# The three paths AI could take from here - Shawn Wang SWYX interview [Podcast #208]

Channel: freeCodeCamp.org
Video: https://www.youtube.com/watch?v=kQqrMNviM9U
Duration: 1 hr 16 min
Language: English
Words: 15273
Transcript page: https://viewrankai.com/tools/youtube-transcript/kQqrMNviM9U

---

[0:00] Welcome back to the Free Code Campamp podcast. I'm Quincy Larson, teacher and founder of free codecamp.org and we are interviewing Sea Wang, aka Swix, who coined the term AI engineer and has been teaching devs how to fully leverage LLM tools for the past four years. I hung out with him last weekend at Stanford

[0:20] University's Tree Hacks Hackathon. 15,000 developers entered this hackathon. Only 1,000 were selected to compete and Sean got to judge and help choose the winners of the hackathon. I also got to hang out. I shot a ton of footage. We've got a documentary about the hackathon coming soon. Before we get to the interview, some community news. Free Coke Camp just published a comprehensive course that will teach you how to use security focused KI Linux operating system. You'll learn how to identify, exploit, and defend against realworld vulnerabilities. You'll also build a solid foundation in penetration testing, network security, vulnerability assessment, and more. Most importantly, you'll learn how to think like a

[1:05] security engineer, and leverage tools of the trade like end mapap and wire shark. We also published this guide to passing the certified Kubernetes administrator exam. Bo Car teaches this course which will walk you through key DevOps concepts. You'll start by setting up your Kubernetes practice environment. Then you'll bootstrap a multi-node cluster and your control plane. Then you'll learn about Helm, high availability, autoscaling, core DNS, and more. That's a free 2hour course on the Free Code Camp YouTube channel. Links to all these things are in the description, but please finish listening to this

[1:40] interview before you jump over to those. I know it's very exciting. Uh, and we published a fulllength handbook on Freo Camp Press last week that you can read right in your browser. It will teach you modern React data fetching best practices. You'll learn how to leverage suspense error boundary, the new use parenthesis API. And if you're interested in web development, front-end development, this is well worth bookmarking and reading when you have time. Again, it is a heavy read. And

[2:09] speaking of heavy reads, a heavy watch. We just published a DevOps course that free co- camp instructor Gavin Lawn finished. Uh you'll learn how to take a full stack app on your local machine and ship it to a fully containerized production environment. Along the way you'll learn about CI/CD, that is continuous integration, continuous delivery pipelines. Uh you'll learn about Docker images, launching containers, and more. By the end of the course, you'll have a professionalgrade pipeline that automatically builds and

[2:41] deploys your updates with every push. Again, 4-hour course on the Free Code Camp YouTube channel. This week's song of the week, 1980 classic, Turn Me Loose by Canadian band Lover Boy. When I was a kid, I thought the name for this band was so cringe, but turns out they're a freaking insane band. Really strong chops on these gents. Uh, so the song is built on top of a super catchy syncopated baseline and it just has the most expressive vocals. The vocalist is really dynamite, really soulful. It's

[3:13] got these cool ' 80s buzzing synths. It's got these really percussive piano and the guitar solo is top shelf. Give it a listen after you finish listening to the podcast. Support for this podcast comes from the 10,121 kind folks who donate to our charity each month. Join them and support our mission at donate.freeccoamp.org. You can also pick up a free Codeamp shirt and rep the free Codeamp community with pride. $20 gets you the shirt

[3:41] shipped anywhere in the US. Shop.freecodeamp.org. And for the main course, today's interview with Sean Wang. He's a software engineer, founder of the AI engineer conference, which is having upcoming events in Europe, Singapore. It's really taking off. like a lot of people are interested in uh the kind of overlap between software engineering and leveraging AI tools, fleets of agents, getting things done. It's pretty awesome. So, uh he's the host of the latent space podcast. If you want to learn about like kind of the most advanced machine learning and u you know just a lot of the leading edge, check out his podcast too. I listen to it a

[4:22] lot myself. We're going to talk on this interview though about how even if LLM's plateau, there will still be paths to better output through the surrounding harness code and three big areas that researchers are exploring that Shawn thinks can further improve model performance. World models, multimodality, and embodied AI. That is taking an AI and putting it into a robot for an egocentric firstperson view of the world. Crazy stuff. Uh Sean also shares which skills he thinks are the most important for developers going forward and why he thinks you should switch to your own self-eing uh you should switch from thinking in terms of just in time learning to just in case

[5:04] learning. Sean Wang, welcome back to the Free Code Camp podcast. Hey Quincy, how's it going? Good to see you. Rocking and rolling, man. I'm thrilled to have you back. This is your third time on and uh we've learned a lot from you in the previous interviews and I'm excited to learn a lot from you this time too. I always it's always a good time chatting and uh happy to entertain where I can, educate where I can as

[5:24] well. Right on. Well, let's start with Tree Hacks Stanford's uh hackathon. Last week, you and I hung out there. You were judging. I was recording a documentary that will go live on the free co podcast, uh the free co YouTube channel, probably within the next month and a half, two months. Uh but it is a very exclusive hackathon that 15,000 people applied for and only 1,000 entered. So that's a very uh similar to Stanford's university admission rate in terms of the people that are able to actually participate and uh some banger freaking projects came out of there. I was super impressed. But uh you were judging what

[6:03] was your favorite project? Um I I don't remember their exact name, but it was a project 195. Um, and I just remember that cuz I was trying to enter it in the system and it kept refusing me. Uh, they did like this sort of reverse engineering project where they had this like sort of universal unlock tool kind of that you could throw it at any binary. So, you know, imagine they had a drone, they had like old Logitech thing where you lost the firmware documentation and it was just it's just like an executable file and they were like, "Okay, what what can I do with this?" Um, and so they worked out a

[6:38] reverse engineering script that could just like turn any piece of software and and sort of expose the internals so that you could then create controls and vibe code anything on top of it. Um so it's basically reverse engineering by looking at binaries and I thought like wow I mean like you know the fact that this this can just be a weekend project that um makes any sort of outdated super old software like suddenly instantly relevant but also like you know it can be turned into a hacking tool if uh if

[7:05] you you know know what you're doing. It's kind of like a lock picker for like digital software and like that's kind of very cool. And so uh I really like it also because then they also did some uh fine-tuning on open models uh in order to improve their performance on uh reversing binaries. And I thought that was like a really good example of like this is not just like throw an LM at it and then you know take credit for what the LMS are inherently capable of doing but this is actually applying machine learning and AI engineering to uh push LMS and spike LM in useful ways. Um, and

[7:39] I I thought like, yeah, this is the kind of project I want to see. I think there's a lot of lazy LLM engineering out there that um, you know, it's it's just kind of throwing a nice UI on on what is essentially already in the model. Yeah. The criticism a lot of people have is like a lot of products are essentially just wrappers for foundation models. Is that an accurate characterization of a lot of products

[8:01] coming to market? Yes, but you know, chat GBT is a rapper over the GPT API and that's okay. like there there are valuable rappers out there. Um I just like for a hackathon you want to have something technically impressive. So I do think uh there's a difference between what is technically impressive for hackathon where you should win on merit versus in the business place. Yeah. If you uh make a rapper that is shinier than the other rappers and make more money then more

[8:30] power to you. No judgment. Yeah. Well, let's talk a little bit about the nature of hackathons. Uh, you've probably participated in a lot. I've participated in a lot. I've judged a lot. Um, but I had the luxury of judging before these LLM code generation tools became ubiquitous. Um, and uh, I've seen them be, you know, pretty rapidly adopted by uh, devs at these hackathons, both student hackathons and non- studentent hackathons. uh can you talk about how hackathons have changed and how these LLM tools have I guess

[9:05] been adopted? Yeah. Um I think one thing that's uh pretty obvious so I used to do hack I used to organize hackathons myself. I used to attend them. I no longer do. Um mostly because in San Francisco at least there is a rise in hackathon cheating where basically uh you know the the spirit of the hackathon is you land there you start from kind of a blank slate as as close to blank slate as possible and you do most of the substantive work at the hackathon. uh a lot of pe a lot of winners these days and a lot of hackathon organizers actually kind of just have let that slide and so people

[9:42] mostly come with 99% done work and then they like they're like oh okay like I guess today Google's the sponsor so I'm going to use the Gemini API but other other than that other than tweaking the model string and the API provider everything else stays the same from like a previous work that they've done and then they win because it looks good and it's obviously a lot of substantive work polish and no one checks whether it's done during the the weekend and the hackathon organizers also don't care

[10:06] because they want impressive projects. So it's like it's like a weird like bad incentive uh that I don't really know how to navigate and and obviously it's also unfair to compete right under under those conditions. So uh I think that's one thing that's sort of a non-technical reason why hackathons are are a weird area for me. Um I do also I did also run into nontechnical people at tree hacks who had good submissions and they don't know how to code and they were like yep I I'm a medical student. uh I was working on this thing and I'm like well I you know I would love a weekend uh to

[10:38] to code up something that I I would use and they did and they don't know how to code and I think that's like fascinating because now a whole generation of people are enabled to do like you know software informed by their domain expertise but they don't need a developer like the the the the lady I judged was just a single individual and she was a medical student and she showed like some radiology thing I didn't understand but great for her like she could do it and she could make that vision come to life without hiring

[11:09] an expensive Stanford CSgrad. Yeah. And I I think it's awesome that people from outside of software engineering as a discipline who have a lot of the domain expertise in like the sciences for example or in other fields of engineering can essentially kind of like do their best impersonation of a software engineer just by using superpowered tools. Um, and I want to talk about those tools a little bit because uh, it is definitely tempting for somebody who is a developer already to just say, "Okay, there's a lot of turnover in the tools. There's a lot of fads here." You can see like like it almost feels like trying to chase shiny

[11:50] objects, shiny object syndrome. If you're going out and installing the latest and greatest thing, if you're buying like, you know, a Mac Mini to run uh, you know, claw open claw on or something like that, is that even going to be a thing? like a few months from now or is some new tool going to be just better in every regard like because there's so much turnover. Reminds me of the JavaScript framework wars where it seemed like every week there was some new framework that like did something better than the other ones and and they were inoperable like not interoperable and you had to like essentially throw

[12:26] out like a lot of the skills you learned to do one to do a different one because they'd be using a different paradigm and they were all very opinionated stuff like that. like uh and my reaction there and a lot of people's reaction there is like cool I'm just going to focus on learning the fundamentals and we'll see what happens with these tools way high up in the layer of abstraction and uh later once there's a solidified kind of tool chain I'll adopt it and you could definitely say that like for example React has one NodeJS has one you know um to extent like like express and stuff

[13:00] like that like like uh Yes, there there is there are still some people arguing otherwise, but if you go into an LLM and you start building an app, it's probably going to use Reaction. It's probably going to use Node to build that out, right? Um, and I think that's a sign that like they are winning and they have an incredible amount of momentum by virtue of being so well represented in the training data on GitHub. You know, like most projects on GitHub are JavaScript projects and if they're full stack, they're probably using Node, for example. So like my question, a very long one, but essentially I'm just

[13:34] trying to say should people wait and just focus on things that aren't going to change like computer science concepts, database concepts, network engineering concepts, or should they venture out and start experimenting with these new tools knowing that there is no real terrairma there and that it's going to take a while for these to settle? Yeah, I think you and I somewhat agree and then somewhat disagree and it's it's interesting to tease out the differences. Um, now let me provide the counterpoint. I'm just going to like complete provide the complete counterpoint, right? Like um a lot of this kind of learning is just in case learning. Uh you're you're learning

[14:12] because you imagine that the fundamentals are important to understand the whole system. Uh, but you know, did you learn how every single part of the car works before learning how to drive? Well, not really. You just get in with, you know, some someone like watching you drive and you press the pedals here, turn the steering wheel there, you're good. Um, and in in in in some ways like software is becoming more abstracted, right? Uh, most people uh will not will never learn assembly. Free code has never taught assembly. Uh, right. I

[14:43] don't know. You guys teach a lot. No. and and like I don't even believe that we have like a extracurricular video course on it and we've got extracurricular courses on practically everything including like cobalt but like we do music now you know. Yeah, we have we just published a chess course the other day I'm very excited to

[14:58] watch. Yeah. Yeah. So, so you know what I'm saying? Like, so um there's layers that just get solidified and like they're so reliable that um you just don't need to know them like or you can learn them just in time when you actually need to know them, not just in case when you just because you think it's it's fundamentals. But like 30 years ago, assembly was fundamentals and then like 20 years ago, let's let's call it C++ was fundamentals. Um and now now this is the the the phase where like well maybe databases and maybe that kind of infrastructure maybe Node.js JS, maybe React. These are all fundamentals

[15:36] that maybe get abstracted away. Like it it really like we start to get into this field where um things are definitely moving faster than they used to be and timelines are accelerating and it's actually not that clear to me what is a fundamental that there's a lot of things that will never change, but what must you learn in order to be productive and like to get yourself out of trouble, especially when LMN's run into trouble, which they do. Um and uh and what it's like a a basic component of literacy where like yeah I don't use calculus every day but like am I happy that I did

[16:11] calculus in school? Yes. Like it helps me think about derivatives and helps me understand like gives me at least what I know I don't know so I can at least have a basic tool set to go and go and learn everything else. So I think that's like what you want to aim for not to learn um too much of the fundamentals just in case because that can take up too much learning and actually you just want to build and then learn just in time and especially because in this era learning just in time is actually a lot easier than it used to be. Right? In the past

[16:41] learning just in time meant oh I need uh I suddenly run into this problem where I need to level up. I need to stop what I'm doing and I need to go take a free code cam course, sit down with Andrew Brown for like four hours and like learn AWS, whatever. Um, and then I come back and I and I can I you know like I know kung fu now and I can go do that. Um, that used that used to be the case. Now it's just an LLM query, right? So the the just in the the the the speed of just in time learning has dramatically

[17:11] decreased uh increased and I think that that means that your just in case learning strategy should adapt right because then you don't need to frontload so much learning. So I think that's that's one uh important thing. The second thing is a career strategy thing. Um before I move on the career strategy thing you want to respond to that I don't know if you want to you have any difference on the just in time just in case. I mean all this is very interesting and uh I like your just in case learning as a as a kind of like

[17:40] analogous term to just in time learning. Just in time being like a kind of manufacturing production terminology that we imported into education. Um one thing you mentioned is like you don't need to know how the car works to be able to drive the car. But if you look at like the kind of people that drive cars or work with cars, like you have of course like drivers, people driving driving like Uber or something like that and they're going to make dramatically less money and probably have a lot less job security than mechanics and mechanics are probably going to make less money and have less career security than you know automotive

[18:13] engineers who are actually like building and testing the different systems that power the cars. So there is like a very clear kind of like hierarchy or there's a clear kind of like pathway upward uh to where like the value is so to speak. I mean there's certainly value in driving a car and having people drive you around and stuff like that but uh and having people be able to fix your car and having somebody local who you can trust who can do that like that's

[18:39] why mechanics make pretty good money. But being like an automotive engineer where you're maybe making hundreds of thousands of dollars and you're being able to do like like kind of almost like artistry uh when designing these cars and stuff like that like that requires a whole lot of additional skills. So you could almost look at like each of those is like maybe like a a four-year jump in the amount of training necessary uh between those different tiers. Do you think like similar tiers are going to exist with software engineering going forward where yeah there there could be like a you know the equivalent of like the WordPress dev who installs Woo

[19:12] Commerce and helps build out like the the very simple e-commerce experience that like a small local business needs versus the person who's building like software as a service tools versus somebody who's you know training machine learning models and stuff like that. Yeah, I mean absolutely and there always has been and there will continue to be,

[19:28] right? I'm not saying the levels change. Uh there will be levels and obviously the more specialized you are in valuable fields, the more money you'll make. Uh but the the mix of just in case versus just in time has probably shifted at every single level. Um yeah, that's very helpful. That's that's very helpful perspective. So uh I think

[19:45] we can move on. Yeah. Yeah. So then the the second part is the career strategy part, right? So the the the the so the main statement that you made was like well I would just wait right for people to figure it out cuz there's a lot of fighting right now and I don't know what's going to win and I'll just learn fundamentals right so um I debunked that on the just in time versus just in case side of view and then the career strategy thing is something that I have preached in the sort of the learn in public uh stuff that I did before like my book and

[20:12] everything where you know you cannot have you cannot beat uh other people in years of experience in Rails because Rails has been around for 20 years, right? So, uh but you know what what you can do uh is specialize in the thing that's only been out, you know, you can't have more than 5 years experience in a thing that's only been out for 5 years. Um and so like sometimes being early in a field and being um you know, genuinely like good and right and throwing a lot of effort into it makes you a leader in that field and that's how a lot of the leaders in the field

[20:44] got started anyway. like how did you think like you know the the the sort of early uh uh sort of Django people um you know stood out like they didn't they didn't sit around and wait for like people to figure out whether whether Python will win. They were just like Python's cool I'm going to build Django

[21:01] and Django won. Yeah. So there's some risk and reward like you could absolutely have adopted one of the many frameworks that we don't even know the names of before and said I'm going to be the early mover to that and maybe they have a single conference and then they've lost adoption and uh you know it was like a nice one-year fever dream that this was going to be a thing but now it's something people

[21:20] barely even remember. Yeah. So I do think like you one there's some luck involved. Yes. But there I think there's a lot more strategy which is something I focus on. uh for people who are interested they can search my how to create luck article where I think a lot of people you can sort of break down luck and you can make yourself more lucky if you have good strategy and uh that's a nuance that a lot of people miss anyway um I uh I do think that like sincere hard work on a thing even if it's wasted ultimately is impressive to other people who will then hire you to

[21:52] say like okay like you you bet on you bet hard on a thing it didn't work out but I see the effort and I see the talent. You just were unlucky. But here, come work for me on this on this new thing. Uh I'll give you a shot because you like there's so few people actually put genuine effort in betting on

[22:08] anything. Everyone wants to wait. Everyone wants to be like everyone wants to like just jump around and like be shallow, whatever, right? But um if you just put deep work on anything, I don't care what it is. Um and and then you change your mind later, people will recognize the the sincere effort in that thing anyway. So, um I do like that and I do like the career strategy and then I think there's a there's an economic fun meme that may not be that uh funny in the sense of uh I do genuinely think that there is some amount of acceleration in tech where the people

[22:45] who for example can command tens and maybe uh in future hundreds of coding agents and do be a lot more productive will severely outstrip the productivity of people who refuse to learn anything about coding agents and therefore their economic value will be not like double like 10x 100x. Uh and that means that like the the the disparity and inequality will increase and so if you wait you will you will necessarily choose to be behind and I think the the opportunity cost of waiting has increased over time like it used to be the case in the 2010s that you could just wait it out because JavaScript doesn't change that much front-end

[23:28] development and browsers have a like a yearly refresh cycle let's say right like of like well the browser is mostly the same from year to year like it's not going change that much. Like if you're in mobile development, yeah, mobile's not going to change that much. Um, in AI, it's it's very different. And you you can't just be like, "This is how it used to be like when I was younger. Now it's probably the same." Like, no. Like, I think that's that's a very uh sort of naive assumption uh based on like not really looking enough at the fundamentals of what's going on. Um and I and I do like and I say this knowing

[24:02] that like yes I I I worry about hype and yes um some of this will will be like it is very bubbly and some of this will will come will look back and like say that oh that was that was like a hilarious like joke and all that and um people have had lots of mania before. Uh I do I do think there's enough here though that um you should probably take it seriously uh in in the sense of um I I do think like it's much more personal than it ever used to be in every prior technology shift. Um and it's much more relevant to your to your uh to your to

[24:36] your to your to your career outcomes um if you if you're early versus if you wait. And um I think maybe a last like Yeah. So, the meme I'm referencing is the permanent underclass meme where like if you don't get on board, you're you're always going to be like part of the permanent underclass. Uh meaning like you're going to miss forever and uh yeah, the fear of missing out. Um yeah. Yeah. So, I don't I don't I don't strongly believe in that, but like I see it, man. Like I see the argument and I'm not sure it's

[25:07] wrong. Let's talk about the breakthroughs. So, we know that there have been some breakthroughs. Uh, reasoning models that has been a major unlock for me. Like the quality of my LLM response has gone up through the roof. Uh, who's the guy that you had on the latent space podcast a while back? Nathan something. Uh, we've had Nathan Lambert on RHS, but also uh Nome Brown who led recent

[25:31] models. That's the guy. That's the guy. He's the uh diplomacy champion. Mhm. Yeah. I'm gonna link that episode because that guy is fascinating and his contributions are it's very funny. Yeah, the sequence is weird. Uh so Diplomacy is like this very uh unusual game that's more complex than chess or go. Um which which people understand like complexity is a game makes makes it harder and harder for AIs to solve. Uh in diplomacy you have to talk and and like uh lie and convince and and collude with your fellow players in order to win. And so it involves like conversation and so this is a uniquely like only LM could have done this

[26:07] because before this and it's it's hard to generate like convincing robotic voice. Um and so yeah no like one diplomacy with an AI with with with real human players and the human players didn't know and then after that he won as a human in person and he he won the he he was a world champion on that too. So he's the world champion both through cheating by leveraging an LLM and cheating air quotes like not doing it

[26:32] the way it's intended. Yeah. Yeah. Different. One was an online Yeah. One was an online contest. The other one was like an inerson World Series type thing. Yeah. Yeah. And because he had the domain expertise of being like the best diplomacy player, I'm sure. And also had like a whole lot of uh machine learning

[26:45] chops. He was able to combine the two. The one thing I was trying to get out of him was like, did your play improve because you had an AI sort of training you in the same way that people play chess now? Everyone plays with a chess engine because why wouldn't you? Uh, and he's like, "Yeah, of course." Like, why would you not when you have a we have a teacher that's smarter than you and perfect? Um, feedback loop that's virtuous because he can then turn and say, "This this isn't working. Let's improve this and uh further improve his own tutor." Yeah, I I do think Yeah, I I do think I have

[27:15] some minor disagreements with him on just like the the way you model other players. So, the um we talked we can get into about this about like what is next after reasoning models, but clearly reasoning models were enough for for for him to do what he did. Um what is next after reasoning models? Exactly. Right. So uh reasoning models are actually a little bit old had now like they were they were first introduced end of 2024 uh with 01 um and then we spent most of 2025 uh basically deploying the reasoning paradigm to effectively GPT5 Gemini 3 and claude 4 and now we're like at the

[27:50] limit where like look it always takes more time to reason and sometimes you don't want that like it just reasoning models are very slow like my girlfriend's using uh Claude Opus in cursor and she complains now because human wants are infinite that every time she types a new command she has to wait uh for for code to come out and you can upgrade to accelerator it's like openi did a deal with cerebras to 10x their inference speed but you can eat that 10x um as quickly as you got it uh this is called worst law which used to be called gates's law like software will take away

[28:27] all the performance advantages that better hardware will give you Right? Like we have much better laptops today than 10 years ago, but I don't feel my computers any faster, right? It's because developers can be lazier. They don't have to optimize their code as much. That's a that's a that's a desire. That's an example of human wines being infinite. Like like like I want my

[28:44] developer experience to be better, too. So, I'm going to build an Electron rather than build a native Mac app because like the Electron's good enough and therefore I ship a Chrome instance in every single app. Uh and that it eats a ton of memory, but like my life is easier. you don't know this as a as a consumer because your your machines get more powerful anyway, right? Right. So that's workflow if anyone's looking at uh show and uh electron amazing tool uh that just dramatically reduces the need to have a lot of different you know platform specific app implementations like you don't necessarily need uh one

[29:18] for Windows or um one for Mac or one for Linux you can just have an electron and it essentially leverages like the V8 engine from Chrome and like all that other stuff. It bundles a Chrome browser into every app. This is why every app is now at least 300 megabytes maybe sometimes a gigabyte uh where it used to be 30 or 50 and

[29:38] your computer can tank the hit. I guess I don't know performance like I got like 2,000 tabs open on my Chrome browser like I I have every app open because Mac never closes any apps and like I I run into memory outages all the time, right? And and so anyway, um what was I saying here? Like I so so improvements uh beyond reasoning models. Okay. So there's there's three uh that people are focusing on for this year. Uh and obviously I stay close to the research community. So I'm I'm relatively aware. Uh one is in the general category of memory and continual learning. Uh because LMS don't uh by

[30:12] default learn they they just always inference from static weights. Um two is let's call it world modeling where they are uh understanding theory of mind and playing out different scenarios and being more efficient in their learning. Um and I think three is just a general category of multimodality where they used to be very focused on text and now they can operate computers and uh have have like very good video uh understanding and vision is split into let's call it OCR which is the most classic form of vision like recognizing handwriting that's what emnest uh used to be that is like the very very first application of computer vision to uh

[30:52] screen understanding where they can watch your screen and like operate a computer for you and that's a uh I've been very enormously impressed by cloud co-work uh this is called computer use and then the last part of vision is embodied uh where um you put it into a robot and it can uh have this what they call egocentric view of the world where um you can sort of uh understand what you're what environment you're in and manipulate objects in the environment u because obviously like a lot of AI is like contained within hardware uh software within your laptops and within your internet um once it starts

[31:26] impacting uh physical worlds, you can start doing dangerous jobs. You can start doing manual jobs uh that nobody nobody wants to do. And so like that's that that's very important. So I think there's this sort of these three directions and we can go in order if you want. Okay. And so just uh to um like LM by default they don't learn but you know

[31:47] giving them the capacity to truly learn. I mean not learn the way a human does but like just like a submarine doesn't swim but it does get through the water just the same right uh you know so basically like that that kind of learning capacity is one the next is world modeling where they understand like they have kind of a theory of mind and theory more broadly philosophically of how things fit together understanding of physics and time rather than just um predicting one word after

[32:12] the other. Yeah. And then the uh third one you said is multimodality which you know incorporates you know their ability to uh you know like traditional classical computer vision optical character recognition uh screen understanding which I'm very excited about and uh you've been impressed with cloud co-work there and uh then of course like the embodied so this is also part of the multi modality but embodied and you said egocentric view of the world which basically like viewing the world in first person kind of like trying to reason about your place in like which could be a literal physically like where am I in this warehouse for

[32:49] the purpose of a robot. Yeah. So, so those are the three things. Uh yeah, let's just let's rip into those all three of those. If you can walk through some of the big work going on in all three of those because unlike most of the guests on the free co podcast, you're spending a lot of time actually pouring over these research papers that are coming out of institutions like Stanford that we were just at the other day. Um yeah, bring us up to speed. give us the skinny. So, so you know um I you know one of the questions that we prep was information diet. Um I as part of

[33:19] inspace I interview people uh who are like the researchers and the the founders in the field. Uh and obviously I I have to prep for those conversations but I also run the lane space um discord where uh people actually share links that are relevant to each of these fields that I just mentioned. Um the discord meets once a week uh as for the the inspace paper club and there we go over these papers that that people are are excited about and and and having some impact in in the field. Um and so and obviously for AI engineer I field uh thousands of submissions of of talks and

[33:57] like have to like pick like well who's legit and who's just trying to sell their company and I think like yeah so so in terms of memory um and um continual learning this is like quite an open question. uh a lot of the database companies sell themselves as improving memory and that is a uh the sort of systems approach right like there's a very classic divide in AI on whether you improve the model or you improve the harness or the wrapper or the system around the model and uh a lot of the last 5 to 10 years has been just improving the model you we figured out

[34:32] how to scale it and we need to keep scaling it and every time we scale it um some part of the harness above it falls away. Um, I do think like basically like we've run into effectively the limits of context like we can't really extend the context lengths of LLMs a lot further like and and I obviously I might regret saying this but we've had million token context windows for two years now since Gemini first introduced it and no one uses it. Everyone stays under 200k. Um so like if I gave you 100 million that's just practical reasons because the context rot and the fact that the

[35:06] more context you feed it generally the lower the quality output will become. So, so yes, that that is that is up there uh in in the number of reasons, but there's also cost reasons, right? Like you are charged for every single context token that you put in there. So, like why would you voluntarily make it more expensive for yourself? Uh and then there's also just order of magnitude reasons, right? Like uh sure I have a million token context window today with let's say let's say I give you a million token context window today with zero context. It's perfect recall across all over, right? Absolutely solved one

[35:34] million. Then I then I come along and go, you know what' be really great? 100 million token context window. Uh so you just 100x and like also perfect context ro like not you know uh amazing great but I'm sitting over here I'm enterprise and my document uh database and storage and nodes start at 8 trillion tokens.

[35:55] What is your 100x going to do for me? Nothing. Absolutely nothing. Yeah. So there's no economic incentive to focus on that aspect of improving. It's just like it's just wrong. It's just like not useful. Like you're scaling a thing that like like it's great for research, but like you know look in any practical enterprise uh thousands of people talking and working and producing documents like are are you like are you really going to sit here and tell me 100x is enough? No. Uh you know so like you're going to have a different way. Uh and so like I I I I think that is the bullc case for the

[36:31] sort of database memory people. And then the machine learning people obviously hate this. They're like what do you mean it's it's so offensive like what do you mean you tell me I have limits you know the machine learning people are like okay well we'll just keep updating the weights and having pluggable architectures and all that and so that's what they are working on and at some point you know probably the right answer is both of them will win in different elements and I have to sort of like be the neutral Switzerland arbiter of like yeah like you guys have good points and then I turn around and go like you guys

[37:00] also have good points and like the truth is like they both need to believe strongly in view of the world in order for the world to sort of arrive at like something better and like we won't truly know what it is but I can see both sides. Okay. And and just to be 100% clear those both sides are you need to improve the harness the application layer versus you need to improve the capabilities of

[37:20] the underlying model. Yeah. And both I mean the answer is both need to be improved but there's no reason why they can't be concurrently developed and and built out. I think like to to to dedicate your life and career to working on something, you do have to believe the other side is probably not going to scale, not going to work. And so like yeah, people people are more extremist just because they have to to to like really uh create those innovations. And yeah, if they were more pragmatic, they might slow down and just wait to see, right? That's what we've been talking about the whole time is like why not

[37:54] just slow down and take a breather and let somebody else do all the heavy lifting and then see whether that's the the road that we actually wanted to drive down. That's a great way to be a follower the rest of your life. Yes. You know, at some point you do need to actually have some conviction to uh to jump in,

[38:09] right? Yeah. Uh so you've mentioned your information diet and it sounds like you've got like a fire hose of interesting uh information coming from uh your discord like so these three different modalities you talked about uh that are uh improving uh the um the three different like ways in which AI is moving now that we we had the initial just utility of the foundation models and then we had the reasoning models and now we have uh just again to recap those three because I think it is important the world model um the uh learning capabilities and the multimodality multimodality uh like are how would you say like the

[38:50] the research and the enthusiasm around those three different things like are are there people that focus just exclusively on one of those areas and like they're mostly sharing articles or research focused on that within within your community or is everybody like equally excited about all three? Um, pretty much only the content creators like me are excited about all three because we don't have to make bets, but other people, yeah, they do have it to make it their entire career, right? Um, so yeah, there's some of the smartest people in the world are working on each of these things. All of them have at least $100 million each to like

[39:24] go pursue whatever their their view of the world is. Um, yeah, very very smart people working on all of them. There's a fourth sector, by the way, which I haven't mentioned that we've made a big bet on, but I wouldn't classify for a generalist audience, which is AI for science. uh the the the main idea is that um general office white collar knowledge work manipulating Excel files, powerpoints, scheduling meetings, taking meetings, what is all this is mostly saturated and AI can already do it. Uh and the next frontier is human expert scientific work which is both verifiable so that's good for training data for LLMs. Uh but also like it would

[39:58] generally advance humanity and and I think that's something that like this year people are investing a lot more than last year. So we've actually gone gone ahead and started a second podcast. Our first podcast was focused on uh software engineers. Our second podcast is focused on educating people on the movements in AI for science by because I got two actual scientists hosting the

[40:20] podcast. I don't have to do it myself. But um I do think I I I do think it's like um an an important shift that most people will never really touch, but you'll see it trickle out. Um you'll see people really mention like here's how we're sequencing the immune system. So we know how to create like uh and solve diseases that as rare as one is is the is the tagline from the Chan Zuckerberg Institute from the Mark Zuckerberg podcast I did. And um yeah like the vast majority of human diseases have no FDA approved treatment. Uh and like the the understanding and solving the human body

[40:53] is something everyone wants and it's extremely complex. It's probably the most complex system in the universe. Um and we don't understand how our own human body works. That's like absurd. Well, I mean it's it is certainly complicated to be fair, but we have only had like science and medicine for a few thousand years. Yeah. Yeah. Yeah. But it's it's like it's surprising how like uncurious I am about how my own body works. I'm like, well, I'll worry about it when I'm when

[41:19] I'm sick. Yeah. But like thankfully there are doctors and scientists out there who are proactively but like, okay, where's my free coat for like learning how my body works, right? Like where's my just in case learning of like, you know, like sleep is pretty important. Like this is this is what we're reduced to. Like you know you should sleep eight hours and then like everyone goes like yeah yeah yeah and then they don't and like you should drink more water and then they don't and

[41:42] like this is how bad it is. Well I mean they don't appreciate because they don't have all the context that like a physician would have. Like my you better believe my physician is telling me to drink more water and I'm like isn't that just like some don't I get like half of my water requirements

[41:55] just from breathing humidity in the air? He's like no drink more water. Like not just tea. I know you like drinking tea. Uh not just decaf coffee cuz I had to switch to decaf. Um okay, I was too hyped up. I was too high energy. Uh speaking of keeping this high energy, uh I do want to talk about the uh the world models real quick. Yeah, that is fascinating to me because LLMs the the biggest, you know, knock against them that like comedians throw out and all these like people like if you ever heard of it just predicts the next token, no big deal, right? Like but it doesn't

[42:28] actually know anything, right? uh how does that work? How how do you actually train an LLM which is essentially just like a token predictor? Uh how do you train it to actually understand the world around it? And again, I use understand in air quotes because no machine is ever going to approach most likely approach the world the way that an organism approaches it. But we can get similar effects through through these approaches. So I don't want to be semantically splitting hairs and stuff

[42:54] when I use the term learn for example. Yeah. So we we uh we published a post on this recently that people should read and I'll put in the show notes. U but they're uh the first thing you should know is that world models are already an overloaded term and there are three kinds of world models and people are ignorant of the kind that they are not working on. So you should know who you're talking to when you say world models because depending on the person you're talking to, they will completely assume that the only kind of world model is the one that they know. So the first

[43:23] world model is the video world models. Uh those this is like Genie 3. Uh Tesla also has one and OpenAI has one that they haven't released yet. Um but like uh basically these are the video like you you sort of inherit you sort of come into a world and you can look around and you can see like 3D objects placed in

[43:42] space and you can manipulate them. Functionally it looks no different than a game. But um what's different is that everything in there is generated by an an AI. Um I do have to make the distinction that uh a lot of these models are not LM based. Uh so for example we did a episode with Fei Lee who is the creator of ImageNet uh that was the that basically led to uh the ImageNet moment Alex net that eventually led to Transformers and and the the the whole deep learning era was started by ImageNet and and Feay was the progenitor of that. um and she's working on spatial

[44:17] intelligence and she has a really good essay uh everyone go look it up on like why LLMs are limited and why she's working on the next frontier of world models and spatial intelligence but her world models again also nerfbased um uh video sort of 3D world models and they're very very different from the other two that I'm about to talk about uh but you should be aware that like most people in AI uh AI research when they say world models because of fee because of um gem uh because of deep mind's research and world models, they usually mean videogenerated world models. It's kind of boring. It looks

[44:54] like a video game. They're all like it's no different. But I I do think that spatial intelligence is very important and we explained it in that podcast. The second one is the Yan Lakun School of Thought, which is who's from Meta. He just left. Um but he's been very much trying to basically have this like Aristotalian or Plato's cave view of the world where like um you know uh models learn to predict text. So they they're rewarded and punished on the their ability to predict the next token uh or the next series of tokens. Um and um and that is like the representation of the world, but the map is not the territory,

[45:29] right? like they're they're like I I'm I'm rewarded based on like the things you output as a published final article but I I actually need to learn one step before that which is what you're thinking before you wrote that article but I don't have access access to your thoughts so um what they what what Yangun has is the joint and um join embedding prediction architecture ja uh where you basically project or you try to learn a common latent space uh among your uh among your embeddings among the the things you're trying to predict and uh that is the the correct machine learning pill way of doing this

[46:05] which is saying throw your hands up put it in a black box say I don't know what this is but you machine go figure it out all I know is that there is a black box between um what what you are your input and what gets output which is the thought that is somewhere somewhere in there and the reasoning models try to do this very very crudely by u having a chain of thought um piece which uh which is effective like a piece of research that's been done since like 2022 let's say uh from Jason we uh people can look up the the original chain of thought

[46:37] paper that is like that is all reasoning is which is very very amazing that it's scaled this this much uh and then I think the the final piece um is what I argue in that in that post uh that I put up in the show notes which is the adversarial world models where you have a theory of mind and you try to form like uh you you're sitting at a poker table or you're sitting in a negotiation or you're sitting even in this podcast interview. I'm trying to model what you think. I'm trying to model what uh the audience is thinking and I'm trying to respond in a way that uh elicits the

[47:08] information or like transfers things in information in a way that I think will suit you the best because I have a theory of what you are interested in. I have a theory of what the audience is interested in. I have a theory of what I want to say and I'm playing all those world models against each other in order to arrive at um what is optimal for me to say. And so, uh, that article as a guest article, it frames it as a sort of adversarial negotiation thing, but I think it can also be collaborative in a way that like we're both trying to get to a better understanding of what's

[47:40] going on in the world and I'm trying to articulate as much as possible and you can chime in with your point of view and I can like disagree or agree. Uh, but like that is how intelligence works. It is not by like trying to produce like the median expectation of what a good article is based on a prompt. It is me like having my point of view, you having your point of view, me understanding your point of view and probably getting a little bit wrong. You you probably getting my point of view a bit wrong, but we talk and we we figure

[48:11] it out. Um yeah, man. Uh a lot to unpack there. So, uh first I'm just going to recap what you said. Yeah, I take detailed notes as I listen to all my podcast guest but here certainly uh typing a lot. So video world models like G3 uh essentially like 3D objects in space. I've seen like a simulation of like this famous uh Counterstrike map called like I think it's called like

[48:36] Dust 2 or something like that. Dust Yeah. or something like that, right? And uh basically like you're walking around and it looks like Counter-Strike, but it like sometimes like a box will be like slightly out of place. You'll turn and you'll turn back and it'll be like all warped and stuff and the model starts to break down after

[48:51] a few minutes. Correct. So that's an that's a bad video. Uh yeah, that's a bad world model, right? So So video games have perfect world models because they're they're programmed in by hand. Like every every detail of that map was was like I don't know if like exactly hand, but like whatever like some human intended it to be there, right? They were taking like assets and positioning them in a certain place and

[49:12] it was being stored as the map. Yeah. Uh the video world model gets inconsistency. So it's it's like exactly exactly what you say like it's almost like a horror movie. Like I look at this screen and I look away. I look back and like things have changed and it's horrible. But that's what generative models do right now because what and why is that? Because well when I turn back from that previous screen I have no memory of the what's what happened before. and what I render as the as the new look is just as likely as the previous one. So like it's fine to me as

[49:41] a generative model, but because I don't care about the world model, I've never been incentivized to have a consistent world model. Why should I care? Okay, so that's world models. And uh your argument is that like a lot of those things that have been showed off so far with like you know generating a video game as opposed to you know having everything you know compiled and rendered and and following deterministic rules laid out by a programmer like that is obviously a work in progress. There's a lot of work to be done there. Second thing is you talked about uh like Plato's cave view of the world. Yan lun

[50:16] essentially like he said uh you have to like blackbox everything like if you're not like throwing up your hands and just accepting like hey I don't know how this works but this ML is this model is going to figure out you know kind of f from first principles by spelunking in and like figuring it out essentially. Um they're going to go into the latent space and figure it out. I'm not sure if I'm using that term correctly. I know

[50:40] the name. Yeah, this is this is horribly abstract for people. Anyone who's never dealt in latent spaces and never dealt in hidden layers in machine learning. Um, but just trust that machines do learn much better than us if you let them. And this is Yan Lakun letting them. Yeah. Getting out of the way and not being too uh I guess uh prescriptive uh and not trying to put it on rails to achieve a specific goal, but like giving it more

[51:06] slack essentially. Is that accurate? Yeah, I should also mention that this is the least successful one. Uh like basically no one outside of meta believes that this is going to work. Um and in fact within meta like he had so much resistance that he had to quit and go go start something. Um so like this is a dark horse like you know I I've learned to just never bet against like legends in the field like that. Sometimes they are right, sometimes they're very wrong but he has enough conviction to keep going. Um so it's yeah by by nature of of his advocacy of that kind of role model he gets a

[51:37] mention even though no one as you said earlier you have to have an opinion you have to stand for something uh otherwise there's no progress to be made if everybody the moment something starts to work there they all jump ship and just jump on the bandwagon right so it is important to have people kind of like pushing their own you know darkhorse agenda uh to try to improve the field one adversarial world models essentially

[52:00] like a salesperson type mentality. You're trying to figure out what the person needs to hear to sign on the dotted line. Uh like through metacognition. Also, you could say like a teacher at the front of a classroom. Exactly. Looking at their I don't like the word adversarial. Yeah, I don't like the word adversarial. The Cuz like adversarial you think it's like is this only useful in like poker or like no, you know, thermonuclear warfare or something like that. No, it's it's it sounds like it's just like basically shaping a message and it's just good

[52:26] good communication skills. I I think it's multiplayer is what it is what it is. A lot of most most environments in LM are single player. Um, and so then you then you need you need that. But uh the player doesn't have to be human. The player can be other AI, but the player can also be an inanimate object. Uh the player can be this this Riverside session we're in. The player can be this this microphone. I need to have a model of what the microphone can do, how far it is away from me uh to know how I sound, to know that he has a

[52:53] mute button, all that kind of stuff. These are world models of everything in my life. Uh, and they all they're consistent because they're just facts and I need to like plan around them and be intelligent about them. And most LMS will will just don't have that because they've never been taught that, right? They they've only been taught to predict the next token. And there is a there's a there's a so there's a fundamental question here of like do we need to change anything about what we do. Uh because um and we get to to into this in my discussion with Nome Brown. Um and his assertion because he's he would say

[53:26] that is everything can be modeled just scale up the model give me more compute give me more money and we will build implicit world models by just by just training it as we have done as as history has shown um that uh you know just just scale up the reasoning and it'll it'll work. He would say that because he leads reasoning right um but obviously he's smart enough to understand that there is a counterargument which is that it will never scale to what we need. Uh and so the argument in that essay was that like he's wrong. Uh which is a which is a both the cost of trying what he's saying

[53:58] I mean as you scale up it gets more and more expensive like not in a linear manner right uh and and so there's there's great risk like we could literally take all the compute and all the energy and all our funds to build data centers to try to test that and if we're wrong we just

[54:16] wasted an immense amount of resources. Yeah. Right. Um so it is risky to pursue that. Yeah, I'll give you a simple analogy which is like um uh so don't don't take babies but like maybe take um when you come in when when you're um when when let's let's say let's say you you you finish college, you finish high school, you you you you start coming into work, how long does it take you to learn your job? You know, a few weeks, right? To to like at least like start start doing start doing your job. uh a model would have to require the equivalent of the data of hundreds of thousands of human

[54:54] years of data just to do a fraction of what you do. Um in other words, this is a data efficiency argument like models because they are so inefficient at uh at at learning um on specific tasks and domains because they don't have don't have world models. They are so much worse than us, right? Like we learn our jobs from very few shots. Like I I see I see one dog. I see two dogs. I see three dogs. From there I

[55:20] got it. I know what a dog looks like. Models pretty much have to see more than 20 dogs at once to like at least have like some vague image of a dog. Probably thousands of dogs. Um and like when when when I when I see a dog with four, you know, all my examples are dogs with four legs and I see a dog with three legs, I

[55:35] don't go, "What what the hell is that? I've never seen that before." No, like it's a dog with three legs. Right. And like um so I think like that that is a model of we know what this abstract concept of dog is and models don't. They just have no preconceptions whatsoever. They don't even try. Uh like every new concept is potentially as alien or as as familiar as as as as it could be. And it's it's very very inefficient. Uh, and so there's probably some uh some efficiencies to gain from modeling it closer to how we know intelligence to work because we have the existence proof of the human brain being

[56:13] able to do all of this. I can model all of this on 20 watts of power in my head whereas they need a giant data center and 500 gawatt. Yeah. Yeah. That is interesting to think about and it it just hints at the sheer improvement possible because I mean these models are so incredibly

[56:33] inefficient compared to the human brain. It's not it's not even funny. Um but look, so we we at least as a species are like a few hundred thousands of years old and they are 10 years old, right? Yeah. And and once they figure these things out, they'll be able to operate they'll be able to run circles around us. But until then, you know, uh we have a lot of work ahead of us if we want to get them to that level. Um so we have learned a lot here. I want to fire off some rapidfire questions because I only have you on the podcast like every

[57:03] two or three years. Maybe that'll change if people want me to have enough back. Let me know like but things are moving fast and I'm learning a tremendous amount. Uh I do have some quick questions for you. Uh if you got hit in the head and suddenly lost all your programming skills, Yeah. Yeah, but you didn't lose any of your other skills or your knowledge of like I guess the world and like where you know you think things are going. How would you go about rebuilding your programming skills and what would you focus on most given

[57:35] the improvements in the tooling? Yeah. Uh definitely head to free co. I love free co. Um I do think like understanding some amount of like how computers work is more important than actually programming every single thing. Um so understanding that there's such a thing as an if statement you know like memory and like how that gets stored to a file system that kind of stuff like um that is definitely the in the category of learn it uh just in uh case you know foundational prep um whereas like you know stuff I used to have to learn like how to configure Webpack for React um doesn't matter like I spent so much time

[58:18] on that stuff like it doesn't matter anymore and the LMS can teach you just in time. um or you don't even have to teach uh to learn them like you can just vibe code it and I do think that that will go away but um I do think like the the sort of free resources are great um you know for those who don't know and uh you you you can listen to the previous interviews but um I changed careers when I was 30 and I learned to code on free code camp and you know and I I do think like that is now as fundamental a

[58:46] requirement of literacy as learning math and learning um I guess Yeah, I'll put it at the level of math. Let's let's call it like it's not the not at the level of learning English to just function in the world, but it's pretty close to learning ma like it's pretty close probably more important than learning math right now. And uh yeah, I think like to the to the to the extent that um you can teach it like I forgot all my programming knowledge and I want to learn I think the sort of free high quality resources are great and free code camp is like top tier. Uh, and then

[59:18] also just like start building a bunch of stuff and then learn just in time because my interests are going to defer from others and um, and you know like I I can't consume all free code resources forever, right? Because that's that's a great way to just rot in in YouTube hell. Um, I just have to get out there and build

[59:37] and so I I would do more of that. I'll emphasize that since you went through free cooking many years ago, uh, we've taken an extreme project oriented bend to our uh, pedigogy. So now we basically always uh everything is in the the context of a project and we have tons of projects. You if you go through the entire full stack curriculum you'll build more than 100 different projects because like 100% and that's a lot a lot of that's informed by you know learning public and a lot of the other uh important kind of educational movements in software development over the past 12 years since we started. And uh yeah,

[1:00:14] 100% on just having a project and building and that's one of the things you've emphasized a lot in this conversation and in your uh you know the many articles you publish when whenever you do talk about uh learning programming you talk about uh the the need to learn in public and to to fail a lot while trying to dare greatly with projects. I I will say so it's been since eight years since I wrote the original learning public essay and I do think like the bar for like what you publish has has risen because a lot more people have done it and so uh I you know then I've developed like what I

[1:00:49] call learning gears where like you well not everything should be public in fact most things should not be public uh so I think like people try people get learning public religion and they publish everything and they're like why isn't this working well cuz like it's crap and like no one no one's like everyone's too polite to tell you it's crap and you kind of know it's crap but like people want to be encouraging and

[1:01:06] whatever. Um, so but you do risk fatiguing people to your cause. Exactly. If you're constantly overwhelming them with mediocrity. No. And and then people blame me because I encourage them to learn in public and I'm like no no no um I had a writer tell me once throw away your first million words. That was his writing advice. Yeah, there's there's there's a lot of the opposite of learning public if you

[1:01:30] think about it. Oh just don't not everyone has to know everything about your life. you know, we're not that interested. Like just, you know, put it on your personal journal as people have done for centuries. Like not everything has to get a thousand likes. Like it's fine. Like do it for the sake of doing it. Doing for the joy. Do it for the learning. Uh and then every now and then stuff will break through because you you are producing and that puts you ahead of

[1:01:52] 99% of people who don't. Yeah. 100%. So uh let's talk about building companies which is something you know quite a bit about. Um Oh god. If if you wanted to uh create a startup these days, like let's say you were just starting over, you you had the skills, the knowledge you have, but you just wanted to start uh and and like maybe you would want to c coach some of your peers in doing the similar thing like have the costs necessary to build a startup gone down? Have the tools improved? Uh like what what is the current status of like I want to get a startup going? Like do you need to get

[1:02:29] funded? Do you need to get into some accelerator program? Uh the answer is no and it was always no. Um I do think the average cost of getting to a million dollars in AR um is is much lower. Yes. Uh so I I have a thesis called the tiny teams uh playbook which uh I've been exploring and I I you know the the original phrasing of this question was um is like a a lot of people are focused on like the one person unicorn like can can you get to a billion dollars of one person and you know my push back is like well do you

[1:03:03] need to be one person? Can you be two? Can you be three? And like sometimes it's nice to work with buddies like the only reason the only reason you focus on one person is because it's egotistical and you like to they like this image of the lone hacker but act it's nice to get there with friends and it's it's it's uh and people have different strengths and like um what what what really matters is that people are a lot more productive than they used to be in the past and that is true and that is enabled by AI um and that means that everything changes around funding and economics

[1:03:31] right people are profitable from day one therefore they don't need as much funding So why would you raise other than like to cause some splash in the news? Um if you're selling to enterprise, it does help to have like Sequoia's backing and all these. Uh and that's why people raise money that they never use and they proudly say like we haven't even touched our series B. Now we've raised a series C, another 100 million. And you're like, well, you didn't need the money. You just said you

[1:03:55] didn't need the money. And that's true. That's just all signaling. Yeah. But it is a big splashy PR win if you can raise a bunch of money. it it signals. It's signaling to the it because like the the startups are very shaky even as a series C series D like later stage startup and so if you're a Fortune 500 you're like well do I want to be business partners to these guys who might not be around next year. Um saying that you have $100 million in the bank puts away a lot of those concerns,

[1:04:22] right? Um so like that's why you do it. Uh because you want to become one of the larger businesses. And what does large business mean? Uh I one of my favorite interviews that I listen to uh uh if you want to if you want to like get in like top tier CEO tier like listen to grit from Kleiner Perkins uh but this guy Juben Merza Dicken I've had him on on lat space as well he's great uh like he had one of the CEOs he goes like yeah your your your your business only really starts once you hit a billion dollars in revenue like talk like below that you're

[1:04:54] just getting started uh because these guys you know working like hundreds of billions of dollars in in in the They manage very very large businesses, right? So to them every startup is tiny. Every startup can go away. Every startup is like uh you know inconsequential and like you know unreliable and like instead of relying on them we should just build this in-house, right? That

[1:05:13] that's the default. Yeah. I mean not becoming too beholden to a company that can go out of business. Uh Freo camp uh you know we we used to use Gitterter for example uh as our kind of like Slack alternative uh for for chat. It was an open source GitHub ccentric uh chat tool and of course it got acquired uh and then ultimately shut down and these things switching to different chat tools because they keep getting acquired or they keep going out of business and uh yeah like I I can definitely empathize with that and yet you don't have to worry about like you know oh I guess you even have to

[1:05:47] worry about like Google discontinuing some service that you depend on if you're like built around Google reader you have to worry about Reddit like you know changing the dynam the economics of using their API and putting a bunch of accessibility tools out of business. You know, like like I can definitely see the instinct to build internally and of course developers always uh you know say oh not invented here that's not a good thing. Like you you don't want to be like constantly rebuilding. You want to focus on what makes your product better, right? And not just building a whole bunch of your own tooling. But now with

[1:06:18] LLMs, this is a very hot topic right now. It is in theory possible for you to build more and more of your own stuff and rely less on buying like the build or buy equation has changed. Uh has it changed? Yeah, it totally has. I think um how has it changed? I mean yeah, people are building more instead of buying and that's that's great. I do think like um again back to the theme of like human desire and demand is infinite. Um, you know, just because more businesses and enterprises are building in-house because they can, uh, doesn't mean that you as a tiny team can't also just build

[1:06:57] your own stuff and sell to people who would otherwise not being served. Uh, you know, to me the best example of this is lovable, where they basically like put cloud code, it's not actually cloud code, but you know, cloud code in in a box and put it on a website and like gave nontechnical marketers the ability to make their own websites without uh going through a developer, without using a templated thing like Wix or Squarespace. Uh, and the marketers

[1:07:21] freaking love it. They love this thing. They they used to spend thousands and hundreds of thousands on on developers and contract agencies and all these things and have like really long development cycles and like a product that they didn't actually really like that much and now they have direct control of their over building their custom website that is fully custom, not a Squarespace

[1:07:41] template, not a Wix template. Uh and you know and like lovable came from nothing and now it's like one of the biggest companies in Europe and like more power to them. Like I think like that kind of stuff whereas like it's mostly uh a sort of tiny team success story. I don't even think they're like

[1:07:56] 100 people yet. Yeah. And they're worth like six seven billion. Um yeah, it's definitely an outlier but but it is a very impressive. There's many smaller teams. Um, I had OLED on my podcast and and at AI Engineer, it's a team of four and they they make uh they make like app store apps that are like student aids, teaching aids, uh, homework help, and they market through Tik Tok and they

[1:08:21] make a million dollars per person. Yeah. Right. They'll never be a venture scale business, but they're very happy and they're they're making a lot of money. And what else do you need? Yeah. Not everything has to be a venture scale business. Uh so uh I want to be mindful of your time and our audience's

[1:08:38] time and get to one final question. We've learned so much, but uh there's a lot of conventional wisdom floating around around working in tech. Uh a lot of it may be outdated. What are a few pieces of conventional advice that you think have outlived their usefulness that people need to discard? Yeah, we we went over this in the prep and I was like I don't have an answer. I I I I also I don't know if I know what I

[1:09:06] would say there. I I do I do think like we probably already started this podcast there which was the learn your foundations which I used to be a big learn your foundations guy by the way. I I I used to tell people hey every now and then like once every six months just sit down and like pick a technical book read it from cover to cover. You probably learn things that you thought you knew that you didn't know. uh you you yeah you probably learn things you didn't know that you thought you knew and uh I think that's still like that's rare like I I haven't done

[1:09:39] that in a while now but I used to uh by the way that that was like my first interaction with Boris Churnney who uh used to work on TypeScript at Facebook and now is the creator of cloud code uh which is a $2.5 billion business in inside of Enthropic. uh he wrote a TypeScript book. I sat down and to learn Typescript I read it from cover to cover and I learned my fundamentals of TypeScript from him and that was great but now probably no longer needed. Uh so like I I do think that the sort of learner fundamentals has reduced a lot in comp in accidental complexity to just

[1:10:10] essential complexity of understand how these systems work. Um, and uh, I I do think I I do think like that makes you unlock things that nontechnical people uh, would otherwise be stuck at. Um, I'll give you I'll give you like a really interesting example that I had from today. Uh, can I screen share on

[1:10:30] this thing? Um, yeah, go ahead and narrate for the people who are on the video or the audio only version literally just happened today. You know, like this whole this whole thing like people want to sell you this vision of like, oh, you can vibe code everything, you can just prompt it. Like, no, no, no. There if you're a technical vibe coder, your technical prompter, you can do things that normal people cannot do, right? So, uh here I have uh here's me talking to Devon, the the coding agent that um I I I do work on as part of Cognition. Uh and it's it's like, let me

[1:11:00] check all my social links like, you know, I like I put in a bunch of speakers for my conference. Like, I just want to make sure some of the social links work. a lot of them are hallucinated and they lead to a 404, right? And blah blah blah, check it. And there and and he and the bot comes and says like, I I corrected all the Twitter links, but I couldn't correct the LinkedIn links because LinkedIn blocks

[1:11:19] bots. Um, and it stops there and I'm and it would have stopped there for any nontechnical person, but I as a technical person, I know the next prompt, which is get more creative, spoof UA, spoof the user agent. meaning, yeah, it blocks you, but because you're identifying as a bot, just identify as a

[1:11:38] human. Then he's like, blah, blah, blah. Okay. Then then it just fixes everything. Just violate the terms of service and you'll get there. Yeah, I mean, it's doing more for me as a like I would have to do this manually, but now I can just prompt a bot to do it. But it's not it's not like you're like systematically like scraping them or

[1:11:53] anything basic information. I like it stopped here and a nontechnical person would have stopped here because they don't know any better. But because I'm technical, I know the magic words, maybe spoof the UA. And from like those three words, I was like, "Oh, yeah, yeah, I could do that." And then it just immediately does

[1:12:11] it and get completely everything right. And I'm like, you know, like that kind of thing you probably have to learn just in case uh and like have fundamentals and foundations, but it's a different kind of fundamentals. It's like knowing magic words of like how the system works, not knowing specifically how to do it. like I don't care specifically how to do it. I just know like like this is this is how I would work act as an engineering manager to to a junior employee. I'm like I know the rough magic words to point you in the right direction and then you can figure it out from there. And that's what exactly what

[1:12:41] I did with the coding agent. And so I do think like you still need fundamentals but it's like a systems level understanding of like and like the magic keywords to like prompt models in the right direction which is what you eventually needed as an experienced engineering manager. you're just pulled forward five years in your career. Yeah. And and just a final like quantitative question. Uh before when I talked to you uh in San Francisco, we met, we recorded an episode like two years ago. Um you said AI engineering is about 90% fundamentals and 10% knowing how to use uh AI tools. Would you say that that percentage has changed

[1:13:21] in Yeah. Yeah. So I can't I disagree with my old self. Yeah. Okay. How how what what are the percentage distribution now? Um let's call it 60 40 60% fundamentals 40% uh yeah so so still majority fundamentals but a lot less and like and do you think there will come a day where like an expert is no longer really needed and like domain expertise isn't really needed and it's just being really good with working with LM tools and you'll have these generalists that just

[1:13:50] use AI tools. No, experts are Yeah, experts are always needed. uh until we solve the three things that we talked about world models and memory and all those things. Um awesome. Uh but you know experts are always needed but you know they they may just be a lot less of them. Like we used to have a lot of expert like farmers and now we have a less expert farmers but who do you call when your your farm goes

[1:14:10] bad? Like an expert farmer. Um there's just a lot fewer farmers and and I I do think like uh you know one of the questions that we prepped was like also you know what's the job market going to look like? Yeah, I do worry for the junior developer. Um I think that is a concern. I think like uh the the smart ones, the high agency ones are always going to figure out a way, but the mediocre junior developer who like is just like barely scrapes by, barely learns a tutorial, and like doesn't know how to generalize beyond just a tutorial, you're in trouble. You you do

[1:14:38] need to step up. Um and like that's just how it is. Like you know, um you no one owes you a living. You got to go earn it. And um you know, the tools of the trade are changing very very very dramatically right now. And uh you like no one owes it Quincy doesn't owe it to you to teach you. Like you have to go

[1:14:58] figure it out yourself. Well, I'm going to do my damnest to teach you anyway. Uh and so is the free cocaine community, but I 100% co-sign that sentiment. The world does not owe you a living. Uh got to figure out how to make it work. And there are a lot of people out there that are making it work, including many of the people whom I've had here on this podcast. Sean, I want to thank you again for making time to share your expertise and share your insights and and even your wild speculation. It's like super helpful, even though I often butt heads and push

[1:15:29] back. I mean, this is not scripted. Like, this is not edited. Like, I don't know how he's going to react to these questions. He doesn't know how I'm going to react. Uh, but but yeah, like this has been a blast. I hope everybody tuning in learned quite a bit from this. Be sure to check out some of the links that I'm including uh to articles that Sean has uh recommended that that he mentioned during the call. Uh this has been like super helpful for me. I've learned a tremendous amount in a very short period of time. I hope you have

[1:15:55] too. Until next week, happy carting.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt