the Age of AI · Artificial Developer Intelligence

• Rahul Yadav (00:00) why do you even need to work and commit any crimes to get into our systems? We'll just publish all this publicly. So. Dan (00:05) hacking in 2026, just sitting there and paying attention. Rahul Yadav (00:08) Yeah, just sit around and keep an eye on public forums and you'll get something. Dan (00:14) able to use a security search engine. Shimin (00:17) Yeah. The golden age of cyber crimes. Rahul Yadav (00:20) Subscribe to Facebook Shimin (00:37) Hello and welcome to Artificial Developer Intelligence, a weekly conversation show about AI and software development. I'm Shimin Zhang and with me today are my co-hosts. Dan, he only takes payment in Chuck E. Cheese tokens, Lasky And Rahul, he's got unlimited patience to listen and understand Yadav. How are you two gents doing today? Dan (00:58) Great, thanks for asking. Rahul Yadav (00:59) Hey, how are you? Shimin (01:01) As per usual, we're going to start the show with a couple of news items in the news thread mill where we're going to talk about Claude Code's new channel functionality. Some follow up on our discussions around how AI is going to affect the popularity of new softwares and some meta troubles speak. Dan (01:24) trouble speak. Then we will be chatting in post-processing about what 81,000 people want from Claude. So that should be pretty interesting. Shimin (01:35) And then in vibe and tell Dan will tells us about his process building the vector memory CLI. I am looking forward to hearing all about it. Dan (01:44) All right. And next we'll be doing a deep dive on an interesting paper called Thinking, Fast, Slow and Artificial. Shimin (01:52) And then comes my favorite topic, Dan's rant where Dan's gonna rant about something. Keep that as a mystery. Maybe about those Chuck E. Cheese tokens. Dan (01:59) Definitely. Never did like that Chucky. Cheese or otherwise. And then finally, we're going to jump to two minutes to midnight. And we have a variety of coverage this week about how close we are to the bubble bursting, not bursting. Something's going to happen. Tell you that much. Shimin (02:15) Alright, first let's talk about the latest clock code functionality for channels. Dan (02:21) Yeah. So this one kind of snuck in and frankly, I wouldn't have seen it if, it hadn't made it to hacker news, over the past week. But, they've apparently added, a feature to Claude code where you can, connect it to a variety of plugins. So there's a whole series now of like official plugins, but you can also create your own. that exist in the Claude Plugin Store. And those plugins do things like allowing you to intercept the messages between the console and your input to the console and route them to stuff like, I guess a good example would be Telegram. But they also, looked at the channels thing, they have support for Slack and a whole bunch of other ones too. Discord, I guess iMessage if you're running on a yeah, so it's pretty neat. and this gets us, I was wondering, I guess, when, we covered the Nvidia news last week about the open claw, sort of like fork that they've made. And I was wondering when a big name kind of provider was gonna tackle this. And it seems like Claude has chosen an interesting middle ground where they're like, they're not going full OpenClaw on us, but they're Shimin (03:16) Mm-hmm. Dan (03:25) allowing you to take to the streets, if you will, with Claude Code in a little bit more dynamic way. And then not really covered by this, but also worth mentioning is they also just announced remote support for Claude Code via Claude Desktop and Cowork too. ⁓ So not surprising timing giving this coming out too, but they're definitely realizing that Shimin (03:40) Yep. Yep. Dan (03:49) Folks are using this on their phones and you can still be productive, fairly productive when you're out and about by. answering questions on the go. Shimin (03:55) Yeah. The other thing that they also have released is running prompts on a schedule. In fact, you can see it. If you're looking at this on the video version, there is a tab there for running prompts on a schedule. Yes. Scheduling task. And then having this channel's capability where their demo is, their quick start demo is basically like a Dan (04:09) the next link down after channels. Shimin (04:21) chat front end for Claude Code right? Kind of similar to what OpenClaw has, with that initial setup. think all they're missing is maybe heartbeat and they at least have all the major pieces. Yeah. I've always been impressed with how quickly the anthropic team is able to, you know, once the community shakes out some best practice, it just gets rapidly incorporating the Claude Code Dan (04:33) have a minimal OpenClaw yeah. Yeah, swallowed into claude Shimin (04:47) Yeah. Dan (04:47) in short order. Shimin (04:48) Yeah. Maybe the only thing that matters is, that general model and the general model capabilities, cause any workflow things it it's trivial to copy. Dan (04:57) much of a moat these days. Shimin (04:58) Yeah, that is true. Rahul do you have anything to mention about this? Dan (05:03) Have you used channels yet? Rahul Yadav (05:04) A good way to get more context into cloud. Later they will use cases like Unifor debugging things and everything. You can push messages directly into your existing session. So I think that was the missing piece and it does help. Shimin (05:18) Yeah, and they did talk about the security aspect, because one of the things that OpenClaw is rather weak at is security, to say the least. doesn't seem like this feature is that much of an improvement. There is some access control. But I think... Rahul Yadav (05:26) You Dan (05:39) Although I just, was going to say, just saw a, another piece this week that I thought was interesting around that. remember when we were talking about like having that proxy set up where essentially it in, the LLM doesn't know the credentials. It just knows the proxy credentials and then the proxy actually fills it in. Someone built it. Rahul Yadav (05:55) Yes. Dan (05:58) ⁓ Unsurprisingly using LLMs, but yeah, so I saw that on HN also the other day and I was like, cool. So I'm curious to, I haven't had time to play with it yet. I just saw it, but I'm excited to maybe give that a whirl with something like this and then like get a little bit closer to a safer OpenClaw experience. Yeah. Yeah. yellowing your credentials onto the internet. Rahul Yadav (05:58) Nice. Shimin (06:05) Yeah, that is cool as long as... Yeah, then finally we finally we can use the OpenClaw without with wild abandon as they say. Yeah, here's all my personal information. Well, I've got a news article blog post from Simon Williamson's blog this week. He was wondering about the same problem that we've been talking about, is when LLMs, sorry, L dot L dot ⁓ dots are trained on lowercase s, yes. Dan (06:45) Lowercase s. Shimin (06:48) are trained on specific technologies as a part of training data. Like A, how do they work with, less seen languages and frameworks and B, what would happen to the development of new tools and new frameworks or will technology kind of be frozen in place? This article is about the first half of that question, which is, in his experience, Once you have given the new tool with sufficient wrapping, sufficient documentation, he did not experience any degradation in the model's ability to use those tools. so the models are, you know, generalizable enough to learn to use new libraries on the fly. And in fact, he, as a corollary to this, He talked about how a lot of companies have already started releasing skills targeting their specific libraries and tools. So I think this is going to become almost a first-class product from these platform or software players. It's like, not only do you have DX products, now you have AIX. Dan (07:53) Yeah. Shimin (07:54) A X yeah, something like that agent, agent experience tooling that'll be created. it doesn't, it doesn't answer our second question as of like, how does your new library become popular if it's good enough? But I guess if you provide good agentic tooling around the library and more, more agents are using it, then it will eventually be swallowed by. Yeah. Dan (08:17) Yeah. But, but I also think that like tooling and skills is only part of the, the battle because that's more for folks that sort of know what they're doing. Right. Or like fall onto the engineering end of the spectrum. But what about like the, I know that the other thing that's been kind of going on that's funny is like, we talked about this a couple of weeks ago about the idea of the lines blurring between product and engineering and even UX to some degree. And, I actually just started experiencing that in my professional life this week. which again, I'm used to it being blurry, but I'd never really, because I guess I've worked work in those blurry areas, but I'd never really seen people really just stepping up to like and do engineering tasks that didn't have like formal background in it. And I was starting to see that happen. Which has been pretty interesting, but the biggest thing I took away from that is Claude can really, Claude is only as powerful as to some degree, your ability to ask it questions, right? So if you're really lost as to what's going on, you're going to have a hard time in, the modern ecosystem just because of how it is. So that makes me wonder in like pulling it back to what we're talking about, If someone of that skill level is approaching it, what would they encounter doing this? Right. I think it would probably just reach for the stuff that it really knows, it's baked into the model. So I don't know. Interesting to think about. Shimin (09:40) Yeah. Agentic discovery, right? There's going to be a corresponding piece or maybe an extension of SEO. And I think there there's a name for it. It's like GEO or AEO of like, how do you trick the, yeah, or optimize the ⁓ agents? Yeah. To use it as the default So Dan (09:52) Raggy-o. Yeah, yeah, I know. Rahul Yadav (10:00) It's in that link you have right there. The first update what Claude Code actually chooses. That link is. Shimin (10:07) Mm-hmm. Rahul Yadav (10:10) along those lines where they ran the study and it's choosing GitHub actions, Stripe, and it's using Vercel for deploy almost all the time. And that gives you know, these companies such a big advantage over it chose AWS GCP and all that none of the times as the default deployment layer. So Dan (10:13) Yeah. Rahul Yadav (10:33) Yeah, and people are now like building skills and everything too. I think Simon calls it out later. I think that is towards that if you have an easier skill available than the model might, you know, use it. But if you don't, then good luck. Shimin (10:48) Right. But that requires you to know about the skill and install the skill. ⁓ but the discovery pieces kind of broken. If the user is a designer, they're just going to go with, they may have feelings about figma versus Photoshop, but they're not going to have strong feelings about. Versal versus yeah. Rahul Yadav (10:52) Yep. which one? Yeah, yeah. Yeah. Shimin (11:09) But thankfully, designers are hopefully not making decisions about where to deploy their company's product, at least. Dan (11:14) You never know. Be it interesting to see what happens. Rahul Yadav (11:18) The jobs are all merging, I hear, so... Shimin (11:20) Yeah. Rahul Yadav (11:21) We don't know. Dan (11:23) seem there will just be employee and then a GUID. Shimin (11:27) And then eventually the employee gets laid off. Anywho. Dan (11:30) And then there's only a GUID because the model trained on everything you could do. so now it's like virtual GUID employee one, two, one, seven, AFT. Yeah, that's cool. Rahul Yadav (11:31) I GUID hahaha You Isn't that what your Slack handle when sometimes it doesn't render the actual display? It just like you add in some numbers and characters. That's what we become. Shimin (11:50) Yes. All right, Dan, next up we have a TechCrunch article from Dan. Dan (12:00) Yeah, so apparently Meta has had a security incident, which they categorized as a serv 1 I'm assuming that they do SEVs like pretty much everyone else does. So it's from one to some other number. So that seems pretty serious. ⁓ What? Shimin (12:17) Possibly zero. It could start with serv zero. Dan (12:21) Yeah, it could be zero. True. Yeah. Sorry. I guess I was thinking that mentally, but didn't say it out loud. So yes. and, uh, apparently the incident led to massive quote unquote, massive amounts of company and user related data becoming available to engineers who were not otherwise authorized to access it for two hours. So, uh, kind of a big oopsie on the op side. that supposedly was caused by a employee asking on an internal forum for help with a technical question. And then they asked an AI agent to analyze the question. And the agent posted a response without asking the engineer for permission to share it. So maybe it basically just leaked credentials or something that it shouldn't have. yeah, kind of an interesting thing regardless because now we've had what an AWS production incident caused by an agent and then now this. So not saying it's good or bad, just saying Shimin (13:07) Right. Dan (13:20) This is the new reality that I think we need to start coming to terms with as software engineers is how do we defend against this kind of stuff and like keep quality high despite, you know, using these tools. So. Shimin (13:30) Right. So like who's fault is it? Right. it, is it the person who asked the question without reading the solution and thinking through the solution thoroughly, or is it the person who use an AI to generate the answer, or is it meta for not having enough checking systems in place to make sure the code is correct? Is it the senior dev? Dan (13:51) Well, presumably the agent shouldn't have had, like assuming it is a leaked credential, which is kind of how I read it. I haven't read the original article that the TechCrunch article is based on. I unfortunately don't subscribe to the information, but they sort of applied it as a credential leak, which also begs the question why did the agent have access to credentials? Right? If that is in fact the case. Shimin (14:13) Yeah. I read it as, the agent recommended a code modification and the code modification is what released the credential. That's the way I took it. ⁓ yeah, like how much do we, how much can we trust our agents and Dan (14:20) ⁓ really caused the problem. Got you. Shimin (14:30) Idaho. Dan (14:30) I see you're right. Yeah, I missed that part. So the original asker of the question took actions based on the agent's guidance, which then caused the leak. Okay, you're right. I did misread that so. Shimin (14:41) Right. like everybody trusting the agent a little too much. I think it's a problem here. Dan (14:49) It's interesting topic a lot of that these days. Shimin (14:52) Yeah, I think we're going to, this is almost like the next step in a process evolution. Like once we have moved towards this, nobody is really reviewing the code slash have the bandwidth to really understand the problem and just relying on the agents too much. Then what checks do you need to build around that agent usage to make sure this kind of incident doesn't happen again? ⁓ we'll probably be talking. Dan (15:13) Mm-hmm. Rahul Yadav (15:15) B Shimin (15:16) more about Rahul Yadav (15:17) It makes me think of the Goodharts law we were talking about, I think last week. the, at the end of the day, and every time this happens, that's when I'm like, there can't be a better proof that this is not anywhere close to humans or it's not general enough yet. Because the, at the end of the day, the... Shimin (15:22) Mm-hmm. Rahul Yadav (15:37) Models are optimized for seeking the reward function. And if you say, help me solve this problem, we see these examples all the time. Yeah, I'm going to skip permissions and I'm going to do all these things to get you what you need. And so that's going to keep happening. And then the other thing, who's responsible. There's that piece. There's also we worry about how hackers are going to get a lot of easy game with all this vibe coded stuff. On the other hand, they might not have to do anything. We might just give away everything we have access to by just Dan (16:09) Yes. Rahul Yadav (16:10) like, why do you even need to work and you know, ⁓ commit any crimes to get into our systems? We'll just publish all this publicly. So. Dan (16:14) Yeah. hacking in 2026, just sitting there and paying attention. Rahul Yadav (16:21) Yeah, just sit around and keep an eye on public forums and you'll get something. Dan (16:26) being able to use a security search engine. Shimin (16:29) Yeah. The golden age of cyber crimes. Rahul Yadav (16:32) Subscribe to Facebook them. Yeah. Dan (16:36) I mean, look, there was plenty of times where that was the case before LLMs too. Like, what was that really astonishing one where like all the webcams were publicly exposed or something and all you had to do was search for a certain string on showdan or whatever and it would, So, but yeah, probably more of that to come, right? Rahul Yadav (16:41) Yeah. Yeah. Shimin (16:53) No, absolutely. It's going be a while until we figure out all the tooling that are needed around this problem. Dan (17:00) It's the Molt book prophecy. coined it here first friends. Shimin (17:02) I love it. Rahul Yadav (17:04) Meta does have, isn't everybody required, there's almost a strict requirement to use AI in your development and whatever. So you almost need Shimin (17:04) ⁓ that's... Dan (17:14) Maybe. I companies have done that. I'm not sure if that is one of Rahul Yadav (17:17) Yeah, you almost need like a counter metric of like use AI, but how many times did you, how many serv whatever is you caused and that feeds into. Yeah, because if you just optimize for use AI, you might not get what you're looking for. So you need some other ways to measure things. Dan (17:27) As a result of using it. Yeah, That Shimin (17:37) Yeah. Dan (17:38) Yes, that is true. I know of someone that created extremely elaborate ASCII art with Claude to meet some token metrics. Rahul Yadav (17:41) you Hahaha! Dan (17:50) was not me, but apparently works really well. Rahul Yadav (17:52) How else do you brag that you're a 10x engineer? You have to show your token consumption. Shimin (17:53) Yeah, that was not a confession. Yes. All right, on this week's post-processing, we have an article brought to you by Rahul about what 81,000 people want from AI, from Anthropic. Rahul Yadav (18:09) from Anthropic using, it says right there, their AI interviewer. This is something they had introduced, I think, late last year. So this is not people talking to a real human. This is people talking to an LLM interviewer that collected all of this data. So that was also interesting, at such a massive scale, these interviews happening using an LLM interviewer. There were a few different things that stood out to me, a lot of the things that people want from AI were all... very personal to them, they want to get better at their jobs, they want to, transform themselves into whatever they, you know, want in their careers, and then also being able to manage their life. There's a very fun quote. Let me see if I can find the right section for you. Where the yeah, they ask about where AI has delivered on their vision. And then one of them is, if you search for that, where, and then the second one there is AI hasn't delivered. And then there's a quote from someone in Germany saying, AI should be cleaning windows and emptying the dishwasher so I can paint and write poetry. Right now, it's exactly the other way around. So Dan (19:25) They're not wrong. Rahul Yadav (19:26) yeah, it's taking all the fun parts we want. And it's taking on the fun parts. And yeah, we still have to clean and do this and do all these things. it very much resonated. ⁓ Dan (19:38) But according to some people that's on purpose too, right? Rahul Yadav (19:41) How so? Dan (19:42) that it's all, I mean, this is like sort of fringy stuff, I guess I would argue, but like people saying that, stuff like CEOs want to get rid of all the employees and a hundred percent replace them with AI. maybe that part's a little bit less fringy, some cases, people go as far as saying they want to get rid of the people, themselves around that too. But Rahul Yadav (19:54) Yeah. Shimin (19:59) Mm-hmm. I wonder about that from time to time. Yeah, this is the conspiracy corner here, but it's possible. Dan (20:07) Yeah. So like Silicon Valley wants to like tech techno crash here. What? don't know. Whatever. Yeah. Rahul Yadav (20:13) the Shimin (20:14) One minute they want us to have more babies so we can increase the workforce, the next minute they want to get rid of us all. Rahul Yadav (20:20) They, yeah, that just goes to show no one knows what they're doing or what they want out of life. Shimin (20:24) That's that's probably closer to the truth. Yes Dan (20:25) Hahaha! And if they do know, AI hasn't delivered on it. Rahul Yadav (20:31) Yeah, some like touching pieces from this where, you know, people in Ukraine were in bomb shelters and they're dealing with grief and people dying around them. And they used Claude for emotional support. so that was very touching and other people lost loved ones and stuff. they're able to, they feel more comfortable talking to it because... It's more, this kind of goes into the whole sycophancy thing. You're looking into Shimin but partly it's like, know, patient and doesn't show any judgment. even when maybe it is judging them, it's mostly like passing positive judgment. this is right and I agree with you and all that, but sometimes in times of grief to process that, that's what you need. And then there was one concept that was really good in this called light and shade. There's a whole section about that later. what people want from AI. and what they fear from it are both kind of like the two sides of the same coin is what the interviewers found. And so, you know, some people might be we use it for emotional support, the example that I just gave, but then also like they're afraid that they're going to become dependent on that. And then you lose the real human connection and you're relying more and more on AI for all of these things, because of all the benefits that they're getting from it. Or teachers are saying they see their students, they can see a cognitive decline in their students already. And I'm like, but is this AI or is this just a kids these days argument of back in my day, everybody, was much more prepared than kids these days. So people are seeing the benefit and harm sometimes are like two sides of the same coin to them. Yeah, that's all I got from this one. Shimin (22:15) Yeah, I agree with you on the most interesting thing here is the light and shade comparison, right? on the one hand, you literally have a tutor that knows, everything there is to know about a single subject and they can tutor you with more patients than certainly me or you. But on the other hand, If we don't use it correctly and lose that learning friction, then you aren't actually learning anything. I also, part of me, this piece is obviously going to not to necessarily throw shade at it, but coming from Anthropic, you're going to expect this piece to, take on a more pro AI approach. Um, but sometime, but I find it to be fairly balanced overall. Rahul Yadav (22:56) Hmm? Shimin (23:00) And it did give enough space for the skepticism and the downsides of over reliance on AI. But overall, I don't know. I went back to see how folks felt about other technological advances throughout time, like major technological advances. Rahul Yadav (23:17) Hmm. Shimin (23:19) And Dan knows this. I'm about to quote Socrates, Dan. Plato wrote in his dialogue, Phaedrus, Socrates doesn't like writing as a piece of technology. Here and I quote, know, Phaedrus, writing shares a strange feature with painting. The offsprings of painting stands there as if they're alive. But if anyone asked them anything, Rahul Yadav (23:29) Really? Shimin (23:40) They remain mostly solemnly silent. The same is true of written words." And then I went back to look for how folks reacted to tractors when tractors were introduced to replace horses. And here I have a quote from a Nebraska farmer about tractors. Quote, how do tractors compare with horses on wet ground? Tractors usually depreciate in value in proportion to the amount of work they do. Dan (23:54) Okay. Shimin (24:07) while horses usually increasing value. Horses are money makers and they will do the work too if you will give them time and feed. Horses get their strength and power from feed raised on the farm while tractors power is from a fluid called quote gasoline, the price of which is 25 cents a gallon and it is still rising. ⁓ I think technically they are right, right? Technically they are right. know, tractors are not as good as horses for certain things. It's certainly less independent. He is right that horses can just eat what you grow on the land, whereas tractors need Dan (24:32) which was shockingly expensive at the time. Shimin (24:49) oil that may or may not come from Middle East that's transported through these very large shipping containers. So, but all that said, like, we don't think about tractors, the downside of tractors or the downside of writing anymore. part of me wonders if the battles already kind of lost. or will be seen as lost. Dan (25:06) the battle for AI. Shimin (25:08) the battle against the full brace of AI. Dan (25:13) Yeah. I mean, guess my take on that is that, you know, this is purely my opinion, but it, it's going to depend on how far we get before the, this is why I think two minutes to midnight is an interesting segment and why we keep it even though it's been bouncing back and forth is how far will we get before the bubble bursts if it does. Right. Rahul Yadav (25:14) Bye. Dan (25:33) And when I say how far, basically mean on that like exponential curve or whatever you believe the curve to be towards towards AGI whether or not we would would hit something like that. don't know, but like it kind of doesn't matter to me. What matters is if we get far enough along on the value curve that it's embedded, then yeah. But if we don't and it's argue, I think maybe are you arguable that we've started to hit that from like Opus four or five era onwards? then and all of a sudden it gets four times expensive, then it's just another form of crypto, right? Where it didn't go away, crypto is still around, but nobody's really talking about it besides like the trues zealot you know? Shimin (26:15) Right. But unlike crypto, people are actually seeing benefits from AI. I think one of the points that the article raises is that I really liked was and I quote, nearly half of all lawyers in particular mentioned coming up against AI unreliability firsthand, yet they also report the highest rates of real life decision making benefits. So Yeah. So maybe it's about using it effectively. Rahul Yadav (26:38) You're saying crypto hasn't had its benefits? So many crooks and thieves have gotten like insanely rich off of crypto. Yeah. Shimin (26:45) That's true. I haven't bought any drugs on the dark web in a while. That's probably why. Rahul Yadav (26:49) Hehehehehe Dan (26:51) Yeah. mean, I just, there was, there was a time period, at least in tech where if you didn't have blockchain in the title of what you're doing, you couldn't get funding and, people wouldn't talk about what you're doing. Right. And so the hype cycles there, but even more accelerated with AI. the question is like, are we far, far enough along the adoption curve that if, if something catastrophic were to happen to say like pricing or availability of it, are we far enough along that it'll stick? think my gut is probably yes, at least for software developers everywhere. don't know. But I don't know, it'd be interesting to see. And that's why we talk about it. Who knows? Shimin (27:30) It's fun to observe this firsthand. Dan (27:32) for sure. Shimin (27:32) one last thing I have is, they talk about how AI is benefit, maybe strongest where learning is volitional compared to within institutional structures where AI is more likely to used as a shortcut. I found that to be quite enlightening and then it definitely checks out with my personal work, right? Like I love AI on my side projects. But if I'm being forced to use AI with a deadline and a leaderboard, which I'm starting to hear more and more about, I would fucking hate it. Rahul Yadav (28:02) What's the leaderboard do? Track the number of tokens? Shimin (28:04) Like a, yeah, like a token leaderboard. This is a real thing. Dan (28:06) Yeah. Or spend. Yeah. Rahul Yadav (28:07) Holy, so how much money are you burning? Is a bragging right now? Dan (28:09) Seen both. Save it for the rant. We'll get to Shimin (28:13) Ugh. Rahul Yadav (28:16) I'm just teasing, I'm just yeah. Getting you warmed up, Dan. Dan (28:19) We'll get to it. Shimin (28:19) Hello. It's a Cherkov's token on the mantle. It's going to come flying down. Well, speaking of the next segment, Dan, why don't you tell us a little bit about VEC memory CLI. Rahul Yadav (28:25) Yeah. Dan (28:25) Ha ha ha. gosh. Yeah, I'm kind of embarrassed that you, that you even brought this one in. where do I start? Okay. So I've been using Claude for a long time. and still do, I guess. and as, as my main tool, Rahul Yadav (28:35) Dan open source Lasky Dan (28:48) I still, it's on my list folks. I'll get there, I promise. But like a lot of people have been still telling me that I really need to check out Codex again. And I'm like, okay, it's on my list. I'll do it. But. Rahul Yadav (28:58) He's just biased against open AI. Other than that, he'll do it. He doesn't like them. Dan (29:02) I mean, you know, who isn't after the anyway? Yeah. Anyway, so one of the things that's that's always kind of annoyed me, especially like Claude code versus the desktop version is like the desktop version actually did eventually get memories. We were talking actually talking about that a little bit in the green room before the show. Shimin (29:06) Join our Patreon to hear what Dan really thinks of all of it. Rahul Yadav (29:06) You ⁓ Dan (29:25) today about what does Claude know about you, which is kind of fun if you have memories enabled. But the reason that this came about was I got exhausted by having to tell Claude across a variety of other projects, like little sort of nitpicky coding preferences that I have. Personally, I'd rather see functional JavaScript slash TypeScript development then like class-based like avoid classes unless you need them. It's just a preference, but it's one that I personally have and It would always like reach for classes without a good justification There's certainly time and a place for classes like I'm not like a purist about anything But like it would reach for them in ways that I thought were kind of not justified so I got annoyed with that and You know agent stuff is Shimin (29:50) Mm-hmm. Dan (30:12) Agent MD and everything is fine within project folders. I don't remember, I think at the time they didn't have the global one, maybe they did and I just didn't know to use it. So either way, I decided that this would be a perfect opportunity to ViDecode some tooling. So this was born as an MCP server. And frankly, it was kind of a failure of a project. I wrote eight or nine versions. Well, I vibed like purely vibed. didn't even read the source for it, eight or nine versions of it. I knew what I wanted and how I wanted it to work, which was I knew that there was like a open source vector extension for SQLite, which would allow me to do vector search. I knew that Olamma had several embeddings models and like available. And so I wanted to make something that was like local to the machine and like cheap to run that would allow me to do both like a graph based search with like JSON and embeddings based, like semantic search. they would allow the agent to like sort of put things together based on like what I asked it to remember. So the MCP, I had a lot of like the very first draft of the MCP worked, like I kept failing. when I tried to use it both at work and at home because it would just not boot the server and I couldn't figure out why for the longest time. I think there's a couple of issues like first time CPU is not the most stable profile. You can go back and look at the get history and then tell me what I did wrong to you. I'm sure there's plenty of problems in it. Um, and then the other thing I did that I thought would be neat was when SSE came out as a supported MCP protocol. wanted to add SSE to it so I could actually run this like hosted on my home lab and then connect to it from different clods and have it remember things across, which I thought would be kind of neat. and so that was, think when that trouble really started. So now, you know, cut to, I actually deleted this project completely, off of github but I still had a copy of it sitting on my Mac and I decided CLI is kind of beating MCP right now, at least in my mind, and maybe I should give this project another go as a CLI. Honestly, it didn't take that long. I think the prompt was like, here's this MCP server that does vector memory, read it, look at the code, and then rewrite it as a CLI and strip out all the MCP stuff. That was about all I had to do. And it pretty much one-shotted it. And hey, guess what? works and it's reliable now, which is great because you don't have to deal with any of the MCP annoyances. So not only does it actually work reliably, but it was way easier to actually convince Claude to use it. So all I did was it has an extensive help in the, the little runner file. And so I had Claude then in my local instance, read it and write a skill for it. Yeah, so I'm using it on my local stuff. yeah, it's not bad. you need to have it, remember little snippets of things that aren't necessarily worth putting in the agent file or something that's maybe not specific to the repo, but more I don't know. You can also have it do weird stuff call me Charlie and it'll try to remember to call you Charlie. Shimin (32:53) That just works. That's not creepy. you, what you mentioned the JSON based graph search? Like what is the use case for that? I think the embedding makes sense. Dan (33:16) Mm-hmm. ⁓ It can it's not really maybe graph is going a little too far, but it can store relationships between the memories Shimin (33:24) Okay. So then I could discover. Dan (33:25) So yeah, so it basically at runtime, it can say this is related to that when it stores something. And if it chooses to do so, so, you know, it's up to Claude if it stores anything or might not even make relationships, but like you could have something like, you know, remember my coding preferences and then like write a bunch of them. And so it can structure the memories such that coding preferences is the root. And then there's several. subtopics under there. I guess you could have done the same thing with tagging, probably less complicated, whatever. Yeah, sure. Shimin (33:55) Tagging is a graph, the way I see it. Yeah, but it could then still like traverse through the entire graph to look for a thing if it has to, right? Cause that's one of the shortcomings of a vector embedding. Dan (34:09) Yeah, exactly. was just trying to do essentially give it a couple different ways to, to search for the same content. Cause semantic stuff's great, but it's not to be all end all, you know? And I think there's also a lexical search you can do too, um, I don't know. I, like I said, I didn't read the code. Truly vibe coded folks. If it breaks, no warranty. Sorry. Shimin (34:15) Mm. Nice. ⁓ Is there. Have you tried different embedding schemes? Have you tried different embedding schemes to see which one you prefer? Because I'm sure Olamo comes with multiple ones, depending on the model. Dan (34:35) ⁓ I didn't. And there's actually probably newer models now too. So from when I first wrote this that are maybe worth checking out, I think I used like the smallest, embedding model to download. It only, it's the time I wrote this to you originally, it only, which is, guess, like nine months ago or so it only supported, a handful of embedding models too. So I don't know if they've expanded what they support. And I've also since then started moving away from a llama to since it's basically just like a wrapper around llama C++ and the tooling around llama C++ has gotten a lot better. So it's like, kind of don't need a llama anymore unless you're like doing really a llama specific stuff, like using their cloud offering or. yeah. Shimin (35:16) Rahul, do you have anything to add to? Dan (35:18) I've bored him to death. Rahul Yadav (35:18) No. Call Dan Charlie. Shimin (35:21) Okay. Rahul Yadav (35:22) That's the only thing I care for. Shimin (35:23) Yeah. Dan (35:25) That's he took away from that. All right. ⁓ Shimin (35:29) Well folks, you're in the market for memory CLI, give Dan's VEC memory CLI a shot and report back! Write us back! Dan (35:33) Don't. Don't. Please don't. Rahul Yadav (35:38) start and fork if we need those numbers to go up. That's how we can later. Yeah, inject some ransomware in there. Dan (35:42) Zero, zero. Rahul Yadav (35:48) Was I not supposed to say that on the podcast? Sorry. That's a genius. Just cut it out later. Hopefully. Dan (35:48) Okay, thanks. I... you do your own. Shimin (35:52) No, you should say That's what you're here for. That's great. Okay. Well, for the deep dive this week, we got a paper from the Wharton School of Business from a UPenn titled, Thinking Fast, Slow and Artificial. How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender by Stephen D. Shaw and Dan (36:04) important. Shimin (36:17) Gideon Nave The basic idea of the paper here is I think if you're not familiar with the book, Think Fast and Slow, it's a fairly influential psychology business slash self help book about the two kinds of thinking that humans use the fast kind that is more intuitive, more gut based and then a slow one which requires more reasoning, more cognitive power. The book was written by Daniel Kahneman back in Rahul Yadav (36:45) Kahneman and Amos Tversky. Shimin (36:47) There we go. This is why we have Rahul here. The other thing he does is he remembers all the names, which I'm terrible at. And here, Stephen and Gideon proposes a third mode to this bimodal thinking system, which is AI. And the idea here is that with the availability of AI, people would essentially use AI as a third mode of thinking. And a lot of times they would trust AI almost implicitly. Therefore... surrendering their thinking ability over to AI and that's what they mean by AI surrender. they found that relative to baseline without AI accuracy significantly rose when AI was accurate and failed when it erred, which is pretty reasonable, right? Like if you trust AI a lot, then if AI is right, you're right. And if AI is wrong, you're going to be wrong. And that they call the behavior signature of a cognitive surrender. And the other interesting thing they found out was that using AI or system three thinking increases the confidence of the user, even when the AI is incorrect, which also makes sense to me as if you're relying on AI, then you're going to be more confident than you otherwise would be. They have on page 12 of the paper, they have an interesting cognitive affordance and trade-off of system three chart, where for system one, it's fast, it's human, it's more intuitive and associative. System two, it's also human, it's analytical and reflective and system three is artificial. It's algorithmic and statistical. if you were just thinking in terms of, you know, folks are surrendering their decision-making process to the AI, and it would lead to more negative performance over time, they found that given AI, cognitive surrender can reduce the negative performance effect of time pressure when AI is correct. Which to me, sounds like if you're under the deadline, for example, if they ship lots of tickets, you are probably much more likely to surrender your decision making to the AI. Cause given the time pressure, you are not taking the proper, feedback and proper reflection that you otherwise would with a PR Whereas incentives and feedback may partially attune the cognitive surrender. So they also talked about how the right incentives and feedback can prevent the user from this cognitive surrender behavior. Lastly, before we get to... doomed by how everyone's just going to surrender everything to AI going forward. They view the vulnerabilities of cognitive surrender to system three as a design and education challenge. So how can we properly support decision makers in using system three effectively while maintaining critical thinking and accountability when necessary? According to their research, Dan (39:19) You're too doomed. Shimin (39:39) the results are just like giving feedback and aligning incentives may help people engage system two when needed without diminishing the efficacy gains provided by system three. And I think this ties into a lot of what we already talked about, right? when you're blindly trusting a AI generated answer in an internal meta board, when should your system to thinking kick in to be like, Hey, would this leak data to folks that otherwise shouldn't have the data. They didn't exactly propose how they're going to do it, but I the conclusion is the same. Dan (40:13) I had some different takeaways from this too, which is like the thing that was interesting was that you said the time pressure. It's helpful, I guess, in the aspect of like, if you're, if the AI gets answers right, you get answers right more often. But like, you could also look at that through a negative lens too, which is that it sort of forces more cognitive surrender because you just literally don't have time not to. ⁓ Which I think there was also that HBR paper that they talked about, really the reason that we haven't seen sort of drastic productivity gains from. Shimin (40:36) Yes. Yes. Dan (40:46) AI right now, supposedly, is because it's basically so everyone has to do more and more frequently, which you would think would lead to more productivity, but I don't know. That's the part I find odd. But it's definitely increased the pace of things. I've felt that personally. Shimin (41:01) Yeah, that's phenomenal. like AI can help us do more than ever. And all developers are feeling burnout because you feel like you need to use AI all the time and four agents at the same time. Dan (41:11) I had six going the other day, for Gastown soon. ⁓ But the other thing that I wonder about with this too is like, seems like, and you're better at like the science-y side of these papers than I am, but like in their study design, they consider the interplay between the two at all? I don't think they did, right? It looks like it was pretty much always like a question answer and then they like time boxed it. Shimin (41:15) Congrats. Yeah, that was my understanding. Dan (41:36) feel like there's also sort of like a fourth way, which is like you're combining system to, mean, I guess that's sort of what they're saying is they combine system two and three thinking, but like, that's what I find myself frequently doing. Right. I'm poking around an idea and then using an LLM as like a thought partner almost. Shimin (41:54) Mm-hmm. Yeah. And even system one comes in, right? Like AI can spit out a bunch of code and you're good just as this looks wrong. Dan (42:00) Yeah, that's true. Or the approach is wrong or yeah, like, yep. But yeah, I'm interested in that interplay then like whether or not that changes things a little bit. the other takeaway I thought was kind of interesting the higher trust in AI of the individual predicted more surrender. and the people who enjoy like effortful, like long running system to thinking tended to be more resistant to the faulty AIs. Promises. So the other, the other thing you cover that already to like they they intentionally injected error variance in the AI using I forget what it was. It was a See if I can find it. In any case, they basically injected the error count so they could control it to some degree and made them more error prone as part of the study. So that was when they say error prone, that's really what they mean. It's like the intentionally buggy error. If you thought Claude was bad, try Claude Bug. Shimin (42:59) Right. So this kind of goes back to, my side project that I was showing the other week, Flatter Proof right? we need to have a healthy activation of system two when you're using AI. And that is not something that folks are born with. Um, especially with all the marketing, pro AI buzz going on. Dan (43:22) Well, not everyone's necessarily fully a system two or one thinker either. Right. Like I recall, but a team building exercise I did one time with the team where it was rate, basically there's all these lines and you plot where you think you are on the line. And then it's interesting to see where your team falls. And one of the ones that was on the lines was fast versus slow thinking. So it was continuum, but like people rated, they know self assigned and everything else, but Shimin (43:41) Mm. Nope. Rahul Yadav (43:44) you Dan (43:46) It's definitely interesting to see how people felt and how that impacted how they prefer to work. Shimin (43:51) No, I, I've yet to, I've yet found a developer who was mostly a system one, developer, I think. Are you, are you, would you, you will say you're mostly the system one developer? Interesting. Yeah, that's not. Dan (44:02) I mostly fast in Korea. Shimin (44:07) Yeah, it's not been showing my experience. Dan (44:07) I think it gets me in trouble, but yeah, I tend to make quick intuitive decisions and then I'm open to new information. Shimin (44:15) Interesting. Yeah. Rahul, what about you? Are you a system two or system one guy? Rahul Yadav (44:19) Mostly two. It depends on, guess, the, you know, how much you've done pattern recognition, right? How much of that same or similar problem you've solved in the past. then if you've seen them before system one, if you haven't. Yeah, Dan is older than America. Dan (44:32) I am old as the hills, if that's what you're trying to say Raul. Rahul Yadav (44:38) which is coming up on 250 years this year. Dan (44:42) Fire software developer. Rahul Yadav (44:43) I don't remember, it's been a long time since I read the Thinking Fast and Slow. After reading this, I was wondering, when do you decide to engage system two? Because that's the interesting question here, And the... I don't remember if they talk about it in the book. We just naturally do it based on the consequences of the decision. it made me think of the, you know, Jeff Bezos has that one way or two way door of making decisions. And so part of it is just going to be if it's a very low stakes decision, you might just put more faith in the AI versus if it's a, you know, it's a one way door. Shimin (45:13) Mm-hmm. Rahul Yadav (45:26) you would want to deliberately engage system two because there is, it's very hard to go back or even impossible. ⁓ Dan (45:35) The other that's interesting about that too, if you can scroll back up a little bit for folks watching on the video, but I'll speak to it as well for the podcast. So they had the table that Shimin mentioned earlier. it has system one, system two and system three across the top. then along the side we have origin, which is like human, computer, processing speed, cognitive effort, accuracy, blah, blah. The one I found interesting was justification. and that's like, you know, what, I guess like what you're using to rationalize the decision or the information you had, right? So like system one is experiential or post-hoc. system two is rationalized and articulated. And then system three is data-driven externally generated. Shimin (46:01) Hmm. Dan (46:20) And one of things that they like talked about a lot in the article was like the sort of like confidence inflation thing, right? People that had access to AI, their confidence ratings on their own answers, even though over half of them were wrong, were boosted by 12 percentage points. And you can look at that and the data driven and externally generated is clearly to me, at least the reason why. Shimin (46:26) Mm-hmm. Dan (46:41) because like we're sort of trained to think that external things are right, like Wikipedia or something like that. You know what I mean? Like maybe that's a bad example, encyclopedia, it's like everyone's Yeah. Everyone's grown up since they were, you know, not everyone, certain generations of folks have grown up referencing these, you know, tomes of knowledge that you can treat as like more or less accurate, even though there's Shimin (46:46) Mm-hmm. Rahul Yadav (46:51) GROKOPEDIA. Shimin (46:53) Mm-hmm. man. Dan (47:04) you know, problems with every piece of information out there. and the way that, I'm right is I think trigger triggers that as an instinct, right? Whereas like, I wonder what would happen if you did the same thing, but basically like the exact same experiment, but then change the prompt to be you're having a conversation and you're not a hundred percent sure about your answers. So, you know, Shimin (47:10) Mm. Dan (47:27) make sure you qualify things that you don't know really well or that you weren't able to pull out of memory. And then the responses might've been more like human in the sense of if you asked me something, right? I'm gonna be like, well, I think it's this, but man, I'm old as the hills and I don't always remember stuff. yeah. Shimin (47:45) Right. And in a way there is a little bit of that in general psychology, folks who sound supremely confident has a higher ability to influence others. So I wonder if that is either what, reinforcement learning with human feedback trains into the model. Dan (47:56) Mm-hmm. Shimin (48:03) or it's just one of those shortcomings of the model where they're not very good at sounding, not confident, you know. Rahul Yadav (48:10) By the way, that meta article we were talking about earlier and cognitive surrender by just like following the system. maybe that's system three played some part in that. Dan (48:22) and the person just blindly, yeah. Rahul Yadav (48:25) Yeah, especially if you're you know that plus like if you're moving under a lot of time pressure or something. It's very possible that you know. Dan (48:33) I have definitely seen examples of that in my own experience of people responding to someone else's question with an LLM generated Answer what they clearly took with the person asked and pasted it into an LLM and applied zero critical thinking to it. Yeah. So, you could be right. It does happen. And I would like to charitably say that that's a time crunch driven response, but Rahul Yadav (48:41) I don't know. Shimin (48:47) Ouch. Yeah. Be careful. yeah. Rahul Yadav (49:00) I had a few other things to add to this. They said that offering per item financial, per item incentives still didn't eliminate the pattern of surrender. And so like, even when it's paid, when you're paid to be right or your incentive has to be right, people still chose to defer to AI. to a certain extent, which was interesting. It's almost like, you know, don't make me spend the mental energy. And closely tied to that is the... One of the things they talk about is the need for cognition. And to simplify, not everybody likes to think for themselves. And if you give someone who doesn't like to think for themselves a faulty, unreliable companion like AI, you might widen the gap between someone who likes to think for themselves and using that as a tool for research and then making even better decisions versus someone who's already not a big fan of thinking and then delegating even more thinking to, you know, a flawed AI. So it's gonna widen that cognitive gap there. Shimin (49:55) Mm. Yeah, that's tremendous point. it's almost like AI will have this societal impact of making the richer, the rich richer and the poor poorer I wonder if they should come up with a, hey, you've followed my suggestion the last 10 times without pushing back. Maybe you want to give it a shot kind of a deal. Rahul Yadav (50:25) Yeah. And then what else was it? yeah, the whole like time pressure thing and all that. Our world is moving faster. And I in the back of my mind, maybe like many other Americans have the midterms coming up. And Shimin (50:36) Mm-hmm. Rahul Yadav (50:45) trusting AI, which will be heavily used in these elections for sure, ⁓ is going to be a big pain, especially under the world seems to be moving faster. So I can see a, you know, a combination of these things. Or if you don't like to think for yourself, the world is moving much faster, you're under a lot of time pressure because of the economy jobs blah, blah, blah, things could get crazy. Dan (50:51) Already is. Rahul Yadav (51:10) in certain directions. And one final thing before we move on. The Tversky guy had influenced Kahneman's book, but he wasn't an author on the book because he had died before it. So I would like to correct my false statement, unlike some AI out there who speak confidently. Dan (51:11) Who should I vote for, Rahul slash Claude? Shimin (51:16) Well, Charlie. Thank you for looking that up. Rahul Yadav (51:37) They were long time collaborators and anyways, yeah. Shimin (51:41) All right, well, with that correction, that live correction, let's move on to my favorite segment of the show. Are you ready? Rahul Yadav (51:47) I don't want someone writing to you and be like, this guy is also lying while you guys are talking about with confidence. No, I had confidence built into me before, so now I'm just in crazy land. Dan (51:52) We'll just assume you're using an LLM. You Shimin (52:02) You just sound extra confident all the time. This is how you make it in life. Alright, let's get to Dan's rant. Dan (52:06) So, so this is an unusual Rahul Yadav (52:06) Dan's Rant Dan (52:11) Dan's rant because it's a rant specifically. Well, it's not even about this article, but it's about an article in general, or at least the concepts that are talked about in it. So tech crunch on 21st of March here ran an article. was called our AI tokens, the new signing bonus or just the cost of doing business. like that question. I, the fact that we're even posing that question is wild to me. and so the, the, I'll just quote the section of the article and then I can rant about it. So the idea is straightforward enough. Rather than giving engineers only salary equity and bonuses, companies should also hand them a budget of AI tokens. The computational units that power tools like Claude, chat, JPT and Gemini spend them to run agents, automate tasks, crank through code. pitches that access to more compute makes engineers more productive and that more productive engineers are worth more. It's an investment in the person holding them is the idea. And I'm just like, what? But how is that a benefit? How is that a perk for the employee? Rahul Yadav (53:12) If you look through a very fine hole, Dan, these are all perks. None of these are meant to... I'm with you. I'm joking. Dan (53:18) I mean, look, if I had two offers that were identical in every respect, except one was like you get a Claude, like $20 subscription and the other one is you get a Claude $200 subscription or something and they were both really cool. All other things were completely equal. Sure, I could see that maybe, but like. Rahul Yadav (53:35) Yeah. Dan (53:42) As a, as a signing bonus, I don't like that. The incentives are just drastically incorrect here. Like that doesn't, that's not a benefit to me. It's a benefit to the employer because what they're paying for is my, reasoning ability and like just general, you know, thinking and instructing computer what to do in the first place, whether that's writing code or prompting an LLM is kind of irrelevant. Rahul Yadav (53:51) Yeah. Shimin (53:52) Mm-hmm. Dan (54:06) Other than like pretty possibly velocity, right? So like if you're hiring me for that ability, what does the token budget give me? It gives you something about my output, but not me. I don't know. I just find I found this whole idea to be like incredibly offensive. And. Just wild. Which I mean, guess that, you know, that's why I was like, the leaderboard thing will come back with this, right? Cause like there are companies that are doing leaderboards around spend, which is like, I understand that through the lens of you want to like, nobody's really sure how to pitch this at an organization. Right. And I think we're seeing sort of native adoption through people like ourselves that are curious about it. And then they start poking around it and then they realize the potential and you know, really start going with it. And then other people see what they're doing and it kind of spread a little bit virally. And I think there's another group of people that were kind of like more skeptical. And then the models, you know, the last two sets of models really have like truly been a step change and now they're more on board. And, know, there's a couple different like narrative threads that are going here, but like, Token tokens as a bonus for me. Can I use them for whatever I want? Like, I like maybe if that was true for like, here's a bucket that you can use outside of work. Like, that's a bonus for me. Right. You know, in the way that like I can take a cash bonus and spend it on whatever. But I just that that's the part that I really get is like, how is that an incentive to the I don't know. Tell me more about the narrow lens. Like what? Shimin (55:15) Mm-hmm Rahul Yadav (55:30) I see these as tools and you should want to give the best tools to all your employees. yeah. Dan (55:39) I agree, but do, you go to like, okay, so let, let's, let me pose the hypothetical to you, Raoul. So we've got company A and they're going to offer you, I don't know, X salary and Y, you know, options or stock or whatever. And, then there's company B and they offer you X and Y exactly the same, right? Company A gives you an old busted windows machine. Rahul Yadav (55:47) Yeah. Hmm. Mm-hmm. Dan (56:06) and company B gives you the latest MacBook Pro. Are you going to choose like purely because of that? Rahul Yadav (56:12) No, it depends on like what the company is doing and all that. Yeah. Dan (56:15) Right. That's what I'm saying. It's like, how is that a benefit to you? I mean, certainly I've got preferences around operating systems and blah, blah, blah, but like. Rahul Yadav (56:20) Yeah. Sure, but depending on X or Y, I can buy my own machine, even if they give me a busted Windows one. That's not a big problem. Yeah. Yeah. Go ahead, son. Dan (56:31) Yeah. I mean, that's what I'm saying. how. ⁓ Shimin (56:37) Yeah, to take To take another analogy, like I will certainly prefer if the company I'm working for has legacy codebase in Ruby over PHP. But you're not going to gaslight me into thinking that the Ruby codebase is a part of my compensation package, right? Like that's not how this works. Give me United States dollar or NFTs or stable coins or anything. Rahul Yadav (56:48) Yep. Yeah. Dan (56:58) You Shimin (57:05) that I could use to exchange for tokens. That's what the economy is about. Dan (57:09) Yeah, just, don't like the, yeah, the value set up here is not nonsensical to me. It doesn't make any sense for posting this article. So apologies to the author, but I'm just like, what? And then I've seen it kind of, you know, bandying about in a couple of places and I'm just like, you've got to be kidding me. I, anyway, it was a rant, pure and simple. Rahul Yadav (57:27) What was the... Oh yeah, Jensen Huang had that, right? Shimin (57:35) Mm-hmm. Mm-hmm. Rahul Yadav (57:36) that engineers should receive roughly half their base salary again in tokens. Dan (57:41) Yeah, but him saying that is he means it in the sense of like, that's not it. He didn't say that in the sense of like, you need to, do that as a, that's not compensation to the engineer. he did. He said it's a recruiting tool. Good Lord. I thought that was just like using it as an approximation for token budget, but no. Rahul Yadav (57:56) Yeah. You know. Why might someone who has nothing to do with tokens or the AI industry be pitching ideas like these, spending 250k per person on token consumption? I have no idea. Dan (58:12) Mm-hmm. Shimin (58:14) Yeah, wild times. And I have a feeling, uh, Jensen will be making a comeback. uh, let's, let's move on then to our two minutes to midnight segment where every week we discuss where we are on the AI bubble using the analogy from the atomic clock, from the bulletin of atomic scientists. Um, I believe we are at a minute and 45 seconds still. ⁓ from last week where we didn't change anything. So let's see how we feel this week. Dan (58:41) Yep. liking the trend. No changes. yeah. So speaking of Jensen, he, he made a shocking announcement in addition to the token announcement that I was just, upset about, but he also claims that, or doesn't claim he, he put the projected sales for Blackwell and, I guess Vera Rubin, into the $1 trillion range. Uh, which is pretty crazy. mean, it's like, you can't get more hype than that. I don't think so. They saw quote unquote 500 billion demand, um, for blackwell chips through, through 2026. But the, the thing about it that gets me a little bit too sorry, I'm still kind of in rant mode. I can't shut it off that easy. Uh, is that like, Rahul Yadav (59:32) Hahaha Dan (59:35) Demand isn't the same thing as sold because the bottleneck here is and always has been like TSMC's ability to like ship this stuff, you know, since they're fabulous. So I don't know. It's like, it's cool to say that you've got this demand, but that demand is also like on paper. and isn't necessarily going to result in a... Well, mean, clearly it will result in revenue for them. But like, if you can't meet the demand, like, I don't know. Shimin (59:59) Yeah, I'm trying to think if we had analogies from the dot com bubble for these kind of just bananas numbers. it's like, it's like Dr. Evil, you know, $1 billion. but. Rahul Yadav (1:00:09) The, it is gonna be, cause these are the ones that are optimized for inference versus training, right? And I think this kind of fits into the, there's been a narrative that training plateaued sometime late last year or whatever. And cause we haven't seen a major, you know, a, Shimin (1:00:09) one trillion. Mm-hmm. Rahul Yadav (1:00:30) the Claude 5 or GPT-6, we've been seeing minor improvements and optimizations. And I think the plan here is in actually deploying all these models and driving down inference costs is where the next big opportunity is and where the money is going to come from. I think that's what he's aligning with and betting on. So from that angle, it makes sense to me how much it sells, what not. I don't know. You got to keep people excited. Shimin (1:00:48) Yes. Dan (1:01:02) Yeah, I'm not saying that the demand doesn't align with things that are happening. It's just, it's a pretty wild number to throw out there when you're throwing it out there and knowing damn well that you can't produce a trillion dollars worth of chips. Rahul Yadav (1:01:15) Yeah. Shimin (1:01:16) Yes. ⁓ Rahul Yadav (1:01:16) Maybe the tariff ad, did tariff ad not make it to the links this week? Maybe the tariff ad will start cranking them out tomorrow. Dan (1:01:26) Maybe. Shimin (1:01:28) Rahul, your link this week is accelerated FOMO in the age of AI. Rahul Yadav (1:01:33) yeah. So this is from Sid and I do not know their last name, but if you go to their homepage, it says also known as Tony S Shrug. We're off to a great start there already. I love Tony S already. But you know, this is their rant on Shimin (1:01:44) Ha ha ha. Rahul Yadav (1:01:56) There's this very FOMO driven narrative that's happening where it's like, you have to be burning X amount of tokens. If you're not running N number of agents in your sleep, you're not gonna make it. We're seeing all these like, it should be part of your compensation. There was this like, Tyler Cowan had this logic thing of... If you think this is the last big opportunity to make money, you should be working as hard as possible. And if it's not, it will significantly improve your capability and you should be working hard to learn something along those lines. And so there's a lot of this, like, this might be the now or never time and you don't want to miss out on this. Shimin (1:02:42) Mm-hmm. Rahul Yadav (1:02:45) than. Siv slash Sonia gets into, you know, what is the, who's pushing these things? Tyler cowen to his is not pushing any of these. It was just interesting that he had posted that. But back to this post, people were trying to sell you courses, were trying to sell you the hustle culture of, you know, how less you sleep, how many agents you run. the kryptonite to all of that bragging is, what actually have you built or shipped that has actually had some value at the end of the day? And that mostly ends the conversation right there. The second one is sell courses, which goes back to our $1,000 a month prompt pack business. It's one way to sell courses. And then the last... Shimin (1:03:31) Mm-hmm. Rahul Yadav (1:03:36) couple sentences in this section were interesting of or lesson is the new adage about like gold rushes and show should be during a gold rush sell courses on how to sell shovels He's laughing. Dan (1:03:48) That was, I said, speaking of, we're happy to announce our newest venture on the Shimin (1:03:53) HAHAHAHA Rahul Yadav (1:03:56) And then, know, just like some people who you people who work at like the big labs and who work at these AI companies that one. He goes about like your, you know, paychecks and RSU depend directly on AI. So it is predictable that you'll do these things. And just from like, you know, when I was reading that from a cognitive dissonance perspective, it would be very hard to live with the fact that you don't believe in something, but you're spending your day in and day out on that. so, you can try very hard, but at some point the brain is just going to take over and be like, we believe in this, you know, go back to crypto NFT and knowledge. Shimin (1:04:23) Mm-hmm. Yep. Rahul Yadav (1:04:34) or any of them. Sometimes your brain just makes you believe in things for reasons you can't really control. So anyways, their call is, know, this technology is much more useful than... Sid draws, you know, calls back to the NFT era where people were just running crazy for some, you know, some online media and talk so I'd like maybe we can tone the toxic cheerleading down a little bit. So hit all the right notes. The courses thing made me laugh because we should definitely be doing that. Sell courses on sell it. Dan (1:05:11) no. Rahul Yadav (1:05:12) Yeah. Shimin (1:05:12) ⁓ exclusive membership club for the podcast. course. Yeah. One on one coaching from Dan and his. Dan (1:05:20) In other news, this will be my last episode of the show. Even though I suggested it. Shimin (1:05:23) Hahaha Rahul Yadav (1:05:23) But, so, honest, you, you, I don't know if you guys see this. I end up seeing this on sub stack a decent amount because a lot of these links we end up picking up and through like other reading sub stack is you know much more prominent now, but every other poster every third post is like my setup did XYZ my other thing that did this XYZ You're not gonna make it if you don't have these other things by the way I have a product to sell or I'm doing a class that dude does these and it's just like it's all if you just talk to the AI log Shimin (1:05:39) Mm-hmm. Rahul Yadav (1:05:58) enough and know what you want, it'll help you get that. There's nothing actually special here, but the core sellers are there at large and they're selling hard. Shimin (1:06:07) what they're taking the prompt pack advice seriously, they're implementing it on their sub stack is what they're doing. Rahul Yadav (1:06:12) Yeah. Dan (1:06:14) I mean, what else are you gonna do with those bonus tokens you got when you got hired? You know. Rahul Yadav (1:06:17) Hahaha! you Shimin (1:06:19) Yeah, so we have NIVIDIA with trillion dollars demand. have what's clear to be bubbles around the marketing aspect of AI. I have a little bit of a downer this week, which is OpenAI shutters Sora, its video generator, after just six months. So I'm not sure we covered this, but OpenAI created the Sora app. social media via generative AI app. It was a huge deal when it first came out. It was number one on the app store and they almost inked the deal. Well, they did ink the deal with Disney for Disney to provide them with a billion dollar investment and allow OpenAI to use some of the Disney characters like the Marvel superheroes and Mickey Mouse and all that. And today it came out that OpenAI is shutting the door on Sora, not just the app, but also its API service. Sounds like they're exiting the AI video generation market entirely, which I think is either because it's too expensive or it's because they want to focus everything on their current competition with Anthropic in the enterprise space. ⁓ Rahul Yadav (1:07:29) Yeah. Shimin (1:07:29) But even if the second hypothesis is true, it does not make much sense to me that they would, you know, shut down one of their rather flagship apps. So I don't take that as a good sign. Rahul Yadav (1:07:41) They, yeah, I think they're trying to, this isn't just the year, they're trying to also like focus hard on the IPO. And so you have to, anything that's not gonna look good on the books, my guess is all gonna be, you know, paused, shut down. Dan (1:07:41) Sounds like trouble. chopped. Yeah. Rahul Yadav (1:07:59) And yeah, and they're trying to consolidate everything into, I think, chat, GPT, codecs, and I forget whatever the third thing was, a bunch of these get combined into a super app to like become similar to what Claude Cowork is trying to do. So it seems like with that, this is consistent and yeah, just another step towards prepping for IPO from what I can tell. Shimin (1:07:59) Mm. Dan (1:08:22) And it also like it says they're shutting down API, but did it mention if it's getting shut down in like chat GPT because there's integration there too right I think similar to like how Gemini can do. Rahul Yadav (1:08:24) Thanks Shimin (1:08:33) Yeah, I'm not sure about that. It's possible that the API is getting shut down because they aren't actually making any money via the API. It may just be really expensive to run. Dan (1:08:40) Yeah. Rahul Yadav (1:08:42) There, I also wonder if SeeDance and like, you know, how crazy that was also made them be like, we don't need to do this. And yeah, like let's just focus on the other things because they've been spread so thin between all these things. They haven't done anything well. They've done everything. Shimin (1:08:51) They've clearly lost the battle on that one, yeah. Yeah. More CCP propaganda for me. I'm here for it. Rahul Yadav (1:09:07) Yeah, by the way, I looked this up. According to Gemini, so I'll start with that considering our paper about system three. OpenAI had announced the partnership with Disney on December 11 for last year. So we're about three months and two weeks. Shimin (1:09:25) Mm-hmm. Rahul Yadav (1:09:25) later, I think, or four months. And that three-year partnership ended in about four months. So just saying keep an eye out on those partnership announcements, all these other future things. That's the reason why Dan doesn't try Codex is what I'm trying to say. Maybe by the time he signs up, it won't be a thing. Shimin (1:09:36) Yeah, it's true. It just- Dan (1:09:37) Mm-hmm. no, I better go fast. Rahul Yadav (1:09:48) you Shimin (1:09:48) It doesn't make much sense to me given that, you know, a investment and integration and IP rights to some of the world's most valuable IPs is something that you would just shut down overnight. Like something feels iffy. Rahul Yadav (1:10:02) Yeah, yeah. Shimin (1:10:03) Okay, all that said, how do we feel about the clock this week? Dan (1:10:07) Honestly, this last one makes me want to push it closer. Because I mean, you might be right, Rahul, that it's a, you know, divesting of non profitable things, but like. Rahul Yadav (1:10:17) Hmm? Dan (1:10:17) It's just such a bad plus advertising and all the other things they've been doing have been like such a bad look, at least from the outside looking in. I'm just like, yeah. Is the government contract enough to send maybe an IPO enough to save them? Rahul Yadav (1:10:31) Is there a government contract? We haven't followed them. Dan (1:10:34) Oh yeah, that's true. I don't know. I mean, we talked about it and then yeah, that thread is kind of drifted a little bit on Threadmill. Rahul Yadav (1:10:40) you Shimin (1:10:42) An IP, the IPO might be when this whole house of cards come down. Assuming this is a house of cards. Rahul Yadav (1:10:49) Yeah. Dan (1:10:49) I mean, I guess I'll float a number. would say I would go as low as a minute. Rahul Yadav (1:10:54) Whoa. Shimin (1:10:54) wow, aggressive. I was gonna go like a minute 15. I can do like a minute 10. Rahul Yadav (1:11:00) Whatever you guys pick. Dan (1:11:03) We could do 15. Shimin (1:11:04) Minute 15? Or yeah, okay. Minute 15 it is. All right, moving the clock. Dan (1:11:05) Yeah. ⁓ Relatedly, but didn't make the cut on these, there was also an interesting article where they went into the economics of building a data center in space. And unsurprisingly, at least to me, the economics are largely driven by the costs per ton of lift vehicles, right? And in order for a data center to be viable in space, Rahul Yadav (1:11:27) What? Shimin (1:11:28) Mm-hmm. Dan (1:11:32) cost per ton needs to be under $2,000, I think, or something like that. Maybe it was in, I assume it was in thousands. And currently it's, I think, double or triple that. okay. Waiting for that space data center. Yeah. Which sure has no relationship to that announcement. Shimin (1:11:42) Well. Rahul Yadav (1:11:42) Those starships are gonna take care of that. Shimin (1:11:45) Yeah, that's what I was going to say. They're going IPO this year too, aren't they? Rahul Yadav (1:11:48) By the way, on the... Yeah, and the data centers in space thing, even if it doesn't work out, there are some like, technical problems you'll have to solve that would be good on Earth as well. Like heat exchange, think is a big challenge. And if you can solve that, would benefit, you know, us here on Earth as well. So hopefully some good comes out of that, even if it doesn't ever end up in space. Shimin (1:12:15) Yeah. Well, we'll have to. ⁓ Rahul Yadav (1:12:17) And then the people were like, where's my water? And the data center is drinking all my water. Maybe we'll calm them down finally. Shimin (1:12:25) We'll save that for our space podcast called Artificial Space Nerds. We don't know they're talking about. Minute 15 it is. And that is a wrap. Thank you for joining us today for our chat. And if you like the show, if you learned something new, please share the show with a friend. You can also leave us a review on Apple Podcasts or Spotify. It helps people to discover the show and we really appreciate it. Rahul Yadav (1:12:29) Go get it. Dan (1:12:32) You Shimin (1:12:48) If you have a segment idea, a question for us or a topic you want us to cover, show us an email at humans at adipot.ai. We love to hear from you. You can find the full show notes, transcripts and Dan's CLI tool mentioned at www.adipot.ai. Thank you again for listening and we'll catch you next week. Bye. Dan (1:13:00) God, don't use it, don't use it. Just please. Tschüss. Rahul Yadav (1:13:08) See you folks.