Why the Tech World Is Going Crazy for Claude Code | Odd Lots (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) By the way, I don't have AI psychosis. I (00:00:02) have a Claude complex. (00:00:04) >> Why is everyone making that joke? (00:00:06) >> Wait, which joke? (00:00:07) >> The psychosis joke. (00:00:09) >> I thought you were gonna be proud of me (00:00:10) for saying claude complex. (00:00:12) >> Oh. Oh, that is very good. (00:00:13) >> I It's like I do one pun finally for (00:00:15) Tracy and she's just like, "Why is it (00:00:17) ever making that joke?" (00:00:18) >> Well, I was thinking about I was handing (00:00:20) you a I finally make a pun and you just (00:00:22) jump right over it. (00:00:22) >> Well, everyone keeps saying that Claude (00:00:24) code is AI psychosis for smart people, (00:00:27) right? Like, how did that become a (00:00:28) thing? Very addictive. (00:00:29) >> Yeah. All right. (00:00:30) >> But there's a good pun. (00:00:31) >> It's also very brocoded, I find. (00:00:39) >> Hello and welcome to another episode of (00:00:41) the OddLots podcast. I'm Joe Weisenthal. (00:00:44) >> And I'm Tracy Aloway. (00:00:45) >> So Tracy, you're cool. Like if I like, (00:00:48) you know, just start doing this (00:00:50) part-time as I like build out my (00:00:51) software business, right? Like you're (00:00:53) cool about that, right? I was going to (00:00:54) say I've been thinking about AI and (00:00:56) productivity and so far your (00:00:59) productivity has gone down Joe because (00:01:01) instead of doing OddLots things you're (00:01:03) coding your own software (00:01:05) >> except that I'm creating content for the (00:01:07) OddLots newsletter about coding and that (00:01:09) is productivity um accreative (00:01:12) >> debatable (00:01:13) >> debatable but but you're cool with that (00:01:16) you're cool with like me like oh I'm (00:01:17) just going to like check in part time on (00:01:19) odd lots when we have a recording well (00:01:20) like (00:01:20) >> oh of course not good of course not good (00:01:22) that's the right answer. I want you to I (00:01:24) want you to be really sad. But like a (00:01:26) few other people, you know, I have like (00:01:28) caught the sort of like bug of like AI (00:01:30) coding and I'm totally blown away. I've (00:01:33) like played with it from the be I (00:01:35) started playing around with it last (00:01:36) year, but then over the holidays and (00:01:38) I've been writing about this in the (00:01:39) newsletter. Suddenly like my Twitter (00:01:40) feed is like cloud code, cloud code, (00:01:42) cloud code. I used cursor before (00:01:45) >> which I was very impressed by at the (00:01:46) time. And um so when I got home from (00:01:49) vacation, one of the first things I did (00:01:51) is like figure out how to install Cloud (00:01:52) Code on my computer and I was like, "Oh, (00:01:56) I am like hooked." And this is actually (00:01:58) like I I see why half my Twitter feed is (00:02:01) just like people posting about this. (00:02:03) >> All right, so I have to say I have not (00:02:05) tried it because I only have a work (00:02:07) computer and I can't install new (00:02:09) software. Um and I probably definitely (00:02:11) cannot install new software that then (00:02:13) makes changes to existing software. I (00:02:16) don't think Bloomberg would like that. (00:02:17) Um, but I have seen the hype. Lots of (00:02:20) people talking about it. Have you seen (00:02:23) um Claude Co.? Have you heard of that? (00:02:26) >> Oh, yeah. Yeah. Yeah. So, one of the (00:02:27) criticisms of Claude Code was that, you (00:02:29) know, like, okay, you could code, but (00:02:31) you still need some background knowledge (00:02:33) in coding cuz like, you know, the (00:02:35) interface is kind of like 1980s and (00:02:38) >> and all of that or 1990s. Um, co-work (00:02:41) apparently like goes a step further for (00:02:44) for normal people in coding and makes it (00:02:47) super super easy. And the funniest thing (00:02:50) is that apparently Claude code actually (00:02:53) coded co-work. So the so this is like (00:02:56) really relates to my experience last (00:02:58) year and then this year which is that (00:03:00) even last year like trying to use the AI (00:03:02) coding tools. It was an annoying process (00:03:06) because there are various things that (00:03:08) you had to do in the actual command line (00:03:10) of the computer (00:03:11) >> that were like I didn't I don't know (00:03:12) command line vernacular and you have to (00:03:14) like install these libraries and stuff. (00:03:16) So there was this sort of like barrier (00:03:19) that existed and but what's what's (00:03:22) really changed in the last year or with (00:03:25) the with claude code which has actually (00:03:27) been around for a while and I should (00:03:28) have like played with it before is that (00:03:30) like because it sits on your computer (00:03:33) it's sort of takes away it dects it and (00:03:36) so when you talk about like does the (00:03:38) stuff (00:03:38) >> it does it it just like oh it's like oh (00:03:40) we're going to need to install this open (00:03:42) source natural language processing (00:03:43) library it just does it automatically (00:03:45) instead of me try and like figure out (00:03:46) like what are the right keystrokes to (00:03:48) pull that in or why is this not going (00:03:50) into the right file folder or whatever. (00:03:52) And so like what like co-work it's like (00:03:55) all like all of these sort of like (00:03:56) little frictions like these technical (00:03:58) things like command line user very (00:04:01) rapidly are like dissipating. Yeah. (00:04:03) >> And so that like then you have something (00:04:05) like co-work like they know they're (00:04:07) taking care of that and so you get this (00:04:09) like user interface that's just like (00:04:11) it's just getting easier and friendlier. (00:04:12) There's almost no technical frictions at (00:04:15) all anymore. (00:04:16) >> Also, it feels very iterative like the (00:04:18) code is improving upon itself at this (00:04:20) point and I think that was one of (00:04:21) Claude's main selling points. Well, this (00:04:23) is like you've seen like people talk (00:04:25) about like oh is AGI here and this is (00:04:28) like part of the debate because the prem (00:04:30) one of the ideas I guess behind AGI is (00:04:32) like well what happens when you have (00:04:34) software that can train itself and so (00:04:36) forth and I don't really know if I buy (00:04:38) that but you do just see like how fast (00:04:40) the iteration cycles are and I think we (00:04:42) uh want to get into this be in part (00:04:44) they're fast because a bunch of people (00:04:46) are suddenly getting excited so then the (00:04:47) human provides this sort of like (00:04:49) >> we're sewing the seeds of our own demise (00:04:51) because we're so enthusiastically (00:04:53) participating in the evolution, but I (00:04:55) just like it's suddenly clear like oh (00:04:58) this is going to change I think (00:05:00) computing and the other thing is the (00:05:02) code works like it creates code that (00:05:04) like this is like there's no bugs you (00:05:06) know it works. (00:05:08) >> Um did you see speaking of automating (00:05:10) yourself did you see there was a post on (00:05:11) Reddit from a lawyer who said he's (00:05:13) basically used claude code to automate (00:05:15) like his entire job and he hasn't told (00:05:17) anyone. I'm not exactly surprised (00:05:19) because the other thing that I (00:05:20) experimented with (00:05:22) >> is and I haven't 100% verified this but (00:05:25) on jobs day last week (00:05:27) >> I downloaded the full PDF and I just (00:05:29) typed into the cloud code like find the (00:05:31) most interesting details and make some (00:05:33) charts based on and it did it in like a (00:05:35) couple minutes. I have no like ability (00:05:37) to like I've never like built charts (00:05:39) myself by hand or whatever or like (00:05:41) design it or whatever. And I didn't (00:05:42) totally confirm yet that the data was (00:05:44) all correct but I'm pretty sure it was (00:05:46) cuz everything I spot checked. So I (00:05:47) didn't (00:05:48) >> just that crucial detail. (00:05:49) >> Yeah, I know. I didn't that's why I (00:05:50) didn't want to like oh like here's what (00:05:51) here's the today's jobs report and (00:05:53) charts but mo my (00:05:55) >> but what application did it actually (00:05:56) build it in the charts? (00:05:58) >> I don't know. I just had a file like (00:06:00) that's the thing. I had a file on my (00:06:01) computer at that point. (00:06:02) >> What kind of file? (00:06:04) >> Like a PNG file like an image file. (00:06:06) That's the crazy thing. I don't know. (00:06:08) Um, and so there was just this image (00:06:09) that had a bunch of charts and my spot (00:06:11) checks did suggest like I didn't see (00:06:13) anything off and people get paid money (00:06:16) to like build that kind of stuff for (00:06:17) like analysts and stuff like that and (00:06:19) >> right so this is the other big question (00:06:22) if everyone can build their own software (00:06:24) what actually happens to software and I (00:06:26) was reading something I forget who it (00:06:28) was by but someone used claude code to (00:06:30) create they wanted a website that would (00:06:32) basically make them money for doing (00:06:34) nothing and that was the prompt and (00:06:36) >> did they do it? (00:06:37) >> Yeah. So, the idea that um the model (00:06:40) came up with was you can sell prompts, (00:06:42) packages of good prompts and sell them (00:06:45) for like 40 bucks and you'll make tons (00:06:47) of money. (00:06:48) >> And I was thinking about that like, (00:06:49) okay, it's possible to make money that (00:06:51) way, but also why wouldn't I just use (00:06:54) clawed code to do the same thing? (00:06:57) >> Uh there are many big questions that we (00:06:59) as an economy are going to have to think (00:07:00) about and I think my main takeaway is (00:07:02) we're going to have to think about these (00:07:03) sooner rather than later. But what is (00:07:05) Cloud? Why is everyone so hyped about (00:07:06) it? Like what is it about this (00:07:08) particular piece of software that versus (00:07:10) what exists from OpenAI and Gemini and (00:07:13) all this stuff? Like why has this (00:07:14) captured everyone's imagination? We (00:07:16) really do have the perfect guys because (00:07:17) it's someone who unlike me has been (00:07:20) getting their hands dirty in this stuff (00:07:21) for longer. one of the few people that I (00:07:23) know who was into LLMs before chat GPT (00:07:26) existed and was actually using them via (00:07:29) the API and was actually talking about (00:07:31) their technical capacity to do things (00:07:33) like coding even before November of (00:07:36) 2022. So truly the perfect guest. We're (00:07:38) going to be speaking with Noah Brier. (00:07:39) He's the co-founder of Alfic, a (00:07:41) consultancy that helps big companies (00:07:43) deal with AI stuff. So, uh, Noah, thank (00:07:46) you so much for coming on OddLots. (00:07:47) >> Thank you for having me. (00:07:48) >> What's the deal? How were you like using (00:07:50) LLMs before ChatGpt existed? I don't (00:07:52) know. I I know very few people who were (00:07:54) doing that. (00:07:55) >> I had the good fortune of shutting down (00:07:57) a startup in 2022 and so I had a lot of (00:07:59) free time on my hands. (00:08:00) >> And then how are you using it though? (00:08:02) Like how did you like your like how did (00:08:04) you aware that there was this thing that (00:08:05) could be of potential use to you? What (00:08:07) were (00:08:08) >> So my very first thing I was doing was (00:08:09) using GitHub Copilot which at the time (00:08:11) was built into VS Code and it was (00:08:14) autocomplete inside VS Code. So it was a (00:08:16) nice and pretty immediately realized (00:08:18) that there were certain coding tasks (00:08:19) that it could just handle completely. Um (00:08:23) anything that was very patternbased. So (00:08:25) if you write code, you write a lot of (00:08:27) tests. If you write tests, every test (00:08:29) kind of follows the same pattern and you (00:08:31) want it to follow the same pattern. (00:08:32) You're looking for that structure and (00:08:33) and over time because it was looking at (00:08:35) your codebase, it was able to basically (00:08:36) autocomplete it. Uh, I also started (00:08:39) playing with the GPT3 API which had come (00:08:43) out. I think that came out in November (00:08:44) of 2021 (00:08:45) >> and that was the first time it was (00:08:47) publicly available to everybody and they (00:08:49) had a large language bundle as we know (00:08:51) it today (00:08:52) >> available to them. So I was just testing (00:08:55) and building things and I pretty (00:08:56) immediately (00:08:58) realized the very first thing I did (00:09:00) where it it just blew my mind was I (00:09:02) built a web scraper. So, I was I was (00:09:04) just trying to pull pricing data from a (00:09:05) website. And I've done a lot of this in (00:09:08) my career. It's maybe the most annoying (00:09:10) task you have to do in all of coding (00:09:11) because HTML is the most miserable (00:09:13) language to have to parse. And I just (00:09:16) had this thing where I took the page, I (00:09:17) took the content, I took the text, and I (00:09:19) gave it to the AI. And I asked it to (00:09:21) give me back the pricing table. And it (00:09:22) gave me back the pricing table. And I (00:09:24) just thought, I'll never do it the other (00:09:26) way again. That's it. (00:09:27) >> Yeah. That HTML mention just brought up (00:09:30) like memories of me in like the mid 90s (00:09:33) on HTML goodies. Do you remember that (00:09:35) site? Yeah. I wonder if it's still is it (00:09:38) still up? That would be wild. Um (00:09:41) >> does claude code does that count as AGI? (00:09:44) This seems to be the debate, right? Is (00:09:46) it AGI? (00:09:47) >> I try not to wait into what's AGI and (00:09:49) what's not. I think my guess on on AGI (00:09:51) for what it's worth is that it's (00:09:53) probably going to be a conversation like (00:09:55) the Turing test where everybody thought (00:09:56) it was really really important for a (00:09:58) really long time. We thought the touring (00:09:59) test was the biggest thing (00:10:01) >> for 70 years or whatever and then (00:10:05) >> CHBT very clearly passed the touring (00:10:07) test and now everybody pretends like (00:10:09) it's not just that they forgot they (00:10:10) pretend that it never mattered. (00:10:12) >> Oh, (00:10:13) >> and so I I am kind of guessing that (00:10:15) that's going to be what the conversation (00:10:17) is like. it's just going to be a sort of (00:10:18) forever moving goalpost. Um because it (00:10:20) turns out that the idea we had for what (00:10:22) general intelligence looks like is not (00:10:25) quite that. Um but I also think you know (00:10:28) the computer scientists and the sort of (00:10:30) serious AI researchers would say that (00:10:33) much of what's going on inside quad code (00:10:35) is not the model itself. It's the model (00:10:37) paired with a human. (00:10:39) >> And I I think that is a pretty important (00:10:41) distinction but I don't know about AGI. (00:10:43) >> Well, okay. So you were using um GPT to (00:10:47) code prior to the release of Jad GPT. So (00:10:50) therefore coding models have been around (00:10:52) a long time. So what is for those who (00:10:55) haven't played played around with it? (00:10:57) What is clawed code? Because again (00:10:59) coding models have been around for a (00:11:00) long time. They've c people maybe have (00:11:03) heard of cursor or copilot or some of (00:11:05) these other harnesses etc. What is (00:11:07) claude code? (00:11:08) >> So if we back up first and we go to (00:11:10) copilot. So copilot was the first sort (00:11:13) of commercial application of a large (00:11:15) language model (00:11:17) >> by most accounts and what copilot did in (00:11:19) its initial instantiation was just auto (00:11:22) >> it's a Microsoft product (00:11:23) >> it's a Microsoft product so Microsoft (00:11:24) owns GitHub GitHub developed copilot um (00:11:27) it was Microsoft had the partnership (00:11:29) with open AAI and so they they built it (00:11:31) in and what it was doing was doing (00:11:33) autocomplete so if you're writing code a (00:11:36) lot of writing code is boilerplate or (00:11:38) trying to remember the name of a (00:11:39) function and you know the reason Stack (00:11:41) Overflow existed was because you can (00:11:43) never remember the exact name of that (00:11:44) function or or the exact reax that you (00:11:48) need to use in order to find and replace (00:11:50) something. And so you would go search (00:11:52) for it and they realized that you could (00:11:54) just build that into the IDE uh your (00:11:57) code editor and and have it autocomplete (00:11:59) for you. And it it was pretty amazing. (00:12:01) Yeah. (00:12:02) >> Um then uh chat GPT came out and even (00:12:06) before that I had built a simple chatbot (00:12:08) for myself because I realized that hey I (00:12:09) could just ask this and instead of going (00:12:11) and searching stack overflow it was (00:12:13) totally capable of as answering code (00:12:15) questions and it was capable of writing (00:12:17) reax or doing these things and did it (00:12:19) make mistakes? Yes. But like there's (00:12:20) famous mistakes on stack overflow of (00:12:22) incorrect reax that now exists in every (00:12:25) codebase in the world. Um, and so, you (00:12:28) know, I there were a lot of us just kind (00:12:29) of playing with these things and and (00:12:31) realizing they were a huge boon. And so (00:12:34) I think really the next step is cursor (00:12:36) comes out and the thing cursor realized (00:12:37) that copilot didn't was that it wasn't (00:12:39) good enough to have autocomplete. You (00:12:41) also needed the Q&A because you have (00:12:42) these things that you can't just (00:12:44) autocomplete. You want to be able to ask (00:12:46) the question and answer it. And then (00:12:47) chatb came out and everybody was (00:12:48) switching between IDE. Um, and then I (00:12:51) think really the next big piece is that (00:12:54) Cloud Code came out. And what Cloud Code (00:12:56) did that was so remarkable was they took (00:12:58) the same set of models really um and (00:13:01) they took them out of the chatbot and (00:13:03) they really just gave it some very basic (00:13:06) functionality to operate within your (00:13:09) machine, right? And so you know if you (00:13:10) really look at kind of what exists (00:13:11) within cloud code, you're calling out to (00:13:13) a model and you they gave it capability (00:13:15) around sort of two big things. One is (00:13:17) you can read and write files on your (00:13:19) computer. Um, and then two is that you (00:13:22) can operate Unix, the the base commands, (00:13:25) the bash commands that exist in your (00:13:27) environment. Um, and again, because (00:13:30) these models were trained on the (00:13:32) internet, and there's no greater source (00:13:34) of information on the internet than how (00:13:35) to make the internet, they know how to (00:13:37) use (00:13:38) >> Unix commands incredibly well, right? (00:13:40) Because Unix has existed for whatever it (00:13:42) is, 60 years. And the way these commands (00:13:44) were designed, they're all designed to (00:13:45) be very, very simple. there's a find (00:13:47) command and you know there's thing (00:13:49) called GP and it can search through a (00:13:50) codebase and um Unix has this sort of (00:13:53) beautiful way of tying one command to (00:13:55) another so you can take the output of (00:13:57) one command and send it to another (00:13:59) >> and they kind of just gave the model (00:14:01) access to these two or three very simple (00:14:03) things and it kind of turned out that it (00:14:05) unlocked a whole bunch of functionality (00:14:07) that I don't think even the people who (00:14:09) built it fully realize like (00:14:11) >> one example that I I think about a lot (00:14:13) is just the challenge you have with all (00:14:16) of these AI models is that they're (00:14:18) stateless. So every time you talk to (00:14:21) chat GPT, it's sending your entire (00:14:24) conversation history back to chat GPT (00:14:26) because it has no saved history of that (00:14:29) chat, right? And that's fine. It's the (00:14:31) way it works. It's just fact. But it (00:14:33) means that you know it forgets things. (00:14:35) It doesn't know conversation to (00:14:36) conversation. And one very easy way to (00:14:39) save your state is just write it to a (00:14:41) file. Um, and so you give it write (00:14:45) access and it can create files. And now (00:14:47) all of a sudden you've overcome this (00:14:49) like probably the single biggest (00:14:51) challenge that exists inside these large (00:14:54) language bottles which is that they're (00:14:55) fundamentally stateless. (00:14:57) >> So Claude writes itself little like (00:14:59) memory notes, right, to to remember the (00:15:02) entire context of the conversation. And (00:15:04) that's how it solved that problem. (00:15:06) >> No. So there there's sort of two things (00:15:07) going on in clawed code beneath the (00:15:09) hood. There's one thing that works (00:15:11) exactly like chat GPT or any of these (00:15:12) other ones, which is it's maintaining a (00:15:15) conversation history. So every (00:15:16) >> message you send it and every um action (00:15:20) it takes, it's recording to a log, (00:15:22) right? (00:15:23) >> Um which is just one big file. That's (00:15:25) really no different than what chat GPT (00:15:27) can do. Where it gets really interesting (00:15:29) though is it can also write files that (00:15:32) it can then read. So whereas that (00:15:35) conversation history is all saved off (00:15:37) and eventually that conversation gets (00:15:39) too long and it needs to do a thing (00:15:40) called compaction and um when it (00:15:42) compacts it it tries to sort of just (00:15:44) remember the bits because there the (00:15:46) total window is is is (00:15:49) large but I mean it's like 100,000 (00:15:52) tokens. (00:15:53) >> That's what I mean by memory notes, (00:15:54) right? It compacts the information into (00:15:56) the important stuff that it then (00:15:58) retrieve. It does that. It only does (00:16:00) that at the end like once it runs out of (00:16:03) space. Okay. Once it runs out of context (00:16:04) window. So it has 200,000 tokens I (00:16:06) think. And 200,000 tokens in rough terms (00:16:09) is probably 150,000 words. (00:16:11) >> Um it says, "Okay, it's time for me to (00:16:14) compact all of this stuff." And so it (00:16:16) still saves your whole history on your (00:16:18) computer. you still have the entire (00:16:20) message, but for that session, it just (00:16:22) compacts it down to this, you know, (00:16:25) >> maybe 25,000 token memory of what it (00:16:30) was. (00:16:30) >> Yeah. (00:16:31) >> Um, and is this like something that was (00:16:33) not obvious before as a solution? (00:16:36) >> Like this compaction, how important is (00:16:38) it for this being like, okay, as a (00:16:42) human, I can work on this on a project (00:16:44) for a long time. Like how much of an an (00:16:46) unlock was that? (00:16:47) >> I'm not sure. Compaction was the unlock. (00:16:50) I think the compaction functionality is (00:16:52) >> helpful. (00:16:53) >> Okay. (00:16:53) >> The way ChatGpt does it for what it's (00:16:55) worth is they don't do compaction. They (00:16:56) just forget your messages eventually. So (00:16:58) if you're in one chat, eventually your (00:17:01) oldest message is going to fall off the (00:17:03) back. (00:17:04) >> For coding, that's probably less (00:17:07) helpful. (00:17:08) >> But there are trade-offs. I both (00:17:10) techniques work. I think fundamentally (00:17:12) the thing that is special about cloud (00:17:14) code is not the compaction. It's the (00:17:16) It's the ability to write and read files (00:17:19) on your computer. Yeah. Which means you (00:17:20) can always write off memories. And then (00:17:22) >> what does that mean, write off memories? (00:17:24) >> So you could say, "Hey, I it's really (00:17:26) important that I remember this thing for (00:17:27) future sessions. I want to always work (00:17:29) this way." So in a codebase of mine um (00:17:33) uh I have a set of documentation that (00:17:36) explains how I like to do things. And (00:17:38) Cloud Code makes a mistake. And so the (00:17:40) next time I can write a memory (00:17:42) essentially, it's it's written as a a (00:17:43) thing they call a skill. Um, and you can (00:17:46) write it off and you can say, "Hey, (00:17:47) whenever you run into this, I want you (00:17:49) to operate in this kind of way." And (00:17:51) that existing across every session is (00:17:54) really a thing you can only do when you (00:17:56) can store it as a file. (00:17:58) >> Yeah, (00:17:58) >> it's a thing you can't do in quite the (00:18:00) same way when you're operating in this (00:18:02) environment where it's just going back (00:18:03) and forth to the API. So, this access to (00:18:06) the file system is one really big piece. (00:18:07) And then the second is is just the Unix (00:18:09) commands. I mean computers, every (00:18:12) computer program lives on top of these (00:18:15) sort of baseline (00:18:17) functions and the way that the designers (00:18:20) of Unix built them is really elegant and (00:18:23) they're very small. They all do one (00:18:25) thing and they're all composable and in (00:18:29) uh coding terms composable means they (00:18:31) can be they can be chained together, (00:18:32) right? And so you can say, "Hey, look (00:18:35) for files that mention this word." And (00:18:37) then from those files, I want you to (00:18:40) take this second action. And then from (00:18:42) the output of that action, I want you to (00:18:43) take a third action. And that's just (00:18:44) built into Unix. You literally just put (00:18:46) a a little pipe in between. And you just (00:18:48) pipe them from one to another. And and (00:18:50) that's it. And so you give it access to (00:18:53) write these commands and all of a sudden (00:18:54) it gets these sort of second and third (00:18:56) order effects that are just incredibly (00:18:57) powerful and built over a really long (00:19:00) time. So how much of clawed code the way (00:19:02) it's different to other models how much (00:19:04) of that was overcoming technological (00:19:06) challenges versus like just having a (00:19:08) good idea because hearing you describe (00:19:11) it I mean giving access to a computer (00:19:14) seems like kind of obvious like let's (00:19:16) let's just do that. (00:19:19) >> I don't have a good answer to that. I (00:19:21) think that it was kind of just a good (00:19:23) idea. (00:19:24) >> Yeah. (00:19:24) >> I think they did some patterns really (00:19:26) well. are clearly incredibly talented (00:19:28) not just engineers but kind of thinkers (00:19:32) about how to structure it like uh the (00:19:34) the primitives inside cloud code are (00:19:36) just smart and then the thing that (00:19:37) they've done and and Boris Churnney (00:19:40) who's the uh lead developer on (00:19:42) cloudcoded anthropic he talks about um (00:19:45) latent demand a lot right and latent (00:19:47) demand is basically just hey look at the (00:19:49) ways people are using these systems and (00:19:50) then figure out ways to make that a part (00:19:52) of the product itself I think what (00:19:54) they've done brilliantly And this is (00:19:56) kind of easy when you have a community (00:19:58) of developers who are nerds who want to (00:19:59) go talk about all the ways that they're (00:20:01) using these things is they have I am (00:20:03) amazed at the speed in which you know I (00:20:05) have a small community of 15 CTOs who (00:20:08) all use this stuff religiously and you (00:20:11) know when we first started that (00:20:12) community it took them a month to I (00:20:14) would see it in the chat and then a (00:20:16) month later it would get built into (00:20:17) cloud code and then increasingly it's (00:20:19) like a day later it feels like they're (00:20:21) just they're just listening to it but I (00:20:22) think they're just not only tapped in (00:20:24) but they're really fundamentally (00:20:26) um you know they're they're dog fooding (00:20:27) it. They they use their own products. (00:20:29) When you um you know they talk about the (00:20:31) productivity engineering productivity at (00:20:33) Anthropic um you know despite growing at (00:20:36) a crazy clip it continues to go up and (00:20:39) and you know anybody who's built had to (00:20:42) manage large scale pieces of software (00:20:45) large scale code bases knows that's not (00:20:47) the norm. (00:20:49) >> So VS code and cursor these are idees. (00:20:54) Claude code is not an ID. What is it? (00:20:56) It's called a CLI. Is that (00:20:57) >> a CLI? A command line interface. (00:20:59) >> Got it. And the other labs now they also (00:21:01) have CLIs. So why are we all talking (00:21:04) about claude code? And I I chat GPTs is (00:21:07) called Codeex. I don't know what Geminis (00:21:09) is called. (00:21:10) >> I think it's just called the Gemini CLI. (00:21:11) >> Why are we all talking about Claude code (00:21:14) rather than the other CLI that kind of (00:21:16) have the same thing? Like what is the (00:21:17) difference? (00:21:18) >> I think first and foremost they were (00:21:20) first. (00:21:20) >> Okay. So, um, and I think they've (00:21:22) they've had a lot more. And, you know, (00:21:24) for my very personal opinion, I think (00:21:26) they've done some things smarter and (00:21:29) better as far as the permissioning (00:21:31) models. So, you know, one of the really (00:21:32) dangerous things is you've got this (00:21:34) thing running on your computer. You (00:21:35) don't want it to go and delete (00:21:36) everything, right? (00:21:37) >> And, um, they have a very fine grain (00:21:40) permissioning model where you can say, (00:21:41) "Hey, I want to allow this just this one (00:21:42) time. Want to always allow it." You (00:21:44) know, (00:21:44) >> I always click always allow. I'm living (00:21:46) on the edge. you can you can uh next (00:21:49) time you run it, you can just do a flag (00:21:50) that says dangerously skip permissions. (00:21:53) Um and and it'll just uh they call it (00:21:55) yolo mode. Um I think I think more (00:21:59) fundamentally though if I look at at (00:22:01) codeex versus cloud code I think it's a (00:22:04) a difference in philosophy around what (00:22:07) you want AI to do. Um, to me, Codex, (00:22:11) which is excellent, is very focused on (00:22:14) building an agent that you can just give (00:22:16) something to and it'll just go do it. (00:22:19) So, I want to give it that task. I don't (00:22:20) want to intercede. I don't want to give (00:22:22) it any more feedback. And Claude code is (00:22:27) much more designed to be uh kind of a (00:22:29) pair programmer. And so, you know, in (00:22:31) engineering, pair programming has (00:22:32) existed for a while. It's a really weird (00:22:35) sort of productivity thing where you put (00:22:37) two engineers on the same problem and it (00:22:39) turns out that you can get better code (00:22:42) force multip. (00:22:42) >> Yeah. And it sort of makes up for the (00:22:44) fact that obviously, you know, you're (00:22:46) doubling the staff on it, but because of (00:22:48) how many fewer bugs because you have (00:22:49) both sets of eyes, it it has seemed to (00:22:51) work out for many folks. Most companies (00:22:53) don't practice it. But I think cloud (00:22:55) code fundamentally is much more designed (00:22:58) in that way. It's a pair programmer. (00:23:00) It's uh they you know whenever I start a (00:23:02) project uh I start in plan mode. So you (00:23:04) start in plan mode. You put together a (00:23:06) plan. I really I mean I spend a lot of (00:23:07) time in plan mode. You go through you it (00:23:09) gives you a plan back. It asks you how (00:23:11) you feel. You can give it a whole bunch (00:23:13) of direction. And then it's only then (00:23:15) that it goes off and it goes into it. So (00:23:17) you know we're working together. And I (00:23:19) actually have a whole system now that (00:23:20) I've designed where um uh I use a a task (00:23:23) management system called linear. So I (00:23:25) have claude code write tasks off to (00:23:27) linear. And then I've worked uh with (00:23:30) claude code to write a document that (00:23:32) helps sort of decide a set of (00:23:34) heruristics to decide when you should (00:23:35) assign it to codeex versus when you (00:23:37) should give it to claude code. And so if (00:23:38) it's tightly defined enough and simple (00:23:40) enough, I just send it off to codeex and (00:23:41) it does it totally independently. (00:23:43) >> And then if it's complicated enough that (00:23:46) I think it requires my time and (00:23:47) attention, then it it saves it for me us (00:23:52) to do together. (00:23:53) >> Um and we'll work on it together. and (00:23:55) and so if it's it's sort of touching a a (00:23:58) kind of important enough if it's (00:23:59) changing some part of the data model (00:24:01) there's these other kind of you know (00:24:03) fairly basic set of criteria that I use (00:24:05) and um but that that to me is the (00:24:08) fundamental distinction and and um you (00:24:10) know I find cloud code in that way to be (00:24:13) just it sort of fits what I want to do (00:24:17) and how I want to work uh much better. (00:24:20) talk a little bit more about how it (00:24:23) actually impacts the workflow of an (00:24:25) engineer because you know my impression (00:24:27) was people can code right like the (00:24:31) coding problem is kind of solved at this (00:24:33) point and even if you can't code even if (00:24:36) you're not a professional engineer you (00:24:37) can hire someone from like India or (00:24:40) Indonesia or wherever to just write you (00:24:42) a code maybe it'll take them a week (00:24:44) instead of like two days with clawed (00:24:46) code but how much does this actually (00:24:49) change the workflow for an engineer (00:24:53) >> as completely as it could be changed. I (00:24:56) mean I would say that over the last (00:24:58) three months I've written personally I (00:25:02) don't know a few hundred lines of code (00:25:04) like I I am mostly a manager of a set of (00:25:08) agents who are writing code on my behalf (00:25:11) and you know increasingly what I think (00:25:13) is interesting I I've been thinking (00:25:14) about this a bunch lately is like in (00:25:16) some ways it's just bringing me back to (00:25:18) the core challenge that has always (00:25:19) existed in software development which is (00:25:22) how do you manage a large scale software (00:25:25) development project, right? (00:25:28) >> It has become a coordination problem and (00:25:30) um and I spend a lot of time sort of now (00:25:33) designing (00:25:34) my clawed code system to ensure that (00:25:37) code goes through all the proper spec (00:25:39) checks and that it it has all these (00:25:41) things. The other thing that you know uh (00:25:44) makes code a particularly good place to (00:25:46) do this is that code is verifiable in a (00:25:48) way that you know most other work is (00:25:50) not. So, you know, with code you can (00:25:53) verify that the build works, right? So, (00:25:55) you can say, hey, I want to build this. (00:25:57) I want to build this package. I want to (00:25:59) make sure that it's actually going to (00:26:00) build and that there's going to be no (00:26:01) failures. That's a very easy check. It's (00:26:03) either true or it's not true. Um there's (00:26:05) also uh coders use linting. And so, (00:26:08) linting is a way to kind of um look, it (00:26:11) it's static code analysis. So it (00:26:13) basically tries to sort of find um uh (00:26:18) things in your codebase that are not (00:26:19) going to work ahead of time where you (00:26:21) can predict that obviously you can't (00:26:23) predict um uh Alan Turing proved that (00:26:26) you can't predict uh with certainty (00:26:28) whether code is going to run but there (00:26:29) are certain patterns and things that it (00:26:30) it can find it's essentially does static (00:26:32) pattern analysis and um so you know you (00:26:34) have it run all these things but um the (00:26:37) more kind of opinionated you can be (00:26:39) about that and the more steps you can (00:26:41) have it go through so I find you know (00:26:42) Now I'm I'm kind of the designer, which (00:26:44) honestly as an entrepreneur and as a CEO (00:26:46) of companies, like that's kind of always (00:26:48) been my job. Like I've I've been not (00:26:51) I've less and less been a person who (00:26:52) writes code and more and more been a (00:26:54) person who designs a system, in that (00:26:56) case a company with a bunch of people (00:26:57) who write code. (00:26:59) >> Um, one of the funny things it seems to (00:27:02) me is that setting aside Claude code, (00:27:05) Claude itself has a reputation for it's (00:27:09) a nicer chatbot to talk to. people find (00:27:12) it um and you know Chad GPT seems to (00:27:15) really be psychopantic. I still think (00:27:17) it's I know it's approved but I actually (00:27:19) don't think it's improved enough. I (00:27:20) still people like the pros style of um (00:27:24) Claude Claude and I'm curious that in (00:27:26) the pair trading pair trading I'm (00:27:29) thinking about finance the pair (00:27:30) engineering model whether there is also (00:27:33) an edge there which is like here is a (00:27:36) chatbot that is not annoying to talk to (00:27:39) while you're iterating and whether that (00:27:40) is like a meaningful distinction between (00:27:44) you know coding with codecs or whatever. (00:27:46) Yeah, I I don't know. I It still can be (00:27:50) very annoying, I will tell you. And (00:27:51) it'll still sometimes um be overly um uh (00:27:57) overly affusive with me about a a design (00:27:59) choice I made or or sort of noticing (00:28:01) something um which I I could live (00:28:03) without. (00:28:03) >> I So I'm working on this project that's (00:28:06) uh doing this linguistic things and I (00:28:08) eventually had to say like give it to me (00:28:10) straight. How bad is this? And then so I (00:28:12) said I said actually what I said was and (00:28:14) assume for a moment that you are a (00:28:16) quantitative linguistics with a PhD. (00:28:18) Give me your honest assessment of where (00:28:20) we are with this. And it said like (00:28:23) you've developed a nice toy and there's (00:28:25) no evidence that it actually does. And I (00:28:27) was like okay that's nice to hear. I (00:28:29) actually like you know I appreciate (00:28:31) that. And it was like very blunt. Not (00:28:33) you know it's still like polite but it (00:28:35) was like this doesn't you haven't really (00:28:37) shown anything. you haven't really (00:28:38) established at all that your software (00:28:40) does what it claims to. (00:28:41) >> Yeah, I think so. I think stylistically (00:28:44) I I kind of personally agree. I my (00:28:47) theory, by the way, on on Claude versus (00:28:50) uh OpenAI chat GPD models is I think (00:28:53) Claude is actually better at sort of (00:28:55) reflecting what you give it. Um, and so (00:28:57) I think part of why we think it's better (00:28:59) is it it's better at pretending it's us. (00:29:02) And so we tend to like that is this is (00:29:04) purely uh speculation but that's always (00:29:06) been my theory on on (00:29:07) >> so it flatters you in a different way. (00:29:09) >> I think it's flattering you in a much (00:29:11) more subtle way. (00:29:12) >> Interesting. Um but uh for a long time (00:29:16) just uh anthropic has been producing the (00:29:18) best coding models you know I mean (00:29:20) there's there can be some debate there (00:29:22) now but um you know there's a great curs (00:29:24) story from cursor actually where cursor (00:29:27) basically wasn't that good and then (00:29:29) sonnet 3.5 came out and all of a sudden (00:29:32) cursor was amazing and cursor became a (00:29:34) tool that everybody started using but it (00:29:36) wasn't until this other model came out (00:29:38) and they made that the default model and (00:29:40) um you know I for what it's worth I (00:29:42) think the take away from that which is a (00:29:44) kind of big theme we see in the market (00:29:46) as a thing that uh the cloud code team (00:29:49) has talked about is um you just (00:29:50) constantly have to be building ahead (00:29:52) with AI in a way that is very unique in (00:29:55) the world of software where you kind of (00:29:57) always want to build things that are (00:29:59) working at like 70 or 80% because if you (00:30:02) really spend the time to get it up to 90 (00:30:04) or 100 you're going to lose all the (00:30:06) gains you get when the next model comes (00:30:07) out (00:30:08) >> and the you know with the amount of (00:30:10) capex being spent on these models like (00:30:12) there's there's a next model that's (00:30:13) going to come out that's going to be (00:30:14) awesome and and you just kind of want to (00:30:16) be downstream from that and you don't (00:30:17) want to waste six months getting an (00:30:19) extra 3% when that new model is going to (00:30:21) give you an extra seven. (00:30:23) >> Yeah. This is the only certainty with AI (00:30:25) is like there's always going to be a new (00:30:27) model, right? (00:30:28) >> The the the worst model we'll ever use (00:30:29) is the one that we're using today. (00:30:31) >> That's right. That's right. Um are we (00:30:33) all going to become coding illiterate? (00:30:35) Are we just going to forget how to code (00:30:37) if everyone's using, you know, general (00:30:39) language to do (00:30:40) >> forget? I never learned. Yeah. Okay. (00:30:42) >> You know what I've been thinking about? (00:30:43) You know that um uh Scott Karp, the CEO (00:30:47) of Palunteer, and he has that line. He's (00:30:49) like, "When I was young, I was too poor (00:30:50) to have a car or so I didn't get a so I (00:30:53) never learned to drive. And now I'm too (00:30:54) rich, so I never learned to drive." I (00:30:56) feel like when I was young, I was too (00:30:58) dumb to learn to code. And now (00:31:00) >> you leaped ahead. (00:31:01) >> Yeah. Now I'm too smart to learn Python (00:31:03) or HTML or whatever. (00:31:05) >> I have a couple takes on this person. (00:31:07) one personally. So, um first one is I I (00:31:09) just think like this is the worry of all (00:31:12) technology ever. There was a paper that (00:31:15) came out that showed that people were um (00:31:18) uh you know they were forgetting more (00:31:20) things or something because they were (00:31:21) using chatgd but you know uh in Fadrris (00:31:24) Plato was worried that people were going (00:31:26) to forget things because they started (00:31:28) writing things down. Um and you know I (00:31:30) think the trade-off there was pretty (00:31:32) good. We got the scientific revolution a (00:31:33) couple other things. So um uh you know I (00:31:37) think that's the sort of natural (00:31:38) knee-jerk um with that said it is I I (00:31:43) it's very strange when you have people (00:31:46) um you know the claude code team is (00:31:48) talking about how little code they (00:31:49) write. Um, now I I draw a distinction (00:31:52) between the sort of vibe coding and the (00:31:55) kind of amateur people who have never (00:31:57) written code and I think that is amazing (00:32:00) by the way and I think um there's a lot (00:32:02) of software developers who are really (00:32:03) mad about that um because they're uh (00:32:06) they they claim it's for safety reasons (00:32:08) or whatever but I think fundamentally (00:32:09) it's just they've got people on their (00:32:10) turf. Um but I I think that's (00:32:13) incredible. I mean, my I my my (00:32:15) 9-year-old um vibecoded a website uh Oh, (00:32:19) wow. and uh for secret Santa. Um she's (00:32:21) now 10. Uh she would get mad at me if I (00:32:23) called her nine, but I think she voded (00:32:25) when she was nine. Um uh but uh (00:32:28) >> that that's awesome, right? I don't (00:32:29) know. That's amazing. That's a a way for (00:32:31) people to express themselves in a way (00:32:33) that they couldn't before you did your (00:32:35) your linguistics project. That's (00:32:37) >> that's fun and interesting. Um, but (00:32:40) yeah, I I also think the other the the (00:32:43) thing that's happening with professional (00:32:44) software developers when you hear from (00:32:45) Anthropic or or you know when I'm (00:32:47) talking about it's you know the the code (00:32:50) is going through this process and you (00:32:52) know all the code still gets reviewed by (00:32:54) people. We're not letting it get out the (00:32:56) door if it's not at the same level as (00:32:58) human and it's just but what's amazing (00:33:00) is I'm I'm running five of these (00:33:02) sessions at a time, right? And so I've (00:33:03) got like software being developed in (00:33:05) parallel in a way that is unimaginable. (00:33:08) And you know the other thing is just now (00:33:11) the best software engineers wrote the (00:33:12) least code anyway. Um you know the the (00:33:15) sort of classic story of like the (00:33:17) difference between a junior developer (00:33:18) and a senior developer is that a junior (00:33:20) developer gets a problem and they sit (00:33:22) down and they put their fingers on the (00:33:24) keyboard and they start writing code. (00:33:25) and a senior developer gets a problem (00:33:27) and sits there for three hours and tries (00:33:30) to figure out what the best way to solve (00:33:31) it is and then spends five minutes (00:33:32) writing code to get it done. (00:33:34) >> True elegance is restraint. That's what (00:33:36) I say. (00:33:36) >> What are you seeing in the companies (00:33:38) you're working for? Like I find it hard (00:33:40) to believe (00:33:42) and I was maybe skeptical of this but it (00:33:44) feels like right now we're here with (00:33:46) technology where like if I were like (00:33:47) companies like like I said you can build (00:33:50) charts of data in a way that used to be (00:33:52) like someone would have had to get their (00:33:53) hands dirty etc. in the companies that (00:33:55) you talk to. Is right now this having an (00:33:58) effect on how they think about what (00:33:59) positions they're hiring for and the (00:34:01) skills they're looking for and so forth? (00:34:03) >> I think that (00:34:06) it's hard to answer right now. Um I (00:34:08) think that certainly (00:34:10) >> I do think a I personally think if I (00:34:13) look at the sort of layoffs in the (00:34:15) technology industry over the last couple (00:34:16) years I I think some part of that is (00:34:17) just looking at the output of these (00:34:19) models and saying hey (00:34:22) >> uh these models are able to produce at (00:34:24) you know the median and I have a whole (00:34:27) bunch of sort of middle managers who are (00:34:28) producing at the 65th percentile and (00:34:31) it's like I can produce median for a$150 (00:34:34) per million tokens or I can produce (00:34:36) 65th% percentile for however many (00:34:39) hundreds of thousands of dollars a year. (00:34:40) It's it's a sort of fairly simple (00:34:41) trade-off. (00:34:42) >> I think so. I I do think there's a lot (00:34:44) of downstream effects. I think the other (00:34:46) thing that's happening is is kind of (00:34:47) like middle management is under threat (00:34:49) because it's the realization that hey, (00:34:51) like part of what these models are (00:34:52) amazing at is is I think of them as like (00:34:55) a fuzzy interface. They can sort of turn (00:34:57) any data into any other data, right? You (00:34:59) can sort of transform data from one (00:35:01) format to another. You can take a PDF (00:35:03) and you can turn it into charts, right? (00:35:04) And there's whole people who exist or (00:35:06) you know if you think about what product (00:35:07) managers do a lot of what product (00:35:08) managers do is they take how people are (00:35:10) using a product and they try to (00:35:12) transform it into a a format that (00:35:14) engineers can then use to figure out (00:35:16) what to do. And I think a lot of those (00:35:18) kind of um uh a lot of those pieces that (00:35:22) used to just be kind of transferring (00:35:25) knowledge. (00:35:27) I've always said Tracy um I think a one (00:35:30) of the most important roles in any (00:35:31) organization is essentially translation (00:35:33) work and you see it in the newsroom (00:35:35) where it's like here is a team (00:35:37) specialized in emerging market (00:35:40) currencies and then they have to like (00:35:42) they have to then tell the senior (00:35:44) editors what they're working on but the (00:35:46) senior editors who are maybe more (00:35:47) generalists don't really know like why (00:35:49) like some sort of like you know Juan Yen (00:35:52) Kerry is important and that a really (00:35:54) important role within any organization (00:35:56) is essentially the the team that can (00:35:58) translate between the generalist team (00:35:59) and the specialist team. Absolutely. (00:36:01) >> And so I that's an interesting (00:36:02) observation in the sort of engineering (00:36:04) world of like okay these are tools that (00:36:05) are in some sense translation tools. (00:36:08) >> So we talked I agree completely by the (00:36:10) way but we talked about vibe coding and (00:36:12) Joe has this application um that I don't (00:36:15) think you're looking to monetize. (00:36:17) >> No it's I'm just trying to make it for (00:36:18) the good of the world. (00:36:19) >> Right. Okay. (00:36:20) >> When did that become a crime? (00:36:22) >> Um (00:36:23) >> I'm not monetizing it. But like this (00:36:25) opens up massive questions for software (00:36:27) as a service, right? For SAS because if (00:36:30) everyone can write their own software, (00:36:33) um you can replicate anything that's out (00:36:34) there that is currently charging money. (00:36:37) What's going to happen to software? (00:36:40) >> I think software is pretty screwed. (00:36:42) >> A lot of it at least. Not all of it. You (00:36:44) know, you still uh it depends on whether (00:36:46) you call the cloud provider software or (00:36:48) not. Um you know, you still need to run (00:36:49) this stuff somewhere. And I think (00:36:50) there's there's certain kinds of (00:36:52) software that, you know, you just don't (00:36:54) really want to be in the business of (00:36:55) writing. Um, you know, I I as someone (00:36:58) who's tried to build a project (00:37:00) management system, I I'd really rather I (00:37:01) don't think anybody should be in that (00:37:03) business. Um, but I I do think (00:37:06) fundamentally, I mean, we see this every (00:37:08) day inside enterprises. the the sort of (00:37:10) build versus buy pendulum has just swung (00:37:14) and and you know I mean I used to run a (00:37:16) SAS company and we sold to enterprises (00:37:19) and um you know for a long time that (00:37:21) that I think that made a lot of sense (00:37:23) right because like hey it just didn't (00:37:25) make sense to try to build this thing on (00:37:27) your own (00:37:28) >> and so but the price of that was you (00:37:30) know one the price right like and and it (00:37:32) got to be more and more expensive the (00:37:33) other price was that you were paying for (00:37:35) a lot of stuff you didn't need right (00:37:36) because the whole job of building SAS is (00:37:39) you need to generalize problems and so (00:37:42) you build things that are going to work (00:37:44) for everybody and that means either you (00:37:46) have to sort of adapt or you have to (00:37:48) build this sort of very configurable um (00:37:50) software. uh and I think and and what I (00:37:54) see just you know firsthand is that uh (00:37:57) inside these organizations you can now (00:38:00) solve very specific problems that are (00:38:03) highly valuable (00:38:06) >> and not only can you solve them better (00:38:08) than generic software but you can (00:38:11) actually in a lot of ways do it for less (00:38:13) money because you're trying to tackle (00:38:14) less stuff you didn't need the 16 other (00:38:17) features you bought it for the one that (00:38:18) you really really cared about Um, and so (00:38:22) I I think that part of it, you know, I (00:38:24) don't like there's I I definitely think (00:38:25) there are pieces of the software (00:38:26) industry that are gonna, you know, come (00:38:28) out the other side. You're gonna Nobody (00:38:30) wants to deal with payroll, right? Like, (00:38:32) you know, somebody you're still going to (00:38:33) buy some payroll software and and you're (00:38:35) still going to have that. But, um, you (00:38:37) know, I do think there are a lot of (00:38:38) pieces where the software existed (00:38:40) essentially as a a kind of wrapper (00:38:42) around a database and now you're just (00:38:44) going to, you know, with just the (00:38:46) database you can do that. And then you (00:38:48) know the other piece I'd say here is (00:38:49) it's this is not there's a kind of (00:38:51) confluence of circumstances where it's (00:38:52) not just the coding it's also the fact (00:38:56) that you have AI to do a whole bunch of (00:38:58) work. So you know if we pick on CRM for (00:39:01) a second right like you know (00:39:02) >> salesforce.com (00:39:03) >> salesforce.com we can you know you look (00:39:06) at what the interface of that is and (00:39:08) essentially it has existed to get (00:39:09) salespeople to take unstructured data (00:39:12) which is sales meetings and turn it into (00:39:14) structured data that so it can be stored (00:39:16) in a database and now you have AI and AI (00:39:20) is very capable of taking unstructured (00:39:22) data (00:39:23) >> directly from the source. So you have (00:39:25) people recording meetings and then it (00:39:27) can structure it into any data that you (00:39:29) want. This is one of the very first sort (00:39:31) of mind-blowing moments I had was that I (00:39:34) could give it uh a JSON interface. I (00:39:37) could describe exactly what I wanted the (00:39:40) data structure to be and it would give (00:39:42) me back that information and that data (00:39:44) structure. And we've just basically been (00:39:46) having a bunch of humans do that work (00:39:47) for a very long time, whether it's in (00:39:48) CRM or project management or any of (00:39:50) these other places. and the ability to (00:39:52) just kind of get rid of that whole (00:39:53) thing. It I think it really does bring (00:39:54) into question the value of a lot of (00:39:56) these software companies. (00:39:58) >> Well, so we have seen like a lot of (00:39:59) software stocks they look like melting (00:40:01) ice cubes right now. Maybe they so what (00:40:04) is so I want to talk I mean this is like (00:40:06) you know our listeners who are investors (00:40:08) there's a pretty high stakes question of (00:40:10) like what residual value there is but (00:40:12) talk a little bit more about Salesforce. (00:40:14) Maybe this would be a time to learn what (00:40:15) sal (00:40:16) >> what it actually does (00:40:17) >> as it's massively being disrupted now we (00:40:19) get around to learning what Salesforce (00:40:21) is. But I know it's like many things (00:40:23) there are apps that people built on to (00:40:24) Salesforce. But this sounds like we're (00:40:26) hitting on what I think probably one of (00:40:27) the crucial questions for like the (00:40:29) future of the software industry. So talk (00:40:31) a little bit more about like the current (00:40:33) approach and what people are buying when (00:40:34) they buy a package or subscribe to a (00:40:37) service from Salesforce and then what (00:40:39) the unlock opportunity is from having AI (00:40:43) like live in the same world as all your (00:40:45) files. (00:40:46) >> Yeah. So I think if you if we take CRM (00:40:50) as the general category so you know the (00:40:51) biggest players there are (00:40:53) >> that's customer relationship (00:40:54) >> customer relationship management that's (00:40:55) like what you know Salesforce does it (00:40:57) SAP does it HubSpot does it for the (00:40:59) midmarket um (00:41:01) >> you know when I think about that product (00:41:03) and I think about the way we've used it (00:41:04) inside enterprise sales organizations (00:41:06) essentially you know it's a database of (00:41:08) companies it's a database of contacts (00:41:10) it's a database of deals you have in the (00:41:12) pipeline and it's a way to track all (00:41:14) those deals you guys hit on something (00:41:15) before that I think is is really it (00:41:19) which is like inside companies there is (00:41:22) a huge group of people and who exist to (00:41:24) answer the question from management of (00:41:26) what is the status of something (00:41:28) >> right and you know that can be sales (00:41:29) management it can be product management (00:41:31) it doesn't matter right it could be (00:41:32) within a newsroom somebody wants to know (00:41:34) what the status is and somebody else (00:41:36) exists to go figure out what the answer (00:41:38) to that question is and so fundamentally (00:41:41) I think those CRM tools are bought first (00:41:43) and foremost to answer what is the (00:41:45) status right? What's my pipeline look (00:41:47) like? Um, and to answer what your (00:41:48) pipeline looks like, you need a bunch of (00:41:50) sales people putting deals in. Um, and (00:41:52) those deals are associated with contacts (00:41:54) and companies. And they say, "When is (00:41:55) that deal going to close?" And and and (00:41:58) essentially, you were asking the (00:41:59) salespeople to make the updates in the (00:42:01) system to do that. And just very (00:42:05) tactically, I mean, you know, I I run a (00:42:08) company now. We talk to a lot of we have (00:42:10) a lot of sales calls. We record those (00:42:12) calls and they get transcribed and the (00:42:15) AI then looks through them and makes (00:42:17) decisions about where this deal should (00:42:19) be in the process. And it's much better (00:42:23) than having somebody try to go updated (00:42:25) because those people never updated (00:42:27) anyway. The secret of all of this (00:42:28) enterprise software is that nobody was (00:42:30) using it the way that anybody wanted to (00:42:32) anyway. (00:42:32) >> Um, and so, you know, I think that that (00:42:37) is sort of, you know, a lot of what's (00:42:38) happening there. Again, it's sort of (00:42:39) some of it's the coding, some of it's (00:42:41) just the core capabilities and then, you (00:42:42) know, you still need databases, right? (00:42:44) So, it's like, you know, you look at (00:42:45) what data bricks and snowflake and you (00:42:46) know, I think those folks are still sort (00:42:48) of genuinely sitting in a pretty good (00:42:50) place where, you know, all software has (00:42:52) to sit on side on top of some database (00:42:55) that you can sort of read and write to. (00:42:58) Um, but you know, I think some of those (00:43:00) categories that were (00:43:02) >> specifically focused on kind of like (00:43:05) human input. Um now of course you know (00:43:07) Salesforce has a whole AI thing and (00:43:09) they're saying hey we you shouldn't have (00:43:10) humans inputting in Salesforce you know (00:43:12) to at sales is just one small piece they (00:43:14) have a whole customer support thing (00:43:16) which obviously also has an interesting (00:43:18) implication where you know you're doing (00:43:20) support with AI agents and so some of it (00:43:22) comes back to seats I mean you know it (00:43:24) gets to be fairly complicated but I do (00:43:26) think I think the fundamental underlying (00:43:29) thing is anybody who buys software that (00:43:32) is you know SAS you're always buying (00:43:36) paying for a subset of the functionality (00:43:38) that that's nobody is using 100% of the (00:43:40) functionality of SAS. And so there's (00:43:42) always a trade-off that's happening (00:43:43) there where you know you're spending (00:43:45) more money than you need to because (00:43:46) you're not using all of these pieces. (00:43:48) And so you know if you can more narrowly (00:43:50) focus that that is where you could say (00:43:52) hey we could solve this kind of more (00:43:54) narrow problem and not only can we solve (00:43:55) it more narrowly we can solve it way (00:43:57) more effectively because you know the (00:43:59) trick with AI is that the more specific (00:44:02) you are with it the better the output is (00:44:04) right so it's like if you know if (00:44:06) outside of coding if you just ask chat (00:44:08) GPD to write you a story it's going to (00:44:10) write you a very very median story right (00:44:13) um sort of exactly the median but if you (00:44:16) work with it and you you know, then (00:44:19) you're going to get it. The more of your (00:44:20) own expertise you imbue in it, the (00:44:23) further up above the median it's going (00:44:25) to be. And it's going to be, you know, (00:44:26) of course that also means it's less what (00:44:29) where the line is between what's AI and (00:44:31) what's not AI is going to continue to (00:44:33) get blurriier. (00:44:34) >> Joe, how much does clawed code actually (00:44:37) cost? Do you know? (00:44:38) >> Well, I paid for the uh 200 a month $200 (00:44:42) a month version, but like (00:44:44) >> high roller. (00:44:45) >> Yeah, I know. But uh you know I think (00:44:47) it's you can get it with the pro version (00:44:49) of like or whatever the sub the version (00:44:51) of that below $20, but I hit a limit (00:44:53) fairly quickly and I was like I didn't (00:44:54) have my website up. So like and then I (00:44:56) bought the five then I paid $5 for the (00:44:59) extra compute. I was like this is dumb. (00:45:01) I think I just Yeah. (00:45:03) >> Okay. So (00:45:04) >> we going out to two nice dinners in a (00:45:07) month. It's not you know when I think (00:45:08) about that way it doesn't seem that big (00:45:09) of a deal. (00:45:10) >> It's worth it to you. Okay. So, I think (00:45:11) we can all agree this is like a valuable (00:45:14) service that Claude Code is providing, (00:45:17) but we touched on this in the intro. It (00:45:19) seems like the models just keep (00:45:22) replicating themselves really, really (00:45:24) quickly. So, anything that Claude Code (00:45:25) can do, I would expect another model (00:45:28) will come in in like a month, maybe (00:45:30) less, and do the exact same thing. What (00:45:32) does that mean for the actual like (00:45:34) valuations of these companies and the (00:45:37) models? like how are they going to (00:45:38) monetize it when it seems so difficult (00:45:41) >> to actually differentiate yourself (00:45:43) especially for like a substantial (00:45:46) >> portion of time. Yeah. (00:45:47) >> Well, so again here I think we have to (00:45:50) distinguish between claude code and the (00:45:53) claude model. So in Claude Code's case, (00:45:55) if you're using, you know, the latest (00:45:57) version, you're using Opus 4.5, which is (00:45:59) the model, Opus 4.5 has a price of, I (00:46:02) don't know, something in the $150 to $2 (00:46:04) for a million input tokens and whatever (00:46:07) it is on the output, which is like (00:46:08) roughly the going rate for cutting edge (00:46:10) models. Gemini 3 Pro is the same price. (00:46:13) Uh uh OpenAI, Chad GBD 5.2 is they're (00:46:17) they're all the same price. So the first (00:46:19) thing is is you have to differentiate (00:46:20) between those. Um, and so I think a big (00:46:23) part of what Anthropic is trying to do (00:46:25) is they're trying to lock people into (00:46:26) Claude Code. In fact, there was just (00:46:28) some um, uh, controversy amongst some (00:46:31) nerds, um, where, uh, Open Code, which (00:46:34) is a competitor to Claude Code, um, used (00:46:37) to let you use your Claude Max $200. So, (00:46:40) the trick with the Claude Max plan is if (00:46:42) you're just buying those that number of (00:46:44) tokens, it would cost you significantly (00:46:46) more than $200. It is a super super (00:46:48) discounted plan. you are probably you (00:46:51) have the access I I I have the access to (00:46:53) use I would guess in the thousand or 2 (00:46:56) thousand dollars of tokens um uh for my (00:46:59) $200 a month. So it's it's a very very (00:47:01) heavily subsidized plan and open code (00:47:04) which is an open- source version of (00:47:05) Claude Code a sort of competitor um they (00:47:08) had found a way that they would let you (00:47:10) use your Claude Max plan with open code (00:47:13) >> and Anthropic last week (00:47:16) >> shut that down. Yeah. Um, and some open (00:47:18) code people got very upset um because (00:47:20) they said like this is not what you're (00:47:23) supposed to do or I'm not sure exactly (00:47:25) what they said. I I never felt like I I (00:47:27) got a particularly good argument out of (00:47:28) it. Um but uh you know I do think part (00:47:31) of what they're trying to get at because (00:47:33) is (00:47:35) that you know at the very top models (00:47:37) like these are all amazing like the (00:47:40) Google (00:47:41) openai and anthropic their best models (00:47:44) are all on par with each other. I mean I (00:47:47) I would move them around a little bit. I (00:47:49) I still think Opus 4.5 is the best model (00:47:52) out there, but you know, I mean that (00:47:53) might change tomorrow like and that's (00:47:56) where something like cloud code is (00:47:57) really interesting because it's a a (00:47:59) product that is very it's just theirs. (00:48:01) It's not it's a piece of software. It's (00:48:03) not an AI model. Um and so it's sort of (00:48:06) it's less able to be disrupted. Now (00:48:09) again I I think if if somebody else (00:48:11) wanted to copy that exactly they could. (00:48:13) Codeex has one. Gemini has one. I just (00:48:16) think they take a very different tact (00:48:18) with it where it's much less and so you (00:48:20) know I think what they're trying to do (00:48:21) is get developers like me to feel very (00:48:22) comfortable inside that so that when we (00:48:24) go open I still open codeex or try (00:48:27) Gemini or I was playing with open code (00:48:29) the other day and um it just doesn't (00:48:31) feel familiar in the same way that you (00:48:33) know if you're trying to move somebody (00:48:34) from a PC to a Mac it doesn't feel (00:48:36) familiar (00:48:36) >> right they want to own like the (00:48:37) ecosystem I guess environment the (00:48:39) environment (00:48:40) >> work on what a world Noah thank you so (00:48:42) much for coming on a lot I was like (00:48:44) dying to do an episode about this topic. (00:48:46) By the way, I don't have AI psychosis. I (00:48:49) have a Claude complex. (00:48:50) >> Why is everyone making that joke? (00:48:52) >> Wait, which joke? (00:48:54) >> The psychosis joke. (00:48:55) >> I thought you were going to be proud of (00:48:57) me for saying Claude complex. (00:48:58) >> Oh. Oh, that is very good. (00:49:00) >> I It's like I do one pun finally for (00:49:02) Tracy and she just like for making that (00:49:04) joke. (00:49:04) >> Well, I was I was I was handing you a I (00:49:07) finally make a pun and you just jump (00:49:08) right over it. (00:49:09) >> Well, everyone keeps saying that Claude (00:49:10) code is AI psychosis for smart people, (00:49:13) right? Like, how did that become a (00:49:14) thing? addictive. (00:49:15) >> Yeah. All right. (00:49:16) >> But there's a good pun. (00:49:17) >> It's also very brocoded. I find (00:49:19) >> You think so? (00:49:20) >> All of AI is brocoded. (00:49:21) >> Uh, this is true. We should talk more (00:49:24) about this. You know, we should have (00:49:25) David Shore on. He's been doing a lot of (00:49:26) polling about various demographics and (00:49:28) how they feel about AI. (00:49:29) >> We should um and some interesting stuff. (00:49:31) I'd be into that. Yeah, we should uh do (00:49:33) that. Anyway, Noah, thank you so much (00:49:34) for coming on out loud. (00:49:35) >> Thanks for having me. (00:49:36) >> Well, that was fun, Tracy. I really like (00:49:38) I it's obvious to anyone who's been (00:49:40) within five minutes five feet of me for (00:49:42) the last two weeks. I'm like totally (00:49:44) addicted and gone down I never gone down (00:49:47) the rabbit hole and stuff but like I for (00:49:50) the first time unironically I'm like (00:49:52) okay this is transformative technology (00:49:54) beyond being very impressive technology (00:49:57) >> right so I've been coming to a (00:50:00) conclusion which is that you know (00:50:02) >> AI can be both underhyped (00:50:05) >> and overvalued simultaneously like and I (00:50:08) I feel like that's kind of where we are (00:50:10) at the moment (00:50:11) >> is that where you're making your stop (00:50:12) call. Yeah. No, but seriously, like it (00:50:16) >> it's a big deal. It's going to change (00:50:18) the way we work, but is it monetizable? (00:50:21) Can you differentiate the actual models? (00:50:23) The better the technology gets, like the (00:50:26) easier it is to just do what everyone (00:50:28) else is doing. And also like the compute (00:50:30) gets cheaper and cheaper. So, I just (00:50:32) don't know how you monetize this. (00:50:34) >> Well, so that's very interesting his (00:50:36) point, which is that it's the tokens are (00:50:38) heavily subsidized still. Yeah. And so (00:50:40) that if you're paying and actually using (00:50:42) that $200 max program and you actually (00:50:44) use it to the limit, Claude is going to (00:50:47) lose money on this and then the prices (00:50:50) keep dropping. And I know like Claude (00:50:52) Code is okay, they're attempting to (00:50:55) create something that resembles a (00:50:56) traditional software ecosystem that you (00:50:58) feel as a user that you're locked into. (00:51:00) But so far in my various like since (00:51:04) November 2022 when I started playing (00:51:06) with AI, it hasn't felt like anyone has (00:51:09) established lock in with anything and (00:51:11) it's very (00:51:13) >> um uh it's very movable and I suspect (00:51:16) even though I have this file now on my (00:51:17) desktop that has a file called claude MD (00:51:20) that gives instructions etc. I'm certain (00:51:23) that if I open this file with codecs or (00:51:25) Google's I could probably just pick it (00:51:27) up the same. (00:51:28) >> Yeah. I also think there's a fundamental (00:51:30) issue with the lockin strategy because (00:51:32) when you're talking about technology and (00:51:33) the internet like it just feels very (00:51:36) against the grain to try to lock people (00:51:39) into anything and we've seen various (00:51:41) projects over the years and it's it's a (00:51:44) lot harder than it looks. (00:51:46) >> Yeah. I mean I guess I would say it's a (00:51:48) lot harder than it looks but then we (00:51:50) also know the flip side which is that (00:51:51) tons of people are locked into software (00:51:53) that they hate. Right. People are Oh, I (00:51:55) hate people. How many times have you, (00:51:57) oh, I I hate Outlook, right? Or I hate (00:51:59) Microsoft Teams and I hate this and I (00:52:02) spend money on it every month and my (00:52:03) organization can't move off of it or we (00:52:05) can't migrate off of it. So, I do think (00:52:07) that cuts both ways. I do think he (00:52:09) offered the best explanation I've heard (00:52:12) of why the AI coding models are a threat (00:52:16) to a lot of pretty big software (00:52:18) businesses, especially totally (00:52:20) especially the point about how the user (00:52:22) never uses all of the features that they (00:52:24) actually that the software got built for (00:52:26) and therefore maybe the build versus buy (00:52:29) calculation really starts to shift when (00:52:31) they can just design that one feature (00:52:33) very quickly. (00:52:34) >> Um I totally agree on the software side. (00:52:36) It seems like an existential threat, but (00:52:38) just like the locked in ecosystem of a (00:52:41) particular model. Um, I know you said (00:52:44) it's not actually a model, but that (00:52:46) seems like a bigger issue to me. I don't (00:52:48) know. I guess we'll see. (00:52:49) >> We're going to see. And I don't know. I (00:52:50) kind of think we're going to see (00:52:51) quickly. (00:52:52) >> Yeah. That's again, that's the only (00:52:54) certainty is like stuff is happening (00:52:56) fast. (00:52:56) >> Stuff is happening now. Yeah. (00:52:57) >> All right. Shall we leave it there? (00:52:58) >> Let's leave it there. (00:52:59) >> Okay. This has been another episode of (00:53:01) the Odd Thoughts podcast. I'm Tracy (00:53:03) Aloway. You can follow me at Tracy (00:53:04) Aloway. (00:53:05) >> And I'm Joe Weisenthal. You can follow (00:53:07) me at the stalwart. Follow our guest (00:53:09) Noah Brier. He's at Hey, it's Noah. (00:53:11) Follow our producers Carmen Rodriguez at (00:53:13) Carmen Arman Dashelob Bennett at Dashbot (00:53:16) and Kalebrooks at Kalebrooks. (00:53:17) >> And for more Athoughts content uh and to (00:53:20) see what Joe has actually been working (00:53:22) on with Claude Code, check out our daily (00:53:24) newsletter. You can find that at (00:53:25) bloomberg.comthoughts. (00:53:28) And uh you can join fellow listeners in (00:53:30) conversation 247 in our discord (00:53:32) discord.ggods. (00:53:35) >> And if you enjoyed this conversation, if (00:53:37) you like it when we talk about AI, then (00:53:39) please leave a comment or like the video (00:53:41) or better yet subscribe. (00:53:43) >> Thanks for watching.