The #1 skill to survive the AI era - with Carl Brown

Carl Brown (00:00)

AI doesn't replace jobs it replaces tasks

right and what

What we're gonna have to do is figure out basically what does this job look like when these tasks aren't part of that job anymore because the AI is doing them.

Nathan Toups (00:10)

Right.

Carter (00:18)

Well, Carl, welcome back to the podcast. Great to have you on.

Carl Brown (00:21)

Thanks for having me again. This is gonna be fun.

Carter (00:23)

Th this is where we abandon the pretense that this is about books and we just say, We like Carl. We like having him on the podcast and he and you know, we're gonna ground it in all the great books we've read, but you know, we love having you on.

Carl Brown (00:35)

I'm sure some book will come up. Occupational hazard.

Carter (00:37)

yeah.

Well, so I wanted to kind of start out with, and this is something we were chatting even before we started recording, which is you and and to some extent us, I think people might view as kind of AI skeptical voices, but that's not necessarily how we view ourselves. Would would you I mean skepticism? It's not that we're negative on the topic. What do you think?

Carl Brown (01:01)

Yeah, I

mean I I I I have been accused of being a AI hater more than a dozen times, probably this week. but

Carter (01:09)

Ha ha ha.

Carl Brown (01:11)

but I mean I I use it a lot. I I think it's very valuable. the problem is there are so much there's so many more voices out there that are, you know, this is the second coming kind of thing.

Carter (01:25)

Right.

Carl Brown (01:26)

I feel like I have to be that that there are

too few voices that are in the skeptical camp. And so I kind of have to to to to to fill that role. my I mean if this was so my thought is that this thing is primarily a search technology. Right? It's a really cool search technology, but it's not an AI technology so much as it's a search technology. We can get into that. I don't know that it really matters. But my belief is

That if this had been presented to the public as the next Google instead of the next replacement for half of the the white-collar jobs in the world,

Carter (02:06)

Right.

Carl Brown (02:07)

right? I would be incredibly positive about it and it would be being used in a much more productive way and it would be great. and the companies would be valued at one one thousandth of what they are now, and that wouldn't go over well for them. So I I I think it's a it's a

Fantastic technology. I just think that honestly, I think we're being lied to a lot about what the actual capabilities of the thing are.

Carter (02:32)

Okay, well let me let me give you kind of my experience with it. Cause one thing that I kind of kick myself with on AI is when it first came out, because I think a lot of us had been burned on like crypto and the blockchain and kind of like, this is the next big thing. And so then AI comes out. And I I remember the very first time I used ChatGPT and I remember asking it to generate like a a a talk in the vein of one of our religious leaders about like a lion and a circus clown. And like it was a fun little gimmick.

That like it could, right? and I was like, wow, I have never seen something like that before. And I kind of wrote it off and was like, Well, you know, but but what what is one supposed to do with this? And it was kind of when people started when the agentic wave came along and you know, MCP and just the ability that, you know, an AI could could write its own script and then execute it, that I felt I I kind of kicked myself and I'm just like, you know, I think if I could have groked that and understood that one day you know

AI

LLMs would be able to interact with a computer, would be able to, you know, invoke a CLI. I think I could have understood how powerful they could become. And so when you say, well, this is the next generation of Google, I think that is something that is quite different from Google, is the fact that they can operate a computer. They they can call an API, right? And and that's part of what makes them such powerful programming tools.

Carl Brown (03:57)

Well, I mean, we could have done that earlier. Right? I mean, so I wrote I mean, I actually did this on camera. before there were agents, right? I was doing some comparisons of various I don't even remember which ones at the moment, but you know, the the the the like cursor was one of them, codium

Carter (04:19)

Right, right.

Carl Brown (04:20)

was one of them, and then I think

ChatGPT was one of them. I don't know that a Claude Code was even in the list at that time. This was like two years ago at least. And I was using a site that I like called CodeCrafters, which is a thing that you basically build your own version of an open source, you know, web server or, you know, DNS server or that kind of thing. So it gives you some idea of kind of what the, the, the, the, the underlying technology, how it works and the way that the site is set up is basically it gives you.

No, here here's the skeleton of the the initial thing. You go and you comment out the thing that says, you know, okay, here's my here's my first thing, and then you check that in and it runs tests and says, Okay, that was right. You know, okay, now your next challenge is to do this. And so I wrote some code that basically would go scrape the website, figure out what the what the by scrape the website I mean scrape like the chat GPT website, right? Figure out what the output was, put that into a file.

run a Git script, check that in, push that over to code crafters, get the output back, network chat GPT, you all that kind of stuff, right? and so it was like acting as an agent, you know, before there was clawed code or that kind of thing, just because I did I got tired of copy and pasting it, which is how I done it the very first time I did it. but but I could have done that same thing with like Stack Overflow. Right? It wouldn't have been as useful, but I could have said, okay, here's the thing I want to do, you know

Go search Stack Overflow for this, grab the top answer, stick it in the file, post it to code code crafters, see what the errors are, post the errors into Stack Overflow, you know, lather inch repeat, right? It would have taken a much, much longer time before it actually got an answer and it might never have gotten an answer. but then two and a half years ago when I was doing this for the first time with ChatGPT, sometimes it didn't get an answer either.

So

I I I think it's a the the the the agent running in a loop part is really separate and distinct from the underlying AI technology, right? The the underlying

Carter (06:27)

Well that I agree with.

Carl Brown (06:28)

yeah, the the underlying, you know, large language model thing I think is fundamental search technology. the the agent part is just you know, as a Cal is it News Newport says

Carter (06:39)

Yeah.

Carl Brown (06:39)

it's you know, it's it's you know, it's just ask you know, the the agent just asks the the

Ask the the the the chat bot what do I do next, does it, gives the chat bot the output and just sits there and does this kind of thing.

Carter (06:55)

That's something I find very frustrating when with any of the press you get from the labs is cause because I feel like yes, there's there's the advancement of the models, and then there's the advancement of the harness.

Carl Brown (07:06)

Mm-hmm.

Carter (07:07)

And they they talk about them as if they're one and the same, and they're not one and the same. And I I think a lot about when we had you on the podcast, and you kind of said, like, look, every every technology kind of follows the sigmoid, right? Eventually you exhaust some resource and then

We you plateau and and you thought for large language models that resource that would essentially be exhausted was the training data. That they kind of slurked

Carl Brown (07:31)

That's

Carter (07:31)

up the entire internet, everything humans have ever written, right? but I think we are seeing like an an a a boost in capabilities from the harness, the just that ability

Carl Brown (07:42)

Yes.

Carter (07:42)

to call itself in a loop. But you can only invent the harness once, right? And it's unclear to me exactly how much more juice we're gonna get out of the harness.

Right.

Carl Brown (07:53)

Well so so the next the the new thing, right, is the like multiple levels of harness. Right. So I watched your your conversation with Uncle Bob that came out in in in our time yesterday. and he was talking about how I'm trying to remember how he said it, but something about like the the the agents

don't r have sentience and so they can't think about tomorrow at the same time they're thinking about today or something like that. I don't remember exactly how we said it, but

Nathan Toups (08:23)

Yeah,

I think I think he s he said they have consciousness but not sentience. And that yeah, there was some constraint that he gave it, yeah.

Carl Brown (08:29)

Yeah, that that

that but that we have sentience and that's why we can keep what why

Nathan Toups (08:32)

Mm.

Carl Brown (08:33)

we can think about today is at the same time we're talking about the like we're down in the code and we're thinking about how the code's gonna affect us tomorrow kind of thing, right? we despite the fact that we tell ourselves we don't multitask, we multitask way, way more than these agents are capable of, right? They're very focused on what's the next word, right? And we've got a bunch of other stuff. I can feel the fan on my

on my skin and I can hear you and I can see the camera and I I've got a bunch of inputs and a bunch of outputs. I've got all these fingers going. if you want to emulate more of that kind of thing, what you do is this multi-level harness thing, right? Where you've got one agent that's doing the planning and then it hands off stuff and says, okay, you go write this file and you go write this file and you know so on and so forth. and then you've got one agent that's like, you know, hey, hey, hey, hey, hey, I'm thinking about how this is gonna work two days from now and I think we need to do this instead and that kind of thing, right?

The

coordination of that isn't fantastic right now. I think that will get a lot better. and it will get more it will get closer to what an experienced human can do in terms of thinking about what you're working on right now while you're also thinking about other things and you've got, you know, that kind of thing. so I think that's where the next productivity bump is gonna be basically.

It's freaking complicated and one of the biggest problems with it, in my opinion at least, is that these things are basically unmeasurable in a token sense, right? You've got no no concept of like how much money they're spending or how many tokens they're spending. And the only thing you can really do is put a hard limit on it. So it's like, okay, we got, you know, we got to a certain point and, you know, you ran out of your budget, we're done.

it doesn't have the ability. So like with when we're working on something, right? We can think, okay, I've got a week to get this done. And then you get to a certain point as you're working on something and you're like, okay, there's no way this is going to get done in a week. How do I get it to a point where it's like, you know, where where I I can I it it it's functional now. I can build off of it later, right? But let's I I need I need to stop this part. I need to get this other stuff done because it's more important.

And then you when you get to the end of the week, it's like, cool, I got this much of you know, the 80% of what I was trying to get done done, but what I got done is functional, right? And they don't have this concept of like, okay, so here's my budget, right? If I estimate how long it's gonna take to get this thing done, right, there's no way that's gonna fit in the budget. So I'm not gonna do that part, right? I'm just gonna get to a like a stub.

Stub that out, go to the next thing, get it done, you know, that kind of stuff, because that's the more important thing to work on right now. They've got no concept of any of that kind of stuff. I don't I haven't seen any evidence or any even attempt to have any of them estimate how many tokens a thing is gonna take before they start. th

Nathan Toups (11:34)

Yeah, that's interesting.

Carl Brown (11:35)

th there might be a thing that does that. I don't know. I've never seen that. so but anyway, so that that that

awareness of kind of where you are in your timeline or where you are on your budget or that kind of stuff and being able to compensate for that is a thing that doesn't exist yet. So the problem with firing up the big harness and having them all spawn each other as they need to and blah, blah, blah, is it's great if you're, you know, open AI and you're gonna run the thing for two weeks solid and it's gonna spend a million dollars on it because you're trying to, you know, solve a math prize or, you know, hack hugging face or something. Or I shouldn't say that, or, you know, solve exploit gym or something like that, right? But for those of us that

have

a budget, it's like, you know, okay, I'm kinda hesitant to go full bore and see what this thing is gonna do because God only knows, you know, how long it's gonna take before it runs out of, you know, all of my budget. And there's no telling what percentage of what I told it to do is actually gonna be functional at all when that happens.

Nathan Toups (12:32)

That's a really interesting idea in that, you know, we have a lot of things like dorometrics and other things for humans. you know, this is I guess what one of the purposes of Agile was originally. But well, again,

Carl Brown (12:41)

Mm-hmm.

Nathan Toups (12:44)

I said I said original purposes. I didn't say successful, but the idea of no, but I meant I Yes. Right.

Carl Brown (12:47)

I that's a whole other conversation. We can go there if you want, but a a Agile was not for what we're using it for. Yeah Actually, no, no, it has a lot. Agile Agile is for software consulting. If you're

Carter (12:51)

Carl True Agile has never been tried.

Ha ha

Yeah.

Carl Brown (13:01)

if you're parachuting

in and you're doing consulting, it's perfect. It's wonderful. If you've done that kind of thing, it's great.

Nathan Toups (13:06)

Right.

Carl Brown (13:07)

Using it in an enterprise anyway.

Nathan Toups (13:10)

No, but I I guess my point was it'd be interesting to have a a feedback loop that was an estimate and then a reconciliation with how accurate that estimate was in some automated way. That I I I haven't seen any efforts on that either. I again if you if you have, if anyone's watching this video and you've found a technology that does that, I think that would be really interesting. So please share. Yeah, exactly. Yeah.

Carl Brown (13:33)

And if not, there's your next startup idea. You could probably get a billion dollars worth of funding for that tomorrow.

Carter (13:36)

Ha ha ha.

Nathan Toups (13:39)

That's that's absolutely true.

Carl Brown (13:41)

Call SoftBank.

Carter (13:45)

Okay, well th it's funny yeah, yeah. Yeah, yeah,

Nathan Toups (13:47)

Well, I I w so I'm gonna stop here for a second.

Carter (13:50)

yeah.

Nathan Toups (13:50)

Carl, how do

Carl Brown (13:51)

Yes.

Nathan Toups (13:51)

you use these tools? Like which what are th your favorites? What do you avoid like the plague? What's a hard line of like I will never do this?

Carl Brown (13:58)

Okay, so so

lately lately I the last, I don't know, three or four or five days, I don't remember how long I got access to it. I've been playing with Astra and I hate it. It is it is

Carter (14:10)

Okay, what do you not like about Astra?

Carl Brown (14:11)

it is it is so the this is the new ChatGPT six, I guess. it is so yeah,

Nathan Toups (14:18)

Y you hate AGI, is that what you're saying?

Carl Brown (14:21)

yeah, right. it is so over engineering stuff.

It drives me crazy. so I was working on I've been working on this this project for a potential upcoming video type thing, right? And what I'm doing is I'm basically going out and I'm using the C V A C V E database to get statistics on what's the the the the way that bugs have been added and changed over time, right? And trying to do some analysis on

kind of, you know, a as you know, a as this particular agent has happened, or you know, when this time frame it happened, here's what was going on in the AI at the moment. What did that do apparently to the, you know, w w was there a trend that changed in the number of bugs that were reported and how does that, you know, does that correlate and that kind of thing, right? and the the frickin' Astra keeps doing stuff like I'm trying to do like a search of because if you go to the C V E site

And

you say, show me claw show me bugs in Claude, right? It will return every bug whose whose credits say I was assisted by Claude when I found this bug or whatever, right? Which is not what I wanted, right? I want

Nathan Toups (15:44)

Yeah.

Carl Brown (15:45)

bugs in, cause like when when when when Avina says, you know, we're shipping the what what was it? We're shipping eight times as much code as we were

the same quarter last year or something like that, right? It's like, okay, that's not an indication of productivity. Right. I I have I have been through many, many cycles in the last thirty-five years of people measuring productivity by the line l number of lines of code that were generated and it's never ended well. So I'm trying to figure out, okay, so, you know, what does that mean in terms of the bugs that were generated in the in the in the the the thing we're

Eight times more code was generated, right? Does that mean eight times more bugs? Does that mean six times more, 16 times more bugs? Is it actually making less bugs? That would be fantastic. I I haven't seen any evidence of that, but you know. and what it does is I say, okay, so I'm looking for this kind of thing. And it basically is like, okay, cool. I am going to unprompted or unre unrequested, go take the string that you got, I'm gonna apply that to the database, the the ontology of all of the different products that exist.

I'm gonna find the product that matches the strings that you gave me and I'm gonna search just for things that have that product in them. Right? Which means unless I have the product name exactly right, it won't work. And if I want to search for more than one product at once, because I'm trying to search for, you know, all the things from a company or something like that, you know, it won't work either.

And I'm like, okay, no, that's not what I want. I want a full-text search, but I only want a full-tech search of these fields. And I don't want a full-text search of like the credits field, because that's useful. That's useless. That doesn't help me, right? And it's like, okay, I understand, blah, blah, blah, blah. And then it's back to frickin' with the ontology again. It's like, why are you doing that? I didn't ask for that. And I've explicitly told you twice now, don't do that. but it's like, you know, I know better than you do. It's like, no, you don't.

Nathan Toups (17:39)

Well

Carl Brown (17:40)

You really don't. and it's it's the the one I have had I so

In the past, I have had more problems with ones that just like didn't get what I was asking for. Right? It's like, you know, okay, I have no idea. And I still still have this problem where the, at least in my opinion, the code that they write that's test code is like an order of magnitude worse than the code that they write that's normal code. and

Carter (18:07)

Right.

Carl Brown (18:08)

and I th and I think that's to some extent because if you look at like

you

know, GitHub or whatever, the amount of test, you know, unit test code or or or integration test code that exists on GitHub is an order of magnitude or two or three or four orders of magnitude less than the normal code. Right? So it doesn't surprise

Carter (18:26)

Right, right.

Nathan Toups (18:27)

Mm. Yeah.

Carl Brown (18:28)

me. It's got a lot less training data on how to write tests. And so the way that it write tests drive me crazy and I always have to go in and go, no, no, no, that's not what I want.

Nathan Toups (18:35)

Yeah.

Carl Brown (18:36)

I want you to do this and I want you to do this and I want you to do this. but I've never really had one, at least at this level, I've had them

over engineer things a little bit before, but I've never had one at this level of like, like, no, no, I I I know what you're trying to do. You've already tried to do this before. You know, stop trying to overengineer this. I want what I asked for, not what you think I'm you know. So it's like it's like they trained it on like trying to to be a coder for a product manager who doesn't actually understand what to ask for or something. I don't know, but

Carter (19:06)

Yes. my gosh, I feel that so strongly.

Carl Brown (19:10)

it's driving me crazy.

Carter (19:12)

Yeah, the the the amount of times I I I feel exactly what you're saying, where it's like the the the coding agent harnesses can't quite decide if this is a professional using them like as an implementer with a very strong kind of human in a loop presence, or if they are some random person who has never written a line of code in their life

Carl Brown (19:34)

Right.

Carter (19:34)

kicking off a a a task that's supposed to run for 19 hours. I think

I think what the labs want is much more the second. Of course they'd love you to run their agent in a loop for 19 hours, right?

Carl Brown (19:46)

Yeah, well, I mean, the

other thing is that when when you have the multi-stack agent thing, right? And you have

Carter (19:51)

Right, right.

Carl Brown (19:52)

the agent on top saying, okay, here, here's the thing I want you to build, it doesn't have any it doesn't have the same kind of context that we do when we say, This is what I want you to build, right? And

Carter (20:00)

Right, right.

Carl Brown (20:01)

so, you know, the the things that are doing more of the implementation, I think were written with the expectation that the thing that was going to be telling them what to do was another agent that didn't have that level of abstraction.

Or th that didn't operate at the lower level of abstraction. And so it was supposed to basically take the thing and run with it, you know, with few with full autonomy to to make whatever decisions that it wants. which, you know, might be great if you've got a huge budget and you're willing to go through a lot more iterations than I am, but it drives me insane.

Nathan Toups (20:35)

I I do wonder, and again, I don't I don't mean to make this in a conspiratorial sense. I mean I literally mean this in a the sense of you know so much of Silicon Valley's economy is based off of your attention, right? And I would imagine that there would be aspects of the training of the of these data sets on retention of long-lived chat sessions, meaning that the types of tokens that are returned make me want to continue to be engaged.

I don't know how that would be built or designed, but if you're like again, let's use the product manager persona and some engineer chimes in, hey, you know, but does it have test coverage or does it have these other things? And you feel smart because you're interacting with it and you're like, yeah, what I am I do have test coverage and look like I'm able to prototype this so quickly and like I'm getting positive feedback from my peers or even management,

Carl Brown (21:28)

Right.

Nathan Toups (21:29)

those feel like a feedback loop that could be exploited, right?

Carl Brown (21:33)

So I I I have really, really felt that in the not coding sense. Right. So when you're asking it, so like I use various ones. I use AI pretty much as my web search now. because although I have problems with that, but because it does a much better job of understanding what I was looking for. It knows the difference as I say a lot, it knows the difference between not left and not wrong.

And so when I I don't have to be as as clear about, you know, trying to think about what synonyms I want to look for and all that kind of stuff. I'm like, you know, find me a thing that, you know, a lot of times when I'm working on a video, it's like, you know, find me an article where there was an interview where one of the people that worked for an AI company, you know, said something to the effect of this sentence. Right. And then I'll go gather a bunch of those and I'll look and go, okay, so is this a sentiment that's actually, you know, in the industry, or is that just my impression or or that kind of thing, right? and

th I know about lately, but they've just gotten to the point where they've started like you you ask them for the question, they give you the answer. they're really, really bad about actually giving me the sources, which drives

Nathan Toups (22:45)

Mm.

Carl Brown (22:45)

me crazy. I always have to like, no, no, give me the links. I asked you for the links. Don't just give me the answer. but but the other thing they do is they always are like, you know, by the way, I can here are the eight things I can do next. Would you like me to do one of these?

Right. So it's like they they've started having this thing where they're like every time I ask them a question, they're suggesting here are the next four or five or eight or whatever questions

Carter (23:07)

Yes.

Carl Brown (23:08)

I can ask, which I didn't ask you for that. I don't want that. I know what I'm gonna want to ask next. If you give me the stupid answer, would you just shut up and give me the stupid answer? I don't want to spend tokens on that. Right. I mean, not that they're that many tokens. I don't, I don't do the like let the thing run for two or three hours or two or three days

Carter (23:24)

Right, right.

Carl Brown (23:25)

without me looking at it. So I don't really have a budget problem. but I don't.

I I don't wanna put myself in a position where I have a budget problem. Not really because it it's not about the money, it's just like I don't wanna waste all the the time and effort and, you know, I I don't wanna walk away and come back in twelve hours and be like, Okay, what the hell is this mess?

Nathan Toups (23:43)

Right.

W what's your daily driver for the searching? Like are you using Claude K Claud or what would which

Carl Brown (23:49)

I so I used to use Perplexity. I thought it worked really well. I just had to get rid of it like what, a week ago? because they switched to a token model. and so basically what would happen is it works kind of okay in a browser, but on the app on my phone, I would search for a thing and it said, sorry, you don't have enough credits. It's like, okay, my subscription renewed forty eight hours ago.

Right. I've done four searches since then, and you're telling me I don't have enough credits. Okay, I g I guess I'm done with perplexity. So now I'm kind of in this, you know, I use it it kind of depends on what I'm doing. So like anything that's related to YouTube or analytics or any of that kind of stuff, I end up using Gemini for. I I use coggy a lot.

Nathan Toups (24:46)

Mm, yeah, it's

like the private search.

Carl Brown (24:48)

Yeah, it's it's it's much, much better at returning sources. It's not as good at it's not as good at like answering questions. It's not as good as like doing the synonym stuff as perplexity was. and I use not as much for web searching, but for like give me like summarize things for me and that kind of stuff. I'll use

Luma, Lumo, whatever, the one from Proton.

Nathan Toups (25:21)

Hmm.

Carl Brown (25:21)

that's supposedly end-to-end encrypted, so that way my data isn't being used to train other things. And I spent a lot of time using Notebook LM, which is a Google thing, where it what it will do is it will say, here are like you give it, okay, here's a bunch of documents. I'm gonna ask you a question. I only want things that are in those documents.

Right. And so like right now I've got a giant collection of articles that I've saved off about people reporting on the hugging face hack. Right. And I've got all

Nathan Toups (25:54)

Yeah.

Carl Brown (25:54)

of those in a big notebook LM and I'm like, okay, so you know, how many of these were news reports that said, you know, how many use the the, you know, AI went rogue phrasing, right? And how many use, you know, whatever and what's the most, you know.

What's the most extreme clickbaity kind of thing do you think? Give me like 10 examples of that. And okay, that one's pretty clickbaity. Let me go look at what they were saying and that kind of thing. when I'm trying to kind of work on understanding not so much what the what the hack was, I get that, but like what was the media response to that? What were the headlines? What what is it that people what is it that the people that were talking about this that don't that aren't technical, that don't understand it,

What were they taking away from that? What's wrong with what they were taking away from that? And how can I make content that will let people understand what really happened and counter that narrative of like, you know, my you know, God, God, we're all gonna die, kind of thing, right?

Nathan Toups (26:54)

It it's been interesting to see which which tools folks graduate gra gravitate towards and I think a lot of it has to do with like the the ones that you end up having some success with and then you stick with it. I think or I guess I that goes for any product, I guess really. mine was with Claude. I I I use Claude for literally almost everything. I don't really

Carter (27:13)

Yeah, me too.

Nathan Toups (27:14)

spend a lot of time looking at the other the other items, but

Carl Brown (27:20)

Yeah, and I I for the most part I intentionally don't. I

Nathan Toups (27:24)

Mm.

Carl Brown (27:24)

feel like part of my I don't responsibility is way too strong a word, but like what people come to my channel for is Yeah, right. Is is an understanding

Carter (27:30)

Sacred duty.

Carl Brown (27:34)

of like, you know, what how how has the state of the art moved, right? and I've been doing videos for two and a half years

Carter (27:38)

Right, right.

Carl Brown (27:39)

on I don't do as many of them anymore just 'cause they become so much more time consuming and and they don't perform as well. and and I feel like

I I feel like my time is better served by trying to counteract some of the narratives that are causing people to make stupid decisions like, you know, let's

Nathan Toups (27:58)

Right.

Carl Brown (27:58)

lay off this entire department because AI can do it. And then six months later it's like,

Carter (28:01)

Right, right.

Carl Brown (28:02)

okay, well let's have to hire everybody back because that didn't work. and of course we're gonna hire everybody back at less salary than or not everybody, we're gonna hire the same positions back with less salary because these people have less experience of that kind of stuff. And it's that's that's all a mess. And if I can help people avoid that, that's

Nathan Toups (28:15)

Mm.

Carl Brown (28:16)

great. but

Part of what I think I I I try to do and that people want to know about is basically, you know, I don't want to get stuck in a rut of this is the way that I do, you know, AI stuff, right? It's like, okay,

Nathan Toups (28:32)

Right.

Carl Brown (28:32)

you know, cool, Astra is out. I'm switching everything to Astra. and I'm gonna see not everything, everything, like the stuff that's in Notebook LM is still in Notebook LM, like let me stretch everything over there. let me figure out what's different about this. let me, you know.

see what's that kind of thing. So I ne I never get like, you know, okay, I'm doing, you know, I I I'm a I'm just a codex user or I'm whatever. I have given

Nathan Toups (28:54)

Mm.

Carl Brown (28:55)

up completely on cursor. when

Carter (28:59)

Right.

Carl Brown (29:00)

when when it went to, you know, the only AI that I can use is Grok. I'm like, okay, well, so much

Carter (29:04)

Yeah. yeah.

Carl Brown (29:05)

for that theory. but for the most part I just kind of bounce around between things as new models come out just to kind of see, you know, what's changed since last time.

Nathan Toups (29:17)

What about sandboxing and isolation? Like when you're running code and stuff, how how do you how do you do that?

Carl Brown (29:22)

I have so I have a dedicated Linux machine that's an old Intel Mac Mini. And on it, I have a number of I have a four four terabyte sand disk SSD. None of this is sponsored, by the way. and I have multi I I run multiple virtual machines on sitting on the SSD on a dedicated Linux machine.

That's not my primary machine. So it's like in a VM on a machine that's separate over in a part of my home lab that's got a bri a a firewall bridge between it and the rest of my network. which is probably overkill, honestly, but

Nathan Toups (30:02)

Okay. No, I so I

I don't think so. So I I'll tell you that I run, I do I do have cloud desktop that I will use for like queries and searches and stuff. That's mostly research, right? But everything else I use cloud code, mostly cloud code inside of this thing called Herder, which is just kind of like a Tmux replacement that's kind of got some agentic things that are nice. But but I I'm in one of two places. I'm either using exec.dev, which is a

A new startup from the co-founder of Tailscale that makes like VMs that you can spin up real quick. I like I've been using that a lot. And then I also have Proxmox running on a Linux box in my house that has virtual machines to do isolated workloads. And because I do consulting and contract work, those are

Carl Brown (30:50)

yeah.

Nathan Toups (30:51)

by definition isolated by by client, right?

Carter (30:53)

Right, right.

Nathan Toups (30:55)

but I also have it isolated by certain by projects that I'm doing and

and actually use tail scale with it. So I can, you know, I'm I'm traveling right now. I have access to those machines back in Costa Rica. I have, you know, pretty strict ACL rule like network access rules related to those. It doesn't and so but I will tell you, like, I feel like an oddball sometimes because there's other folks who are like I will they like run stuff in YOLO mode, right? Like they just

Carl Brown (31:20)

yeah. I I

I went to a a a conference, I guess what, a couple of months ago here in Austin for YouTubers. And it was about YouTubers and brands and getting sponsorships and all that kind of stuff, right? And the whole the whole conference, the like week long like four day, five day thing, it was all about how to use AI. Not necessarily like how to stuff, but like, you know, the lots and lots of sessions on what do they call it? Either GEO or

AEO or something like that. It's basically the the new version of search engine optimization.

Carter (31:52)

SEO. Right, right.

Carl Brown (31:54)

It's like, how do you, how do you get when Chad GPT says, or when somebody goes to Chad GPT and says, you know, what's the channel I should go to for X, or you know, what kind of

Carter (32:01)

Right, right.

Carl Brown (32:02)

office chair should I buy or whatever, how do you get your product to be the one that Chad GPT recommends or that kind of thing, right? and the number of people that I saw on the on the in on the influencer track, God help me, were

Like there were guys there was a guy that was like, okay, you know, I've got two thousand videos. I've been doing this for five years. I have like 1.1 million subscribers. He was in the like like the carpentry space, I think, some kind of like woodworking kind of how-to video kind of things. I don't remember exactly. and he had an agent, I don't remember if it was Claude Code or c something, but or or open claw. I think it was

Nathan Toups (32:41)

Cowork or whatever. Yeah.

Carl Brown (32:43)

open claw, and it was basically going through all of his videos.

And looking through the transcripts and looking at the video and finding the products that he had used, the table saw, the, you know, the the WD forty, the gloves, the whatever, and putting links in to the to the the affiliate links to the product that he was using at the time.

Right.

Nathan Toups (33:13)

well.

Carl Brown (33:14)

And then searching through the comments and finding people that were asking about, you know, hey, what about, you know, this, you know, what what were we using here or that kind of thing? Right. And then putting in affiliate links there. Right. And I'm like, dude, all it's going to take is one of those links that the Claud Bot thinks is something that it's supposed to execute as a con instruction.

And your 1.1 million subscribers and 5,000 videos are gonna turn into a freaking crypto scam farm and they're gonna delete all of your videos. What the hell? And he was like,

Carter (33:45)

Right, right.

Carl Brown (33:46)

and I'm like, my God. And I never heard anybody but me, and certainly nobody from the stage, that the only thing that I heard that was the least bit negative or cautionary was make sure there's enough of you when the

when the AI writes your script for you so that

Carter (34:07)

Ha ha.

Carl Brown (34:08)

you can read it in your YouTube video, make sure there's enough of you there so you don't end up sounding like everybody else. That's like the closest thing to a cautionary statement that I heard from anybody at the whole conference. It was terrifying.

Nathan Toups (34:18)

Hmm. Well

Carter (34:20)

I've been concerned

lately with cloc 'cause you guys talk about like YOLO mode. I'm like, oof, guilty, which which not but but more like weren't

Nathan Toups (34:28)

Well y

you're you're probably running it on a work computer though, so at least, you know there's your there's your isolated box.

Carter (34:30)

I'm running it a work yeah, I'm running on a work computer. Right. And so yeah, right,

I guess, right. but it it started to do things that spook me, right? Like, for example, I have a question. I just asked it, I was just like, I I got some like kind of on call requests. I'm like, hey, could you just look at the data model and give me the SQL query to run? Because we have w our database is isolated. It's isolated by IP add you know, by IP address. and so even we we run a a copy of PG Web.

Which

is restricted by user account, like you have to authenticate into it. And even that is read-only. And so we just use it to like if you want to make queries against the database, without having to like spin up the bastion host or connect to the bastion host and and do all of that. And it's just nice that it's it's always read-only. So no one accidentally fat fingers a you know, something. And so I'm like, I just need you to look at our our data schema and give me the SQL query so I can go to our sanctioned read-only.

Kind of little console and and run it myself and I'll get the data and kind of understand what to do with it. And instead it it figured out it was like, it looks like the user has a command to connect to a bastion host, right? And then it was like, and then it and then it starts like reading through my like ZSHRC files and being like, good, I found production credentials to log in. And then it starts like, and like, and it looks like the tunnel. Like, yeah, and I'm just like, what? No. Stop all of that. Right. And it like,

Halfway into the database. And yeah,

Nathan Toups (35:58)

That's the stuff it was telling you, yeah.

Carter (36:00)

exactly. I know, right? And and that, and we had another engineer. And then this, you know, he's he's one of our like more diligent kind of security minded engineers. And he freaked out because at some point the agent had decided that what needed to happen was like a staging deploy to like validate something, which we don't know why it needed to do that.

Carl Brown (36:16)

Great.

Carter (36:17)

And so it found his GitHub, he was he was authenticated through the GitHub CLI. Cause like, why wouldn't you be? Because you, you know, if you want to spin up a pull request and it was like,

Carl Brown (36:26)

I know, I know.

Carter (36:27)

I yeah exact right I I'm a

Carl Brown (36:29)

Yeah.

Carter (36:31)

anyhow and so and it like spun up it like kicked off the GitHub action to like deploy to staging and like it wasn't a big deal, like nothing bad is gonna happen if you deploy to staging, but we were like, wait a minute, what why are you deploying to staging? Like how you know when are you gonna decide that actually what you really need is a production deploy and it it it's spooky and has

Nathan Toups (36:50)

What you need is some Monero,

Carl Brown (36:51)

Yeah. I

Nathan Toups (36:52)

you know, crypto miners on your staging environment. I think that would

Carter (36:54)

Yeah, I exactly.

Carl Brown (36:55)

Right.

Yeah, I I

Carter (36:56)

Right. Right.

Carl Brown (36:57)

I'm using a thing called I think it's Git it's either Git T or Git E A. It's G-I-T-E-A. It's a like a a GitHub clone written in Go.

Nathan Toups (37:08)

I always pronounce

Carter (37:08)

Okay.

Nathan Toups (37:09)

it Gitia, but maybe I don't know how actually. Yeah. Mm-hmm.

Carl Brown (37:10)

I I have I have no idea. th since T E A is T, I think it might I

Carter (37:14)

Yeah, yeah.

Carl Brown (37:14)

don't know. I don't

Carter (37:15)

Giddy.

Carl Brown (37:15)

know. But anyway, but I have I have an instance of that set up and I have my agents talk to it and then I basically have a repo on my computer that's got one remote, one Git remote that points at it and one Git remote that points at GitHub. And so I'll pull things down that the agent has done. I'll like review them manually and then I'll push them up to GitHub if I want them.

Right. And so if it goes completely nuts and kills the Git server, it's a you know, it's my own VM that's running, you know, that I can that's got hopefully I've had a recent snapshot that I can roll back to. Hopefully I've remembered just snapshot it recently.

Nathan Toups (37:50)

Well it so this will be our second

startup idea of the day. because this is an area that I think is an unsolved issue. I think we need something similar to a standard like SAML or one of these other standards in which I can I can issue lower privileged credentials based off of like I have a set of credentials and I wanna do a lower scoped credential in some automated way, right? Some way that I can say, I can review something.

spin off a sub agent with some lower prev piece that I know that what the scope is, give it a s small time window, you know, kind of constrain it that way. it I think if we could figure out a standard for that and then you know a lot of tools that you want to hook up to agentic things, I would feel much better about it, right?

Carl Brown (38:33)

Yeah.

Nathan Toups (38:34)

but I'm again you have to be really careful with it. But I I yeah.

Carl Brown (38:37)

Yeah, I so

Sorry, one you just reminded me of something. One thing that I don't do that drives people crazy when I talk about it, I don't like MCPs.

Nathan Toups (38:50)

Okay.

Carl Brown (38:52)

I have and I I haven't when I say, you know, I change models all the time when things comes out, I I tend not to use MDs MCPs and skill files and that kind of stuff. Because what happens is, at least in my experience, when I do an MCP and something goes wrong, I have no idea what the hell went wrong. Right.

Carter (39:10)

Right, right.

Carl Brown (39:12)

And if I'm gonna have the thing do what an MCP would do, what I'm gonna say is write the Python code to do that, execute the Python code.

Carter (39:20)

Right, right.

Carl Brown (39:20)

And

then when it doesn't work, I can open up the Python code and go, okay, what the hell did you okay, that's what you did. Okay, here. This

Carter (39:26)

Right, right.

Carl Brown (39:26)

is how you fix it. Go try again. Right. And

Nathan Toups (39:29)

Right.

Carl Brown (39:29)

if it's an MCP, it's like when when it when it when something doesn't work, it doesn't have enough information to give back to me to say, you know, this is the error that happened, this is what went wrong. Cause it's like I I have no idea what you were trying to do. so you know, I I have it instead of

I I always have it generate the code that's gonna use it's gonna use to talk to whatever it's supposed to talk to and then run that.

Carter (39:51)

Right, right.

Carl Brown (39:52)

and that way I have a much better idea of what it's doing and I can audit it. It takes longer.

but you know

Nathan Toups (39:56)

Yeah. I I've been back

and forth. I personally don't use any MCPs or I don't I'd have to check to see what I have on my for client work though, if there's an MCP server for an internal tool and

Carl Brown (40:11)

Sure.

Nathan Toups (40:12)

it's there, that's mostly and it and most of the time it's like, I want the Grafana read only internal MCP tool so that when I'm at troubleshooting something, it can just like you know, gobble that stuff up really quick.

Carl Brown (40:23)

Yeah. Well part of client work is trying to integrate yourself into the workflow of the client that you're, you know, or or

Nathan Toups (40:27)

Right. Exactly.

Carter (40:28)

Right, right.

Carl Brown (40:29)

or saying, you know, okay, well, you know, I given my experience, right, with the workflow that you've got, I would recommend these changes, right? But but you know, but but

Nathan Toups (40:38)

Exactly, exactly.

Carter (40:37)

Right.

Carl Brown (40:40)

but coming in and saying, you know, okay, I'm gonna I'm gonna do use my complete workflow that's completely different than what you've got, so that there's nothing you can do that's useful to to replicate

the way that I'm doing

Carter (40:50)

Right.

Carl Brown (40:50)

things and you know that's that's job security, but it's also being a jerk,

Carter (40:54)

Ha ha.

Carl Brown (40:55)

right?

Nathan Toups (40:56)

But

and I will and I will say that like again, you know, because everything's a trade-off, I was internally didn't like the implementation of like the Databricks MCP server. And I wrote a short a little skill called Databricks Explorer that was really just a wrapper on a bunch of Python scripts. And again, it was for that exact reason because like I I wanted to be able to troubleshoot and I wanted to lean on the access levels that we could pass in. And I just had more control. It was like much easier, and we just

checked it into the code base for the DevOpsy type people that would be dealing with this. I see, I mean, it's funny, like I see the different values of the workflows and I kind of oscillate. Like I I'm kind of I wish I had like a more like bold answer one way or the other. when people criticize or or or complain about MCPs, I'm like, yep, I I agree. yeah.

Carl Brown (41:44)

It's I

mean, w we have to and and this is true for for you out there listening, everybody needs to understand this is all a moving target,

Nathan Toups (41:51)

Mm-hmm.

Carl Brown (41:52)

right? The state of the art a month ago is not the state of the art now. and so, you know, give yourself a little grace that

Carter (42:00)

Right.

Carl Brown (42:01)

everything is moving so fast, right? And and you know, don't I mean and if you if you want to stick your head in the sand for six months and then come up and go, Okay, so of all of the trends that happened in the last six months, you know, of the ones that died off.

I'm gonna ignore those. How much time did I save myself? Well, it's gonna be a pretty big

Carter (42:17)

Right, right.

Carl Brown (42:17)

'cause there are a lot of times you end up going down a rabbit hole 'cause it seems to be the popular thing and then suddenly it's like, no, we don't we don't do prompt engineering anymore. It's all loop or

Nathan Toups (42:25)

Okay.

Carl Brown (42:25)

loop engineering now or whatever. It's like it changes every two months or something.

Nathan Toups (42:29)

And I'll I'll bring this up too, which is I really hate this do this now or you're gonna be part of the permanent underclass. I think is the

Carl Brown (42:35)

yeah.

Carter (42:35)

Yeah, so

Nathan Toups (42:36)

worst piece of advice because exactly this. Most of what you're learning day to day right now, if you're like staying on top of the trends, is going to be completely irrelevant. you know, there's lots of examples in history where if you just wait until you wanna solve a problem and you're gonna wrap your head around it, those people always win long term. And I it it's not gonna

You know, there also, I think another counterexample of this is there are young people right now not using any of these tools. My my daughter's 11, right? She doesn't use any of these tools. One day, I know, I know, what a

Carter (43:07)

She's not using any of these tools? Wake up, Nathan.

Carl Brown (43:09)

Yeah.

Nathan Toups (43:10)

loser. she's gonna be part of the permanent underclass. So one day she

Carter (43:14)

Yeah.

Nathan Toups (43:14)

might need to use some of these tools at some variation and some point in the future. And at that point, she will learn those tools. Like it and and I'm not like, man, if I don't get her in this right now, you know, permanent underclass.

Carter (43:27)

And that's I I think all the time, I know we we mentioned this when you're on your last time, because one of my favorite videos of yours is is the biggest lie the AI companies are telling you, which is that this is the worst the models will ever be. And you say,

Carl Brown (43:38)

Right.

Carter (43:39)

Well, the models are software. They're not hardware. Hardware does kind of just improve over time, right? software does not. and I think you kind of famously said, like, was Windows Vista better than Windows XP? Right. Like, and I think about that with the model

Carl Brown (43:54)

Well, like for

Carter (43:55)

or or

Carl Brown (43:56)

for what I do, at least at the moment,

Carter (43:57)

Yeah. Right.

Carl Brown (43:59)

GPT six is way worse or GPT five, what the I don't know what that which hell one we're I don't care if you feel like but Astra

Carter (44:02)

What I know, right?

Carl Brown (44:05)

is way, way worse than the codex I was using before Astra came off for what I'm trying to do.

Carter (44:09)

And the thing with me is that even it it's not so much that I can kind of clearly point and say, like, this is worse, it's that I can't tell the difference, right? Like, and and that to me, it's like if you hand someone a palm pilot and then the first iPhone, it's just like, well, clearly this is

Carl Brown (44:25)

Right.

Carter (44:26)

better, right? And then you hand someone an iPhone versus an iPhone four. It's like, well, this is better easily. If you handed me today, like, I I don't even know what iPhone I have, I think it's an iPhone 14. And I think if you handed me an iPhone 17.

Honest to goodness, I could not tell you the difference between them. Right. And and that at least has some verifiable hardware benchmarks where you can say, well, this is better. And that's how I feel when like I go from like Opus 4.8 to 5.1, or sometimes I'll like summon Fable and be like, this is a really important task. And and then it's like, I can't even tell the difference in the output. The only one I can tell a clear difference in is Sonnet versus Opus. Sometimes I'll try to use Sonnet and like Sonnet is just kind of falling over itself and

not useful in the slightest. But to me, like that, that's a bit of a smell. And when I see kind of the advancement that's happening in the AI industry, I see a lot more interesting stuff coming out of the harnesses and the tooling around it. It's the individual models themselves where I I just can't latch on to fable and be like, here are all the reasons this is so much better than Opus. Like it it's unclear to me.

Carl Brown (45:36)

Yeah, it's I mean, there's so many different moving targets and they all interac interact with each other, it's really hard to to understand all the permutations, right? You'd have to sit down and do regressive analysis and go, Okay, well let me try this model

Carter (45:45)

Right, right.

Carl Brown (45:46)

with this harness and it's just, you know. And and and it'll all be irrelevant two months from now.

Carter (45:52)

Yeah, right.

Carl Brown (45:54)

So I

mean so so what what I s would say is if if you do things the way I do, and you know, take that with a grain of salt, right? I'm I'm all an old fuddy duddy, but people do people even say fuddy duddy anymore? They probably don't. Anyway, yeah, right. If you

Nathan Toups (46:09)

They do today.

Carter (46:09)

You just you say fada yaddy. Yeah.

Carl Brown (46:13)

if you skip some of the like big engineering kind of things and you ha you say, okay, my you know, unlike most of the people that are using these tools, right?

I actually understand code, or failing that, I I want to understand code. Right. And so I'm gonna use code as my intermediate language. I don't want MCP crap. I don't want any of that kind of stuff. I want you to write the code, I want you to run the code. If something goes wrong, I can look at the code. My belief is that learning how to debug code that was written by something that was not thinking ahead at all.

Right, and does not remember what the hell it did last time and will not remember what it's doing now the next time you asked it to do something. Understanding how to figure out what that code is actually doing and debug it and all that kind of stuff is a skill that's gonna get more important. I reali when I started doing project rescues of iPhone code, where basically people would go to the lowest bidder.

they would be like they'd come to me and be like, you know, okay, well, you know, a we're 80% of the way done and I'm having problems with my developer. And so, you know, I need you to just get across finish line and then I would go and audit stuff and explain. They have have that conversation that says 80% of your bud budget is gone is not the same thing as your 80% of the way done. Let me give you a a list of all the stuff you're gonna need to do before this thing can ship. but I realized that like having an understanding of figuring out what this code is doing.

Given that it was written by someone who doesn't speak English or, you know, at least doesn't didn't speak English in their code, right? None of the variable names mean anything to me. You know, the the the the patterns make no sense. I don't know, you know, it's like okay, tur, you know, doing some research. Okay, so that would that was an Android pattern that got pulled over to iOS, right? Or that was a React pattern that got pulled over, or whatever. It's like, okay, you know, because people were coming from all different different disciplines. I don't know if that's the right word, but you know what I mean.

That going into a code base and being like, okay, none of this makes any sense to me. Let's dig, you know, roll my sleeves up and let me figure it out. I think that's gonna be an incredibly valuable skill. Problem is it doesn't show up on job descriptions and nobody knows how to interview for it. So

Nathan Toups (48:26)

Mm-hmm.

Carl Brown (48:27)

it's really hard in the like and and we've always had a problem in this industry, always, always, always for 30 years with the things that they use to try to interview people have very little to do with the things that you actually need to do on the job.

Nathan Toups (48:41)

Right.

Carl Brown (48:42)

Right. And and I anyway, the the only way to fix that is to give the actual interviewers agency to make actual decisions, which the companies don't really want to do, and experience interviewing. Right. You don't really get good at being an interviewer until you've done it for a while.

Carter (49:02)

Right, right.

Carl Brown (49:03)

Right. but that the problem with that particular

Letting giving the interviewers agency and giving them experience, that's a thing that the corporations can't make cookie cutter

Nathan Toups (49:17)

Right.

Carl Brown (49:18)

and they can't control and they can't really steer. and they think that they have the they they think that they have the purview over how hiring ought to work. And so they don't let us hire people for our specialties the way that would make the most sense for us. So I don't know how much that's ever gonna work, at least for

companies that are I mean, for like startups and little stuff, you know, l little stuff, smaller smaller medium businesses, it's a different kind of thing. But they're so it those are so much harder to find spots to even interview for.

Nathan Toups (49:50)

Right.

Carter (49:52)

Well, so I want to talk a bit about interviewing and really just people starting out in the industry because I think that's what I hear so much from younger engineers, from people still in college. They're just like, what should I even be working on? What should I how do you even prepare yourself to enter the workforce now when and and this is something we talked about in a previous interview with us where I said, like, you I get worried because the kind of tasks I was assigned as a junior engineer.

AI

can one shot now and you Hey, well, the kind of stuff I was assigned as a junior engineer, you you Stack Overflow and Google probably could have one shot it. So it's it's just different. We're moving up the abstraction stack. But that kind of begs the question of like, okay, well, if we were to start our careers all over again, what would we focus on? Well, it it the the year is twenty twenty six, entering twenty twenty seven, right? What would we focus on? What what what would we do to kind of try to become the best software engineers possible?

Carl Brown (50:50)

So what I've always said to people for 40 years, when you're trying to learn a new programming language, which we kind of don't need to do now, but new tool, whatever, what you want to do is you wanna start with a thing you want to build. Right. And you wanna figure out how do I, cause because what I have found is that when people sit down with the reference manual, or they sit down with the, you know, the

how to program in X kind of thing, right? And they go through all the sample code and all that kind of stuff. it's like they get some kind of understanding, but they don't really get it and they don't really stick with it, at least

Carter (51:29)

Right.

Carl Brown (51:30)

in my experience, right? But if you're building a thing that you actually want to build, right, then you get to the point where, you know, okay, I thought this was gonna work this way. It doesn't work that way. Why doesn't it work that way? Let me go figure out why, how how to make it do what I wanted it to do. I don't understand what's going on. okay, now I get it, right?

Those kinds of epiphanies don't happen when you're just going from tutorial to tutorial, right?

Carter (51:52)

Right, right.

Carl Brown (51:54)

I don't think in the large that advice has changed. If what you're trying to do is learn how to use harnesses or learn how to use MCPs, God help you, or learn how to, you know, use, you know, codecs or, you know, Claude Code or whatever, saying, okay, this is the thing I want to build, right? Let me figure out how to get to the point, you know, the point B.

that I want to get to from the point A that I started with and let me figure out kind of what are the steps along the way. and maybe let me, you know, let me try that with Codecs. Let me try that with you know with you know with Claude Code. Let me try you know, that kind of thing. I think that's a really valuable understanding of kind of understanding what the limits of the tools are, understanding how to use the tools, understanding, you know, what you're trying to accomplish, and again, understanding kind of how to debug the thing when it goes wrong.

Right. the thing that we do different, we programmers do different is we can you we can dig into the code and go, okay, what the hell? The the like the product manager that's trying to code, the best they can do is go, okay, well, it didn't do what I thought it was gonna do. Let me ask the AI to try to fix it. Right. and like Uncle Bob was saying, you know, for me yesterday when the the episode came out, I don't know when this is gonna ship, they go and they they they fix a thing and they break something else.

Right. And the bigger your code base is, the more likely they are to break something else. And and this is a thing called whack-a-mole. It happens in human teams too if you don't have good discipline and you don't have good integration tests and all that kind of stuff. But the the AIs are really, really bad at it. and so figuring out, you know, how do I control that, all that kind of stuff, right? That's how you learn to use the tools and how you learn to be productive. notice it it doesn't really, it's like

it almost doesn't matter what language you're using in working in now, right? Which is kind of cool. the used to so I've programmed in so many languages over the over the years. I like I'll blank on syntaxes a lot of times. I'm like, okay, okay, which which okay, this is go, which the hell for loop do I want in this one and that kind of thing? I don't have to think about

Nathan Toups (53:56)

Mm-hmm.

Carter (53:58)

Right, right.

Carl Brown (54:00)

that anymore, right? I used to have to program with when I was you know, unless I had been on a project for a couple of months and been programming in one language exclusively.

Right. I was always programming with a reference manual or something or, you know, an editor where I could go, okay, you know, four and then okay, who would fill in the for loop for and okay da da. and now I don't have to do any of that anymore. I don't have to think about it. It's like, okay, I want a for loop that that goes does this. It's like, okay, cool. There's your for loop, right? Or or usually I mean at a higher level, it's like I want a method that does this. Right. And

Nathan Toups (54:30)

Right.

Carl Brown (54:31)

then, you know, it'll the loop that's in the middle of the method, it'll do for me. And then I'll look at it and go, Yeah, okay, not really, but okay. and

Yeah, although it's getting they're getting a lot better at that. but that understanding of, you know, i exactly what syntax I need to use for this particular language and those little the the like the riddle crap they used to have for JavaScript about, you know, when you add, you know double quote one double quote to bare one to one point zero, like what do you get kind of thing. It's like, I don't know, it's a stupid language. I don't wanna I don't it's like that that's not

Carter (55:02)

Right, right. Yeah.

Carl Brown (55:05)

that that's that's not useful. I don't care, right?

A lot of that kind of stuff that we used to have to spend a lot of time memorizing doesn't matter anymore, which is, as far as I'm concerned, fantastic.

Carter (55:17)

Right, right.

Carl Brown (55:18)

but it's not, I mean, there there are people, okay. This is kind of a tangent, but there are people out there that really fell in love with the, you know, I I I I really liked figuring out the syntax, right? It was really, really

Nathan Toups (55:33)

Right.

Carl Brown (55:33)

cool, right?

Carter (55:33)

Right, right.

Carl Brown (55:35)

and and I feel kind of bad for those people. But honestly, if you're gonna be in this career for 20 years or 30 years or 40 years or whatever, right, there's nothing really that I did in 1990 on a regular basis that is still relevant now, right? And and that has changed several times.

Nathan Toups (56:00)

Wait.

Carl Brown (56:01)

So, you know, don't don't get precious about.

you know, I really, really like this part. It's like if that's all you like about programming, it's gonna go away,

Carter (56:08)

Right.

Carl Brown (56:10)

which is which is fine, right? Just have an exit strategy because from a career standpoint, that's not gonna stick around. See if you can find the next thing that you think is fun.

Carter (56:20)

I've I've noticed that with AI, where any programmer who primarily defined themselves and can I think derived a lot of their career worth out of I am someone who solves problems has adapted pretty well. It's the kind of people who thought, I am someone who understands all the nitty gritty of of C sharp, right? Or I am someone who, you know, knows exactly how to debug a Docker, a local Docker setup, right? Like.

I'm not saying I mean those are useful skills, right?

Carl Brown (56:52)

No.

Carter (56:53)

But if that's kind of what you thought was this is my unique value proposition, then it it's tough. I think it's been tough for

Nathan Toups (57:00)

I

Carter (57:01)

some of those engineers.

Nathan Toups (57:01)

and maybe maybe it's because I've had to become a royal master of spin in my own career over the years. just because I d yeah, well, cause well,

Carl Brown (57:08)

Welcome to consulting.

Carter (57:07)

Yeah.

Nathan Toups (57:10)

I yeah, well, yeah, also like I had a theater degree, I didn't know I I kind of wandered for a while, right? So like when you ask a question like what would I recommend to a software engineer now who's just out of college, I have no way of comparing how to make myself appealing at 21 years old to a workforce in software engineering.

I kind of went around and sampled a lot of things to figure out what I wanted to do and just fell into or you know, it not fell. I I found this way of thinking and problem solving. But I will say I I think the willingness to go deep on a topic doesn't change, right? Like in the sense

Carter (57:49)

I agree.

Nathan Toups (57:50)

that yeah so so, you know knowing simultaneously that if you really learn how to write Python quite well.

is not necessarily relevant day to day because the expectations of the job have shifted. The ability to do that, to really get into like how does the GIL work in Python or something, you know, because you're interested in it, not because you feel like obligated to. and I'll put it in the same way of like amateur carpenters, I think this is an area I like to think about. There's a beautiful community of hand tool carpentry.

Right. You

Carl Brown (58:26)

Mm.

Nathan Toups (58:26)

can do all of this stuff with these hand tools and use these old techniques that are actually not possible to do if you go to IKEA or if you're going to just get lumber at Lowe's and use some power tools. Now, the expectations of the job are that I use power tools, right? I'm gonna use an earth mover, I'm gonna do this stuff, and that's totally fine. But you might, for your enjoyment at the house, go in and make a dovetail, you know, and use glue and don't use any fasteners.

and put this thing together that you might want to keep and cherish. And you're still in the same industry. And honestly, those things are probably going to transfer over to you being excellent at the construction site. and so I I do think that like

There has to be this curiosity. And I I think that's the part that I would I would still recommend is like if you want to get super deep into like how does a database work or how do you know, how do I solve these types of problems? It's not wasted time. Like please spend the extra time and learn those things. You're gonna be an more interesting person. You're gonna learn how to troubleshoot and problem solve. You're gonna learn how to break through those like I feel like quitting moments, which I think we

Carl Brown (59:30)

Mm.

Nathan Toups (59:31)

all experience. Completely transferable, even if the day to day job expectations have shifted.

Carl Brown (59:37)

And I I I wanna kind of reinforce that this is not new. I mean, I I don't know. Uncle Bob yesterday

Carter (59:45)

Right.

Carl Brown (59:45)

said something to the effect of like this is ten times the change that he's ever seen before. I don't know where the hell he's getting that because it doesn't feel that, you know. but like one of the things that that I get pushback from when I talk to folks is I compare I compare the AI to

and even igentic AI to being able to search Stack Overflow or being able to search Google or you know being able to look up reference manuals online and you know that kind of being able to to ask questions on you know forums and that kind of thing, right? Because we didn't have any of that when I started. And every time something like that has happened, it's been, you know, okay, wow, okay, I've I've got this new power that I didn't have before. And what happens is the folks now say you can't compare AI

To

package managers, because package managers can do XYZ and AI can do all of this stuff and it's non-deterministic and it's blah, blah, blah, blah, blah, blah. And it's like, okay, hang on, hang on, hang on. You're comparing package managers 20 years after they started,

Carter (1:00:52)

Right, right.

Carl Brown (1:00:53)

right? With a, with a, with a bacon ecosystem and package signatures, and you know, a mechanism for how new changes get into them and all that kind of stuff. You're comparing that.

To AI, I'm comparing AI to package managers when they showed up before there was such a thing as NPM or you know pip or any of that kind of stuff. And you were like, you know, that we didn't even have Git. Right? You would download a zip file, you would open the thing up. I mean, it's like it, it's it's a very different kind of thing. and back then people were freaked out, right? People were like, you know, you know, my God, you know, i i if

If all I if all people are gonna be doing is using these package managers and going and grabbing things, and writing a few lines of code to hook those things together, it's like, what the hell am I gonna do for a job? Right? Well, it turned out it wasn't you know, there were lots

Carter (1:01:42)

Right, right. Yeah, right.

Carl Brown (1:01:45)

of things for us to do as a job, right? And and I think that's gonna be the the same here. but don't I mean understand that yes, it feels crazy now. and yes, everything is changing. and yes, everything is is is, you know, uncertain.

But it's always been that way. Right? Every time a new disruptive technology shows up, the thing that's worse about this one is not, in my opinion, is not the agent. And it's not the AI. It's the hype. It's the marketing. Right? And and kind of like I think

Carter (1:02:17)

Right, right.

Carl Brown (1:02:20)

I said earlier, when I I think it was after we started recording, you know, to me, this is like, you know, it's it's not, you know, a replacement for half of the workforce. This is the next

You know, Google search plus Stack Overflow plus autocomplete. when IntelliSense happened, that was the first thing I ever used where you would like type in a a a variable name and you'd hit a period and it would pop up a here's a list of all the methods you could call on that variable.

Carter (1:02:48)

Right, right.

Carl Brown (1:02:49)

It's like, my God, right? I don't have to memorize all this stuff anymore.

Carter (1:02:51)

Yeah.

Nathan Toups (1:02:52)

Mm-hmm.

Carl Brown (1:02:53)

it was a big deal, right? And that was a lot of the stuff that we spent our time on back then, right? and so

The the fact that all these things are changing is not it's not a thing that's unprecedented. we've you know, those of us that have been doing this a while have lived through a bunch of these. you'll live through this one, you'll live through the next one, right? The thing that's scary about this one is the what you're getting from the news and what you're getting from the press releases and what the, you know, the the essays that the CEOs are writing about how, you know, you know

We've re or or you know, or the interviews that you know, Judson Wong is giving that, you know, we've reached A AGI for I think at least the fourth time in the

Carter (1:03:39)

Yeah.

Carl Brown (1:03:39)

last three years that he specifically has said, you know, AGI has, you know, is here or is arrived or, you know, whatever. and it's like th that there's a lot of marketing hype because of the amount of money involved.

and because they're trying to get to the point where they've figured out how to make this thing profitable, which I don't know that that's gonna happen. I don't it doesn't seem possible to me, but who knows.

Carter (1:04:04)

Anthropic

says they're profitable if you don't count any of their expenses. Yeah.

Carl Brown (1:04:08)

Right, right, right. Yeah.

So if you if you if you if you ignore all the salaries they're paying to all of their people and you anyway. but but yeah, if if you can if you can step away from the hype and not panic about what everybody's telling you is gonna change, the actual fundamentals are still not converging, but that chaos has always been there at these at these times of transition.

Carter (1:04:37)

Well that that panic

I I think is so interesting because I

Carl Brown (1:04:40)

Mm-hmm.

Carter (1:04:41)

I I've said this before on the podcast. I've said, look, I'm sure throughout human history there have been people who just didn't couldn't see the writing on the wall and stuck it out too long and then were burned because of that. I'm like, but I bet there's way more people who instead got jumpy and who instead s saw some sort of hype and panicked and said, Well, I I can't do this. You know, I I'm not gonna stick with this career because this this is on its way out.

I

you I think about you you were employed during the dot com boom, right? And it would have been very rational or reasonable at that time to say, like, well, that was the high point, right? Like we're never going back to that again. And so I I better find a new career.

Carl Brown (1:05:19)

A there were

there were a lot of people that exited the the programming space at that point. Right?

Carter (1:05:27)

Right, right.

Carl Brown (1:05:27)

I mean like so several of the guys that I knew that were programmers became like sysadmin types.

Carter (1:05:34)

Right, right.

Carl Brown (1:05:35)

there were a couple bees one of my really good friends changed from being a programmer to a like a chief security officer.

Nathan Toups (1:05:44)

Yeah.

Carter (1:05:44)

huh.

Carl Brown (1:05:45)

Right? And so he like

became

he the he would like work on instead of like writing firewall code like he had been doing when we worked together. he had been he started doing like writing policies, security policies, that kind of stuff, right? Love you, Ben, more power to you, but man, I couldn't do that. but but like, I mean there there were a bunch of people that exited the the actual hands on code space. Some of them exited tech completely. A lot of them didn't they just found other niches that they didn't have to write code anymore or maybe have to is not the right

way of saying it, but they you know. and then some of those people have come back and most of them haven't from what I've seen. it's 'cause a lot of people are like, Okay, I've gotten out of the code thing. I'm not keeping up anymore, you know, I don't really wanna, you know, I I don't really want to try to l to to catch up. So

Carter (1:06:34)

Well, and and I I know we've talked about this before too, which is just like, yeah, the field changes. It's constantly changing. If your if your thought is, I'm gonna learn all my stuff in college and then I'm gonna enter the field and then I'll be set for the next 40 years, it's like, well, maybe there are some

Carl Brown (1:06:45)

Yeah.

Carter (1:06:45)

fields that do that. I don't think there's a ton, but it's certainly not this one. and when I think about

Carl Brown (1:06:49)

Yeah.

Carter (1:06:50)

kind of we we had this great moment where our company hosted kind of like a build with AI night, which was like it was designed for people who were like non-technical to kind of learn how to, you know, write a skill file and things like that. And there was this one.

College

student, bless his heart. Very, very cocky. He had he had set up this kind of like agentic factory or something on his machine. And he was kind of like bragging, like, I bet no one's made something as advanced as you have. And and the kind of like our interns were serving as basically the TAs. I was just there, not even involved in the build net. I was just kind of working on something on my own. And they finally just said, like, okay, let's bring him to Carter, right? Because like, you know, he's he he claims he's just so advanced. We'll bring him to the principal engineer and and and see what he has to say. And so he's explaining this factory to me, and he just says, like,

He's he's gotten to this point where no matter how much he yells at it, it won't do this thing he wants it to do. And he's like, I'm just wondering, what do I do at this point? I'm like, well, this is where we fall back on an old habit called reading the code. Right.

Carl Brown (1:07:46)

Right.

Carter (1:07:46)

Like, and I and I just told him, I said, at a certain point, there is no substitute for understanding a system deeply. And that's what I tell people entering the field. It's like, you've got to fight back against that instinct to say, I've learned enough.

Yeah, you have to kind of be perpetually willing to ask the the next level question and s and say, okay, well, I understand how it works on this level. What's the level below it? And how do I understand that? And and one of the grand ironies in all of this, AI makes that fantastically easier than it ever has been, right? The ability to kind of say, Okay, well, you know, I understand this, and but then but I don't really understand like AWS like networking policies. And so, you know, let me have AI kind of

Carl Brown (1:08:30)

Right. Right.

Carter (1:08:31)

summarize that and point me towards the right blog posts and you know and things like that. Let me feed in our Terraform and it can kind of help me understand okay, this is what we've built. and and I I just think I think a lot about Will Larson and he has some great books on kind of like what it takes to become a staff engineer to kind of rise above that next level. And he says you have to resist the impulse to snack, to kind of just do little projects.

Just

to kind of take on little projects that, you know, yes, they they, you know, maybe deliver a feature and it's it's flashy and people like that. Because like at a certain point, you have to be constantly identifying the highest leverage work you can do. And that's what you need to do and need to grind through. And I think that's good advice for anyone at any career level. But like you gotta ask yourself, what's the highest leverage thing I can be working on right now? And that's gonna look different for different people in their in their career stages. but I think if you can kind of consistently identify high leverage work.

And consistently grill yourself to where you're not just shipping a feature. You're not just as in the we just read Pragmatic Programmer and they talk about programming by coincidence, right? Where you actually understand what you're shipping, you know, that skill has proven valuable to me. And and I'd imagine, you know, in your longer career, that's probably the same pattern you've seen for success in yourself and other engineers is work on important stuff and understand what you're building.

Carl Brown (1:09:53)

Yeah.

and and don't yeah, and and try not to get freaked out by the the changes.

Carter (1:09:57)

Right.

Carl Brown (1:09:58)

let me so let me ask you a question. Have you guys ever done like user interface builders?

Carter (1:10:07)

Are you talking about like kind of drag and drop stuff?

Carl Brown (1:10:10)

No, like like thick c you don't say thick client anymore, like trying to do like a a Mac app or a Windows app or a phone app and like having pulling up a GUI and being like, Okay, I want this control here and I want this control here and I want like whatever, and then it generates all the the code underneath the the the surface that's

Carter (1:10:28)

Right, right.

Carl Brown (1:10:29)

like, you know, okay, so here's what this, you know, all that kind of stuff, right? Because like I started doing that in Windows three point zero.

Carter (1:10:35)

Right, right.

Carl Brown (1:10:36)

and

'Cause there was a there was like an interface builder thing and it would like generate all the, you know, all the menus and all the stuff and you'd be like, okay, I will

Carter (1:10:42)

Right, right.

Carl Brown (1:10:43)

add a new menu here and then you'd like cool. And then you would go into the code and it's like, okay, so here's the the empty method that you type in what you want to happen when that menu gets picked. Right. and that was a thing where it would generate a bunch of code for us. we didn't really understand how it worked until we had to go dig into it because something didn't work, you know, the the interface didn't work the way we wanted it to.

and then you would like accidentally open up the interface file, and it would change a bunch of crap. And then you'd save it again, and you'd be like, Okay, wait, why doesn't it work anymore? And this was back before Git. This is back when, you know, say source control was not what it is now, and rolling back

Nathan Toups (1:11:23)

Well,

Carl Brown (1:11:24)

was a giant nightmare, and it was like it was a whole thing. And it's like, you know, we that largely has gone away. the the

There's

still some of that in like the iPhone Android world, right? But a lot of that doesn't really happen anymore. but it was the same kind of thing. Uncle Bob was saying yesterday something about how what the what the agents can't do is sit where the human is and like look at, you know, like act like the human and like look at what the think about how the human is going to interact with the thing. And I was like, okay.

Well, you know, one I I I don't know to what extent that's I don't know how relevant that really is sometimes. If you're doing the whole product, sure. But like I've spent a lot of my career working on stuff that like nobody that was a user of the product had any idea that feature actually existed.

Right. They would have known that if it hadn't existed, it would have gotten hacked or it would have, you know, their their transaction would have failed or they wouldn't wouldn't wouldn't have known why or any of that kind of stuff. But it's like, you know, if if all the the AI is doing is being where the user is, right? And or if if that's the only thing the AI can't do and it can do everything else, it's like, okay, well, that's most of what I've done in my career, and that's not really my experience with AIs, at least not yet. but it's there's there's it

There's this constant flux of, you know, what tools are in in vogue, what tools are really useful for what we're trying to do, what kinds of programs are, you know, are expected. the idea of Google Docs when I was in college or in early career would be insane. The idea that you could make something that was like Microsoft Word that

Carter (1:13:23)

Right.

Carl Brown (1:13:23)

would run inside a browser, it's like that would not have been possible.

before, you know, all of these giant frickin' JavaScript libraries and all of the the the NPMs and all of the well not necessarily NPM but all of the the the library including and the package managers and all that kind stuff. You know, there there's a bunch of stuff that people can do now that like I I just would have been gobsmacked. I'm I I'm speaking like a like I'm living in the forties for some reason. I don't know why I'm grabbing all these old terms, but you know, it just I

Carter (1:13:53)

Yeah.

Carl Brown (1:13:55)

It just the the idea of that would have been when I first saw you know even when I first started working with browsers, the idea that I could ever be as as as feature rich as something like Microsoft Word was just I I I would have been shocked.

Nathan Toups (1:14:12)

Yeah, and the the C R D T stuff or the you know the conflict resolution of two people editing the doc at the same time too. I think that was like

Carl Brown (1:14:19)

Mm-hmm.

Nathan Toups (1:14:19)

a mind-boggling thing to me. because I was like, not only is this just word on the web, this is a collaborative word on the web, which, you know pretty interesting. To get back to the thing that you were talking about, Carter, about that person, it I've seen this time and time again.

And it brings me back to that Gall, John Gall quote, which is like complex systems that work, a complex system that works is inevitably found to have evolved from a simple system that works, and that the inverse proposition also appears to be true. A complex system designed from scratch never works and cannot be made to work. And I feel like we are a lot of people use these LLMs as like complex system creation machines. And then they can't reason about it,

Carter (1:15:04)

Okay.

Nathan Toups (1:15:06)

and then they get really frustrated. Right. And so I think that what that guy

That you were probably talking to at this the hackathon or meetup or whatever, is he probably had an idea and apparently even a complete implementation idea that wasn't well thought out. He hadn't broken it down into smaller pieces. He hadn't have like taken it down to its essence and then built it back up. And I think what happened was as soon as he had another idea, it maybe even the LLM built a decent, you know, kind of passable implementation of it. As soon as you need to make some sort of modification to it, it just falls apart, right?

I have a couple of consulting clients that I've worked with now that I have seen seen similar things where an idea gets a lot further along that's just not well thought out. And then it all falls apart, right? It gets back to this. I think the dangerous thing is, or at least the the the weird part is I it takes me a moment sometimes to realize that the person I'm talking to doesn't actually know what they're talking about because they've just kind of like

Carter (1:16:02)

Right, right.

Nathan Toups (1:16:03)

absorbed everything that the model told them.

Carter (1:16:05)

They're a meat

a meat proxy.

Nathan Toups (1:16:09)

and Carl, I was I was curious to you, like two sides to this. I mean, again, is this something so have you had similar observations? And do you have friends or former coworkers that have surprised you that have gotten like agent pilled more than you like you're I wasn't expecting that wasn't on my bingo card that so and so

Carl Brown (1:16:29)

I w I was

not expecting Uncle Bob to think these things were conscious.

Nathan Toups (1:16:33)

Okay.

Carl Brown (1:16:33)

but you know, but so the the the not well thought out thought out thing, so I don't know that I would phrase it quite like that. There have been times in my career when we had to pivot. Right? So like during the dot-com bust, the company that I was working for, we had an online product and

By

the time it was over and we had like kind of stabilized, we had switched our technology or our no, the technology the core technology was more or less the same. but the interface to it was a more of an offline thing. So we went from a a B2C thing, right, where the consumers were coming to the website and doing stuff. We started partnering with other businesses, right? And we would run what they were doing through our technology.

Right. So we b we became more of a business to business company instead of a you know B2C thing. There were a bunch of things that we needed for that that weren't a thing because we were planning on doing everything through the website. Right. And so the idea of well thought out, it's like what we had to start with was well thought out. It was just well thought out to do X. And now we need to do Y, because the market has changed.

And from the point of view of why, none of this is well thought out, because we didn't think of any of that crap, right? Because we didn't want to do that when we built the thing to start with. and so even in the sense of like not well thought out, there are a lot of times that what happens is it's like, you know, okay, even if you do have it well thought out, circumstances change, and now you're like, Okay, well, if I'd known then that I was gonna end up doing it this way, I certainly wouldn't have done that. Okay, what th what am I gonna do about that? Right. So it it's it's very common.

through the industry for you to look at a thing and go, you know, either who the heck wrote this? And you look at Git and it was me six months ago. Okay.

Nathan Toups (1:18:33)

Right.

Carl Brown (1:18:33)

I guess, you know, I I I guess I wasn't thinking about what I'm trying to accomplish now. And if I had, then I would have done it a different way. But you know, how do I get from what I've got to where I'm trying to get it to go? and that's a really useful skill. And it's I always found it to be a really interesting kind of puzzle.

the the the the difference between a greenfield project and a a transition to a new thing, right? When you've got all these, it's like you've got all these parts that are all hooked together. It's like, okay, what of this can I reuse and what can't I and what am I gonna have to throw away, but I don't want to throw anything, I don't have to. that's the same thing. That's the same core, you know, architecture understanding kind of thing.

Whether what you're doing is you're picking apart what the agent did, right? Or you're like that that that you know didn't that was built with it, not an understanding of what the heck was going on because the agent didn't have an understanding of what was going on and you didn't do a good job of of describing it. or it's just not capable of understanding your description. Or here's a thing that was written by people that works great, that I need it to do something else now, right? And I need to figure out.

where I can hook in stuff and where I can, you know, pivot and what I need to get rid of and all that kind of stuff. Right. It's the same fundamental, you know, mental model of okay, let me, or let me build a mental model, not like LLM model, but like a let me, you know, build a mental representation of what this thing actually does. And then let me figure out, you know, how do I, how do I hook that chunk up to something else so that I can make it work and what do I need to change about it so that it can understand the new input and all that kind of stuff, right?

that's a really valuable thing to know how to do. and it's a thing that you can learn by picking up code that the agents built and go, okay, so if I was gonna make this work, how would I do that?

Nathan Toups (1:20:35)

Yeah.

Carl Brown (1:20:36)

Right? so you know, yeah, I mean, I mean and going back to your question, it's like, you know, I I've run into lots of people, most of them online. I don't actually

most of the the programming I'm doing now is more for me. And most of the consulting that I'm doing now is way more like business process flow and how

Nathan Toups (1:21:03)

Mm.

Carl Brown (1:21:04)

do we integrate AI into our how do we integrate AI into our thing and how do we write like integration tests around this so when the AI goes rogue, we catch it before it gets pushed.

And, you know, I I'm working wa way more with, you know, clients for that kind of thing than like down in the guts of stuff. Cause I I guess that's I guess given my channel, that's what people, you know, come to me for, which is

Nathan Toups (1:21:29)

Right.

Carl Brown (1:21:30)

fine. It's actually really interesting work. but I don't really have people that are like that I've had that are coworkers for a long time that I'm looking at them and going, you know, are you how red pilled or you know, not red pilled, how how AI pilled are you or not?

it's mostly just

Nathan Toups (1:21:46)

Right.

Carl Brown (1:21:47)

folks I see online that's like, you know, well, you know, I've been following your channel for a while and you know, or you know I I you know, I remember reading your blog posts. why do you think you know why why are you thinking that, right? so yeah, it's been it's been interesting.

and and time will tell who's right. and I there's no I mean, I I I don't have an exclusive understanding of how this is gonna work, right? I mean, my my first reaction to to Mosaic, which was the very, very first web browser, was who the hell is gonna want gopher with pictures? and you probably have no idea what the

Carter (1:22:23)

Mm-hmm.

Carl Brown (1:22:29)

you probably have no idea what gopher is, but gopher was basically a a kind of like a web browser.

But it was all text based and it was cursor

Nathan Toups (1:22:35)

Yeah.

Carl Brown (1:22:35)

and you would like type you like tab around and you would click on a link and you would

Carter (1:22:38)

Right, right.

Carl Brown (1:22:38)

jump to that. And it's like, you know, o okay, you know, I I'm f a lot faster at that than using the mouse. Why would I wanna, you know?

Nathan Toups (1:22:44)

I I learned about the gopher protocol because I'm a Go programmer and the the the mascot for Go is the

Carl Brown (1:22:51)

Mascot.

Nathan Toups (1:22:52)

gopher and I was doing some sort of search and I was like, this is what before the browser, 'cause you know, you still had hyperlinks and stuff. I don't even know if they called them that, but okay.

Carl Brown (1:23:02)

Yeah, they did. Well, we

did. That came from HyperCard. That's a whole other that yeah, that that yeah, that's that's

Nathan Toups (1:23:06)

that's right, in the hypermedia systems, right. Yeah.

Carl Brown (1:23:08)

a whole other conversation. But I mean, the but the the thing that I didn't get, right, when I was looking at mosaic going, why the hell would anybody want this? Is that the target market for mosaic was not me. It was people that read newspapers.

Carter (1:23:22)

Right, right.

Carl Brown (1:23:24)

Right. And at that time, for me, the internet was a place where techies would go grab software stuff.

Right. It was it was not something that I thought my mother would ever use.

Nathan Toups (1:23:35)

Well

Carl Brown (1:23:36)

Right. It was not something that in the thing that I remember that blew my mind, I was I was spent some some time hooking people up to the internet, right? Where hooking people up to the internet for a living was going, getting talking to the telephone company, getting a dedicated circuit pulled into their location.

Putting a router there, putting a server there, you know, getting and when you got to the point where it said this page is under construction, I was done. It was up to the designers, and then I walked away, right? That's what building a website meant back then, right?

Nathan Toups (1:24:05)

Wow.

Carl Brown (1:24:06)

And I remember building one for a company called TriGas in Dallas, and they sold cylinders of like argon and oxygen

Carter (1:24:15)

Right.

Carl Brown (1:24:16)

and nitrogen and stuff to to industrial stuff. I'm sure they sold like big tanks too or something, but the cylinders were what they had. The the the the office

kind of out in an industrial park and I was putting the server there. and there was like this giant fenced in area that was just wall to wall or fence to fence, you know, cylinders of different kinds of gases and stuff. And I'm like, why the hell do you want to be on the internet? What what on earth? And they're like, our customers are asking for it. I'm like, what? Cause

Carter (1:24:43)

Yeah.

Nathan Toups (1:24:43)

Interesting.

Carl Brown (1:24:44)

because to me, it was like, you know, that was as far away from tech as I could really think of.

And it's like if you need to get on the internet, then this is a lot bigger than I thought it was gonna be. and

Nathan Toups (1:24:54)

Interesting.

Carl Brown (1:24:55)

I I remember being just just slack jawed, staring at like, really?

Carter (1:25:00)

We we had Lawrence Kessel on the podcast. He wrote a great little short story, coding machines. And he he said it's it's interesting to ask yourself, kind of in what ways will the world change in in that we don't expect. He says, I go back to the release of the iPhone, right? And if you told someone, hey, 10 years from now, YouTubing will be more like a real deal career for a a larger percentage of the population. It's like that tracks. It's like we just shipped this device that fits in your pocket. It ships with the YouTube app, right?

It it seems correct to me that people will consume more video and that this will be a more viable career. Like totally. Would you have guessed that a whole economy around strangers sleeping in other strangers' houses would materialize, right? Like, and that is is weirder. And and so I think when you talk about with AI and like, well, no one knows how this is gonna go. that that's really what's really interesting to me because there are so many startups really trying to chase like consumer demand for AI. And from everything I've observed.

There's

just very little consumer appetite for it, with the exception of like vanilla chat GPT, which people enjoy as like a, you know, a a chat bot and a and a search engine and

Nathan Toups (1:26:10)

A psychosis machine, yeah.

Carter (1:26:11)

and a psychosis machine. And unfortunately,

Carl Brown (1:26:12)

Yeah. Well

Carter (1:26:14)

AI generated video content, which like there is some consumer demand for that if you just look at like what YouTube shorts and TikToks are popular, right? Well, I don't know if it's demand.

Nathan Toups (1:26:22)

Is it demand? Is that demand?

Carter (1:26:24)

All I'm saying is that like you can you can create a channel that has AI generated content and

Nathan Toups (1:26:29)

Mm. And devalue your entire platform. Yeah.

Carter (1:26:30)

some percentage of the population will watch it. I agreed.

Carl Brown (1:26:32)

The I mean the

Carter (1:26:34)

Agreed. yeah.

Carl Brown (1:26:34)

the the the consumer

demand man the consumer demand is for AI companionship. There

Carter (1:26:43)

Yes, right.

Carl Brown (1:26:45)

is a there is a ton of that and it scares me more than almost anything else.

Carter (1:26:50)

I and I agree, but you keep seeing people trying to say, like, well, this like instinct is the hot new one. Like this is your personal assistant or whatever. And like I know, right?

Nathan Toups (1:26:57)

I I don't even know what instinct is, yeah.

Carl Brown (1:26:59)

yeah.

Carter (1:27:00)

Well, but it's one that like it works through iMessage and can text it and it'll like book

Carl Brown (1:27:03)

okay.

Carter (1:27:04)

reservations for you and like organize your email. And and just one of those things that's like tech people vastly overrate the extent to peop which people need to book over reservations. Yeah. I beats me, right? May maybe it is.

Nathan Toups (1:27:11)

Is is that the one that really is that the one that actually has humans that are it's like not really fully AI? I heard I heard there was one that was

literally like Wizard of Oz, like the man behind the red curtain. Okay.

Carter (1:27:22)

Yeah, yeah. But

but I I have not seen any compelling consumer product aside from what you guys have said. And I agree, and it concerns me more than anything else, which is AI companionship. and so like that that just you talking about like you you realizing like mosaic wasn't for me, right? Or kind of starting

Carl Brown (1:27:44)

Right.

Carter (1:27:44)

to understand the consumer demand. There's tons of business demand for AI. There's tons of demand for if we can.

use these tokens to replace employees, that'd be absolutely fantastic. The the extent to which that's possible is is

Carl Brown (1:27:56)

Mm-hmm.

Carter (1:27:59)

you know up in the air. It's the lack of consumer demand, which I find so interesting. because I think, you know, because you know, four years post-internet or four years post-iPhone, it's so blatantly clear what the consumer demand for it is. And and now we're kind of four years post the release of Chat GPT. And there's still, from what I

can see not really any consumer demand, which is very, very interesting.

Carl Brown (1:28:29)

Well, I mean, so it it has it seems to have replaced well, I guess it had to. It seems to have replaced like Google search for large percentage of the people,

Carter (1:28:41)

Yes, yes.

Carl Brown (1:28:42)

right? And and and to some extent that's because Google said, okay, you know, when you ask for a thing, we're gonna give you AI now, whether you want it or not.

Carter (1:28:48)

Yeah, right.

Carl Brown (1:28:49)

but but there's a lot of th the the demand for it I think a a lot

in in just non-technical consumers is more like it displaced a lot of the interactions that people had with, you know, search basically, right?

Nathan Toups (1:29:18)

Mm-hmm.

Carl Brown (1:29:18)

So people don't, you know, Google their symptoms anymore. They ask ChatGPT, you know, why is my knee hurting

Carter (1:29:23)

Right, right.

Carl Brown (1:29:23)

or whatever, right? people don't

People don't ask Google Maps, you know, what's it gonna, you know, what's the path for me to get from here to this mountain I want to hike? They say, hey, Chat GPT, how do I do this? And they go get stuck on the side of a mountain for two days. but you know, death by GPS was a thing, right? I mean, w the the Google Maps was giving people bad directions that ended up getting them driven off a cliff or, you know, lost or whatever long before there was Chat GPT. But it's just, I think a lot of the the consumer side, it just has displaced a lot of the

internet facing consumer query type stuff that used to be the non-II version of it and it's just it's it's a lot of it is just I read somewhere I'm trying to remember where but basically people saying you know AI doesn't replace jobs it replaces tasks

Carter (1:30:20)

huh.

Carl Brown (1:30:21)

right and what

What we're gonna have to do is figure out basically what does this job look like when these tasks aren't part of that job anymore because the AI is doing them.

Nathan Toups (1:30:29)

Right.

Right.

Carter (1:30:30)

Right, right.

Carl Brown (1:30:31)

but but I think that a lot of that's true for the like the consumer side too, right? Is it like it's replacing the search tasks and it's replacing the navigation tasks and the like, you know, a lot of the, you know, what used to be, you know, okay, let me go to let me go to Yelp, let me go to, you know,

Google reviews, let me go to maps, let me go, you know, let me go through, you know, a bunch of different things in order to figure out kind of where I'm gonna plan this dinner party or whatever. And now it's like, you know, hey, Chat APT, you know, you know, what what's, you know, give me some options for how I could do X, right? and that replaces a lot of the the multiple step things that we used to have to do. and I think that there are a lot of people that would be really annoyed having to go back to do all those steps manually again.

Nathan Toups (1:31:20)

Right.

Carl Brown (1:31:21)

and

the fact that a lot of the times, or maybe not a lot of times, but you know, many times, you know, what you get isn't really what you wanted. Well, that was true for, you know, especially for non-technical people when they were just, you know, searching a thing and then searching another thing and searching another thing and then trying to hook it all together in their heads. So I I I think that's that and companionship are largely where the consumer demand for it is right now. there's a lot of people that are trying to figure out basically how do I put AI behind

like I've got my interface, I've got the the my user using the interface. How

Carter (1:31:56)

Right, right.

Carl Brown (1:31:56)

do I put a hook AI behind that to take care of stuff? and that I I think that's gonna be a thing, but like, you know, a lot of that you end up with much worse outcomes, right? Like the the customer chat bot getting, you know,

Carter (1:32:14)

Right, right.

Carl Brown (1:32:15)

getting moved from, you know, India or the Philippines or whatever to an AI, you know, is even worse than

it was when it moved to, you know, somebody in a different time zone who has no idea of what city you're in and, you know, you're like if they're like trying to answer the question and you don't have any of the context for what you're trying to to ask. and

now that it does have the context,

Carter (1:32:34)

But this gets back to your

Carl Brown (1:32:35)

it just doesn't know, you know, go ahead.

Carter (1:32:37)

but but I was gonna say this gets back to something you've said because I I asked in one of our other interviews, I just like, why do they why do the labs and they backed off this talk a little bit because they realized that they were gonna their public opinion was in the toilet, but I'm like, why do they talk about replacing old jobs? Why do they talk about you know, like potentially killing everyone? Like, why don't they just talk about all the things people like about the product? And you said there's too much money involved, right? Because if if

Carl Brown (1:33:00)

Well absolutely.

Carter (1:33:01)

if you say, it's just a better search.

bought which I we all agree and we all really, really like that about it. Well that's not worth a trillion dollars. Right.

Carl Brown (1:33:09)

Right.

the so my my belief is the it's gonna kill everybody is to distract from the actual people that are getting hurt.

Carter (1:33:22)

Mm.

Carl Brown (1:33:22)

Right. If you keep everybody worrying about I mean like there's I I made a video about this, but like there there's a if you keep everybody worrying about whether it's gonna make us all extinct.

Then you can drive conversations away from what about all these teenage girls that are having classmates that are turning their Instagram posts into nude pictures and then passing

Carter (1:33:42)

Right, right.

Carl Brown (1:33:43)

them around the school, right? What are we gonna do about those girls? we can't worry about that because we've got to make sure we don't go extinct. Right. and that just infuriates me.

Nathan Toups (1:33:52)

Right.

Carl Brown (1:33:53)

Right. So I think that's where a lot of the dooming the doom stuff is coming from. Is it's like, you know, if we can keep everybody focused on that, then we don't have to be responsible for.

You know, the fact that their AIs are hacking other companies or the fact that they're making nude pictures of underage girls or, you know, telling, you know, telling teenagers that they don't need to go seek mental health and by the way, I'll write your goodbye note for you kind of thing. before you create before you, you know, make a choice that's irrevocable and will harm yourself forever, right? those kinds of things they're just trying not and the all the intellectual property theft. they're just trying to keep everybody from from thinking about that stuff.

and I think that the reason that the job replacement thing, well, so there was a period of time a few years ago when that might have been the case, right? Before the the curve started to go sideways, when we were still looking at the exponential, everybody's like, okay, you know, we don't know where this is gonna end. We don't know what the limiting factor is, we don't know what's gonna happen. And the replacing jobs thing is a thing that the managers

That are paying for AI and that are making the like doing the budgets for next year and all that kind of stuff. It's a thing they understand, right? And you know how much money is being spent on white collar work in the US every year. Right. And so if you say we're gonna take half of that, well, cool, you know, the the the GDP of America is $32 trillion a year, right? Anthropic is now saying that they're

Potential addressable market is $30 trillion, right? That means basically all of the jobs in in America are gonna go

Carter (1:35:26)

Right, right.

Carl Brown (1:35:26)

to to you know ClaudeBot. That's ridiculous. But by saying we're gonna replace all the workers in the world, well, now you've got a thing that the finance people can hook into and go, okay, well, I know what that means in monetary terms.

Carter (1:35:39)

Right, right.

Carl Brown (1:35:40)

so I I really think the that the the the discourse around it,

in the media is more driven by the financials and the desire for financials and the desire to keep the hype train going than it really is by the value of the actual product. and I I I don't the thing that that I don't understand

Actually I do, but th the thing the thing that drives me crazy is the way that nobody in the press is skeptical at all. And the way they and and this has been happening for a while, but the way that they repeat press releases as if that was journalism.

Nathan Toups (1:36:32)

Mm.

Carl Brown (1:36:33)

Right? and they amplify things and then the you know, the the clickbait people take the thing that the journalist said that was based on the

press release that was based on a spinned interpretation of the actual what happened. Right. And then the click great people may make that even more sensational so they can get more clicks and that kind of stuff. And then that's what the public gets as its kind of first level understanding of what's actually going on. Then, you know, it's not surprising that, you know, people are up in arms about this stuff. And it's not surprising that, you know, the the

that people aren't talking about what consumers actually want from this, because what consumers are seeing about what they should want from this is, you know, click baity spin of clickbaity spin of clickbaity spin of what actually is going on.

Carter (1:37:32)

Well, we as always, we could talk about this for hours and hours.

Carl Brown (1:37:36)

Right.

Carter (1:37:38)

but it's always great to have you on the podcast, Carl. And and we appreciate the work you're doing. whether it's your responsibility or it's just something you you like to do, right? I I think it's so helpful. I it's something I wish we were better about in the industry, is kind of like it's not paying it forward than like sharing our experience. We we value

kinda like youth and exuberance or whatever so much in this industry when when what something I've found so valuable from this podcast is talking from people who have been doing this for twenty or thirty years longer than I have and realizing that, you know, this is just another change in a field that is constantly changing. And no one can see the future, but you know, if if it's anything like the previous changes, we're gonna be fine. And you know.

Carl Brown (1:38:24)

Yeah. And and I still

would rather be a programmer than, you know, a reporter or a you know,

Carter (1:38:29)

Right, right.

Carl Brown (1:38:30)

a a a stock photographer or, you know, a bunch a bunch of other occupations in this going through this particular transition. So it's not easy, but we're better positioned than a lot of people are.

Carter (1:38:43)

Well, yeah, I and I I I agree. And so if you guys ever want to hear any more of this analysis, I mean, Carl's always on if you're catching this from his channel, you know his channel. and Carl is always kind of talking about the the sort of state of the art frontier stuff and his reaction to it. We're always here on book overflow reading the best technical books in the world and just trying to improve our craft and get better at this constantly changing field we call software engineering. and yeah, Carl, just thank you so much for coming back on. We we love having you on here.

Carl Brown (1:39:11)

Thanks for having me. I love these

conversations.

Carter (1:39:14)

Well, we do too. All right. We'll see you around, folks.