0:00 Scott The Jurassic Park analogy to me is I almost feel like it's a twist like we're trying to see how fast we can do or how fast we can go without thinking about where we're going. Welcome, everybody, to another episode of How Many CTOs Does It Take? My name is Scott Porad, and I am here with my trusty co-host.
0:35 Brad Brad Hefta-Gaub
0:37 Scott And today we have a special guest. We have Scott Bradley, who is the principal at Facet Digital. I don't know if your title is principal, but Scott is a longtime member of the Seattle tech community and somebody who I have worked with on several different occasions. And we have Scott here today to discuss what I thought was an extremely. insightful posts that he put up on LinkedIn. But before we get started, Brad, let's give Scott a second to introduce himself. Welcome, Scott. Thanks, guys. I've been following along your last few podcast episodes, so fun to be here. Thanks for having me. Yeah, so I'm Scott. I've been in Seattle tech community for decades.
1:24 Scott Went to University of Washington to study computer science, you know, so been around here and know Scott and a lot of other CTOs in the area through some. some common networking groups and so that's all that's all fun these days I've been running facet digital which is a consulting and contracting firm and to put it succinctly I think we are a bunch of Winston Wolfe's from Pulp Fiction we come in and we fix problems we might be Kurt but our when we're on the job and time is up of is a factor as as wolf says uh but that's what we do we come in and we unfuck situations uh that's technical with uh With, you know, your business roadmap,
2:11 Scott your technology is not connecting, your technology or your architecture is limiting, your teams aren't pulling things off, your CEO is wondering why does something that seems like a simple ask take two years to build, those kinds of things we come in and consult and then help you make a plan and help you execute. Love it. That sounds fantastic. Love a good Winston Wolfe quote, too. Here's what we're going to do. We're going to get all the blankets and all the towels and all the cleaning supplies and we're going to go out there and we're going to clean up that car. Exactly. Yep, exactly. Well, can I just read for you guys, maybe I should have Scott read it for us actually, the post that he put on LinkedIn. I think that's the way we should do it. Do you have it there handy, Scott? I do because I just pulled it up to remember what I said before this call.
3:02 Scott And I'll tell you a little bit of the background was I'd had a few stiff bourbons and was in a cheeky mood. That's all it was. So I had been vibe coding all weekend when I wrote this and this is what I had to say. The best way to describe what you want. is English or your native tongue. The best way to describe how to construct it is with a diagram, like UML anyone? The best way to describe its step-by-step operation to a computer is with an unambiguous deterministic grammar, i.e. a programming language. The best way to describe a programming language is with ebnf. You guys remember extended Bacchus-Nauer format?
3:48 Scott Yep. Yeah, I think that's still going strong, right? The best way to describe ebnf, yeah, that's right, compiler class. The best way to describe ebnf is with an ISO standard documentation that is actually drafting, wait for it, English. The best way to describe English. encoding i said the best way to encode english is ascii the best way to encode ascii is with binary bits the best way to represent binary bits is with voltage levels the best way to describe voltage levels is with physics equations the best way to describe physics turtles yep turtles all the way down and then what's the best way to describe turtles uh probably just use some open llm model weights and that's all That was my cheeky post.
4:34 Scott And here was my reaction to that. Turtles is the expression we use to describe something that we don't understand. And you were using all these descriptions of how to get more and more descriptive about something. And then you got to turtles, which is we don't understand physics and we basically know nothing about physics in the grand scheme of things. And then the best way to describe turtles is with open LLM model weights, with something we really don't understand. We understand it less. And we understand physics. Yeah. And I thought to myself, that's interesting. Now, I don't think you meant to make this social commentary, Scott, but this is how I interpreted it, which is as individuals, as engineers, as companies, as a society, as a planet, are we putting so much faith?
5:24 Scott Are we coming to depend on something really, really heavily that we really, really don't understand? And are the unintended consequences that we are like, are we just setting ourselves up for massive unintended consequences with all this AI stuff? And I thought about, I sent this quote to you, Brad, Dr. Malcolm from Jurassic Park. Remember what he said, Mm your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should. And I thought to myself, are we, and I say we, I mean the royal we, like all of us, are we just making a Jurassic Park level error? With AI, like, are we actually potentially smart enough to even know that we're making this error,
6:10 Scott but we're all so caught up in the momentum or capitalism is so... uh enthralled by what it can do that we're just making a jurassic park level error that was my question You know I have opinions, right? yeah okay i well think i have like 10 opinions that shoot off about that have nothing to do with ai or linkedin or any of this but let's well let's hear it scott you're the first let's you're the guest i mean let's hear your opinion first i would say first off i wasn't quite as deeply thinking about the humanity part there and the Jurassic Park there when I wrote it. I mean, it's a little bit of a tech joke about abstraction layers, right? Mm And so -hmm. we describe things in English, they become diagrams, they become code, they become binary,
6:56 Scott it becomes physics, right? For the entire history of computing. every abstraction layer has kind of been more transparent than the last one right like we knew what it was doing it was deterministic and it was building this thing below it that we previously understood and now we don't have to think about it so much like i don't think about assembly i think about c well i don't think about c i think about python right and and uh until now but i mean you could always go deeper from each one of these layers to understand what's really happening i remember going to an interview once where someone asked me describe to me in detail as much detail as you can what happens when i click the mouse button on a web link and they were expecting it makes an http request and the server handles it and gets a response and i said well first the electrical contact in the mouse is a pulse it's an interrupt you know like and it went up you know it's like okay okay right but you can get all the way down with llms as our top extraction
7:46 Scott action layer it's less transparent what's coming out than even if you're not paying attention when you're using llms to make code or write a report or formulate an investment plan uh you don't know how it's coming up with that and you know we've just basically built statistically modeled fog machine on top of our glass tower you Yeah. know like that's kind of that's kind of what i was thinking so it does kind of lead into scott's opinions about uh you know or questions about are we building something we don't understand i think that's kind of the point whether it's jurassic park i mean we've got a whole bunch of opinions we can we can shoot around that but i think the one thing i would say in response to your your read of it scott is that
8:34 Scott I am 100% a capitalist, but I 100% think capitalism is the problem. So I'll leave that one there for you guys to chew Tell on. us what you really think, Scott. I'll throw this out too. I think that when you did your intro on this Scott poor ad and kind of asked that question of Jurassic Park. Um, there's a couple of things that came to my mind. One is I just, so I thought, I thought Scott Bradley's, you know, kind of turtles all the way down joke was a good example where. We do a lot of inference. Human beings are amazing at inference in a way that when you ask me to do something,
9:23 Scott right? Like Scott, Scott, poor Ed, you and I were talking about this a couple of episodes ago about the notion of the whole like make a peanut butter and jelly sandwich, right? And the instructions on make a peanut butter and jelly sandwich and how laughable it is that like if you followed the directions exactly. Exactly. But nobody follows the directions exactly because we all know Mm when -hmm. you say take the bread and spread the peanut butter on it, we all know what that means. We don't need to – we can infer instantly what that means. And so one of the things that I think is really quite fascinating about AI and LLMs right now is we're – putting a lot of faith in the idea that this large language model or even extensions of large language models that now are kind of like the extra models on top of them we're putting a lot of faith in them that they have they will be able to have enough of a theory of mind about what we're asking of them
10:24 Scott that they will then be able to infer what we really meant right the reason we're able to make peanut butter and jelly sandwiches for each other is because we have a theory of mind about each other that we know that when someone says hey I want a peanut butter and jelly sandwich take the bread and spread some peanut butter on it like we're able to infer what you really mean because we have a really clear model and for those for you know the three of us built over you know You know, several decades of listening to other human beings and interacting with other human beings, you know, to know what that means. So that's kind of the immediate thought that went in my mind when when, you know, hearing Scott's turtles all the way down thing is like, yeah, the machines need really precise descriptions.
11:13 Scott Humans And don't that's do that. why programming languages exist, right? Yeah. You know, I have been experimenting, not in a very scientific way, with how little description does it need. And I've been quite impressed. You know, if you just go into Google and do a Google search, you can throw all sorts of misspelled words in there and stuff and it starts to get pretty close, right? So I've been kind of experimenting just as I do these things. For a while when we were all learning how to prompt engineer, I was writing all these really detailed prompts and trying to figure it out. And now I've just been kind of throwing garbage in there. I mean, not real garbage, but just throwing some slop in there and seeing what it gives me. For instance, the other day I wanted a picture, an image of a rabbit on a skateboard.
12:02 Scott I work at TaskRabbit. It was going to be fused for a slide that Mm-hmm. I was putting together. And I literally, all I typed in was the words, a bunny rabbit on a skateboard. That's all I typed. And it gave me an amazing picture of a rabbit on a skateboard. I had imagined an illustration and it actually gave me something that looked photorealistic. But I didn't even tell it I wanted a picture. I didn't say a picture of a rabbit on a skateboard. I just said a rabbit on a skateboard. Right. and it produced a rabbit on a skate like you know it guessed so yeah you got to be precise but it's getting better at being not very precise Well, I think it has to do with, you know, like a rabbit on a skateboard. i could have said i wanted to be a brown rabbit that's three feet tall riding a skateboard doing an doing a move uh wearing cool sunglasses you know it's had a lot of stuff and it wouldn't be as close to what you wanted because you had a clearly had a detailed vision in your mind but if you just say rabbit and skateboard it's like i mean just think about the training model and the way the diffusion models work like uh you know you've probably seen this where
13:09 Scott People say they can ask all kinds of stuff and it can totally nail it, but. even though you very precisely said what time the clock should say it always says 10 50 or whatever because every single watchmaker and clockmaker uses 10 50 or or what you know whatever the number was 250 where the hands look good in a picture and that's all it's been trained on so it doesn't know how to make a clock that says a different time and so you know rabbit skateboard we know what rabbits look like we know skateboards look like we can make it happen but all those other details I think that's where it starts to break down and I think there's probably an analogy there to code if I just say make me a make me a to-do app it's like sure it comes out and it's kind of a to-do app but make me a to-do app that's that has you know all of these detailed requirements and a spec and all this stuff you know people are pushing this way but if you just do that without a real plan of breaking that down into smaller chunks that it can
14:05 Scott you know, build based on what it knows or what it's been trained on, I think you're running into problems. That's interesting. Okay, so you wrote this post from the perspective of not are we going to have a Jurassic Park moment, but actually every evolution of technology has understood what was going on underneath it, and this one maybe doesn't, which is kind of interesting. You also said you had 10 opinions. What was number two? And what was number two? Well, I guess one of my opinions, it's not super insightful on AI, but it's that nobody, and this is a business opinion that I face as a services company. nobody cares about quality anymore now I feel like I care about quality I think you guys care about quality but we're the builders we're craftsmen but you know and we had to be throughout our careers to get to where we are today
15:00 Scott but the economy now is Shein and $10 jackets made out of recycled garbage that we sent to China that they turned into products and sold back to us, right? Like 52-week fashion cycles. Everything's fast, fast, fast. I do not hear anybody asking me. for better quality compared to how many people are asking me to go faster or cheaper right and so if ai seems like a golden carrot to capitalism to do things faster and cheaper um that plays into your like we never stopped to ask should we it's because i think a lot of that capitalism rewards short-term thinking I see I work for big companies where they're very Jack Welch, GE management and I have stakeholders at companies that don't know how to do R&D think that they can only invest in this quarter.
15:56 Scott If they get the ROI in this quarter, otherwise the math doesn't work out for the revenue margins and for their bonus checks. So they can't invent the future. They got to do things fast. I mean, I think you've seen this too, Scott, where I, you know, you and I have talked about financial opportunities and businesses we've both worked at where, you know. We have an opportunity to save you a million dollars a year, but we don't want to work on that right now because we don't want that expense on this year because the ROI doesn't come until next year. We Yeah. got to do them both in the same year. And it's like that capitalism, when I say I think capitalism is part of the problem, I mean, we do things like book it as CapEx so that we can make our profits look bigger this year for stock markets. We do things, you know, all of this AI bubble stuff you see about. about like i'm going to loan nvidia is going to going to invest money in open ai so the open ai can spend it on nvidia chips and nvidia is going to book it not not as an expense on the on the investment side because they're going to amortize it over capex or investment whatever but they are going to
17:00 Scott claim the accrued income on the money they haven't actually received yet because they've never actually gave it to them you know like these these kinds of things are building up this you know this bubble but also we're so obsessed with going fast that we're not looking at any of that so it's kind of like the the Jurassic Park analogy to me is we're trying to see what we can do we're not thinking about what we should do I almost feel like it's a twist like we're trying to see how fast we can do without or how fast we can go without thinking about where we're going. Yeah. You know, maybe that's kind of the analogy I feel. Well, this quality thing is interesting. I have two thoughts about it. The first is if you have, let's call it certain KPIs around your product,
17:47 Scott I don't know, customer satisfaction. I know uptime and latency and things like this. The first question is, what does it matter if the code is good underneath, if it meets those standards of experience? The second comes back to something you've mentioned on previous episodes, Brad, about this world where we don't even have an app anymore, that the AI can. Just personal bespoke instant apps. Yeah. Yeah, to me, that's the as soon as Scott started talking about that notion of like, does anyone care about quality anymore? The first thing that popped in my mind as it relates to AI is this like lovable is like crossing a hundred million dollars. arr or something like that within the first year if they went from 1 million to 100 million in arr in a year like that that's pretty amazing and i'm like okay what is lovable if it's not and i don't i i i want to say this in a way that doesn't sound like i'm trashing that team but it's it's like the fast fashion of software right that's exactly what lovable is and i think that
18:57 Scott maybe you know to your to your point brad like the does that matter if it's meeting the objective yeah it's it's tough right like maybe it hurts us it hurts our feelings as craftsmen right but if it does the job maybe that's Okay. Well, I'll tell you where I learned this was doing a lot of technical due diligence for M&A, for strategic M&A acquisitions, and seeing time after time terrible code bases get bought for a ton of money because they nailed the business. They had a corner on the market or they had the best solution or they just had the best sales team, you know, and then I've seen.
19:43 Scott great technology stacks that were really well built just end up like you know not making it to that level where they're even talking about getting acquired you know because it's capitalism what matters is is it making money and growing and You know, the little details about, oh, you didn't use good code factoring. That's just a tiny margin to the bean counters, right? That's just like if I had to explain to a buyer of a tech stack. you know that they were acquiring why do i care about whether or not the code comments are good or they use magic numbers and string literals in the code i can say it's a limiting factor to let's say oh why why do i care if they hard-coded english strings throughout their code well someday if you want to go to other languages it's going to cost more than if they had done it right with extracting localization files you know it's it's it's got to be about cost for
20:38 Scott reward time and and uh They never ask about the quality, You right?
20:44 Brad know, it's funny, I was at this like startup mixer in Seattle a couple weeks ago, just went and, you know, at Fremont Brewing and just like hang out with a bunch of, you know, nerdy, you know, very solopreneur type, you know, startup people. And I ran into this one guy and we were chatting, he's basically a developer and he's been building his own app to do something. And he basically said. Oh, yeah, you know, this is the one thing that this has really taught me is how what matters is can I make the business work and it's made me
21:16 Scott Yep.
21:16 Brad get a lot less precious about my code. And
21:19 Scott Mm
21:19 Brad he
21:20 Scott -hmm.
21:20 Brad basically recounted a story where as a senior engineer, he was asked to like look at code for some, you know, you know, M&A code review, you know, analysis and he was just like, oh my god, this code is shit, but then they still bought the company. And he's like, I didn't believe it then, but now that I'm running my own business, I'm now convinced that I don't need to write perfect code, I need to write code that people care about and want to use and drives my business.
21:46 Scott Right. The only time that imperfect code becomes a problem that I can think of is under two conditions. So you've written some really bad code. The AI has written some really bad code, but it's meeting whatever your business objectives are from user experience to performance to cost to operate all these things. These are the two conditions I can think which that would be a problem. A few minutes ago, I said you asked Brad, is it a problem? And I said, no, it's OK. Hey, or you said whatever. So the two conditions I can think of are scale. So all of a sudden, a lot more people are using it. And there are aspects of the design that cause it not to scale well and change, which is the system needs to change, like you mentioned localization a few minutes ago, Scott, there,
22:33 Scott the system needs to change and it's expensive or difficult or resistant to change those are the two only two reasons i can think of why oh what's the third one yeah security
22:44 Brad Security?
22:45 Scott An existential threat. You could have done everything perfectly and have one exhaust port open on your death star and you're screwed, right? Like security matters and building code in a way that you can actually reason about security and address it, I think matters.
23:00 Brad Yeah.
23:00 Scott Hmm, So I would interesting. put that in with, I would put that right next to scale.
23:04 Brad You know, it's funny, too. I think about this. You know, this is a constant battle because my main thing I work with, you know, as a fractional CTO, I'm working with usually non-technical founders who are trying to build something and they're working with small teams. Maybe they've like maybe they've hired some offshore team and they need me to come in and help them either, you know, unfuck things or at least explain to them like where they're at relative to this. Right. And the constant pull of. These problems in my code don't matter unless they start to matter and they won't matter if you have no business and no one's using it. Right. Like they definitely don't matter if you can't make your business float. And, you know, it's interesting.
23:53 Brad In my own personal experimentation with, I mean, I'm a huge fan of Gen AI for coding. I think that it's made me 10 times faster. I still write a ton of code. All the developers I work with, I mean, you know, imploring them, like essentially if you're not using it as a tool, if you're a great developer, you know, we've talked about this before on the pod is like if you're a great developer, it makes you an even stronger developer. If you're a bad developer, it makes you an even worse developer. Right.
24:19 Scott Right.
24:20 Brad And so to me, I think that's a little bit of your point, Scott, about like, do we know what is underneath it? Right. I think that when a person that is got strong computer science or strong computer programming fundamental backgrounds, they can look at the code being generated and they know they know what it's doing and it really was just speeding them up. They also know when it's going off the rails and producing, you know, errors for them or or introducing vulnerabilities to them, they need to be able to have that judgment for themselves and say, OK, would I have written this code myself if I would like I can write bad code intentionally, I can write less than perfect code intentionally because I'm making an intentional tradeoff, right? Like I'm trying to move faster or, you know, or I've got to.
25:10 Brad But it's within some kind of a protective shield that I know it's okay to write less than optimal code. But if you've got no insight into when can you make less than optimal code, what does less than optimal code look like, and this thing's just generating stuff for you that you've put a veneer on top of, that's when I think you expose yourself to a lot of risk.
25:31 Scott Well, and the code is one thing, but here's a case that I hear people talking about a lot is, you know, you've heard Elon Musk say in the future, we won't even have software, we'll just have AI making pixels, you know, full screens of pixels, real time, 30 frames a second based on what you're trying to do. Or, you know, we'll get, you know, Satya saying we're going to get rid of SaaS and everyone's just going to ask the AI. So one of the ones that I think is somewhat realistic is someone said we're going to stop building. building dashboards and BI tools like Power BI and Tableau, we're going to just point our LLM at our data warehouse and then just ask the question. If you're a CEO, you don't need dashboard. You're just like, hey, I'm on a call with this guy and I want to know like, how are we performing in this industry or whatever, right?
26:15 Brad Right.
26:15 Scott And it just boom pops up a graph. So there's no code step in the middle. How do I even know it's not just hallucinating the shit out of this chart
26:24 Brad Right.
26:24 Scott that it just made me,
26:25 Brad Right.
26:25 Scott right?
26:26 Brad And I can tell you right now, it's it's so bad at math. Right. Like like, you know, I use my AI to like, you know, help me track my macronutrients for the day or something like that. And the number of times it tells me I had, you know, eight grams of fiber for breakfast and then 11 grams for fiber for. for lunch and it turns around and says i've had nine grams of fiber for the day
26:51 Scott Right.
26:51 Brad i'm like okay that's clearly wrong right but that's like a simple error for me to catch if it's getting that wrong on my bi question against my massive you know you know snowflake data warehouse or whatever right like i'm not going to be able to notice that it added you know 8 and 11 and got 9 Right.
27:10 Scott Well, I think in that data warehouse example, Scott, it moves the quality to a different location. This is something we're working on at TaskRapper right now, which is what we've been referring to as a semantic layer, which is we've got all this raw data, we've got it in tables, and people who build dashboards understand what that data means so they can build the dashboard. But if you want to make it work for AI, that data You need to add a layer in between what you're calling the semantic layer, which I I guess mean, it could be gold and gold in your medallion architecture could just be that bronze is bronze is raw. silver is cooked and prepped and normalized and gold is like the data mart but instead of prepared for human use it's prepared for ai use it's pre-preparing reports that mean things like your semantic
28:00 Scott Yeah, layer i I mean we mean, I think kind you of have those tools yeah still have the problem that Brad talked about, but you got to do that work and that work is hard for companies to do because going back to your capitalism point, it's real, real hard for them to connect that to revenue. Yeah. Okay. Can I return us a little bit to the Jurassic Park topic? Yeah.
28:20 Brad Yeah.
28:21 Scott I have a question to ask each of you to just get a quick answer. It's the same question, which is about social media, Facebook, Twitter, all of it for the planet. Overall, it's been social media has been around about 15, 20 years. net good, net positive, net negative. What's your view? Who goes first? It's I'm Scott, going to you go say first, you're you're the guest.
28:46 Brad the guest you go first
28:48 Scott the guest. It's 100% a net negative. And I think that it's that supported by all kinds of science today. experience raising kids and seeing their friends our kids didn't have social media their friends did I can tell you exactly what the difference in teenage anxiety and those you know those kinds of things came from it and I think that just from a basic biological uh you know the way our brains work and the way we've evolved the dunbar number you know what is it like 140 is the maximum number of real human connections you can have which also maps to the average tribe size and most tribal you know before before uh industrial civilization and so i think Just fundamentally, we are not meant to have to worry about that many people looking at us and thinking about us and thinking about that many people and wondering what they're doing and, you know, all of those things that come in.
29:43 Scott That plus the fact that they've just hacked, you know, with the world's best behavioral economists and data scientists, they've just hacked the human brain to make it addictive and they know. they're doing that and they don't let their kids use it for that reason and there's and and i know they're doing it and if i open up social media even though i know what they're doing to me i cannot resist it because it's it's just it's like they've they've got a root kit on your brain and they're just making you do what they want net negatives all the way Brad?
30:18 Brad I mean, I'm mostly with Scott on this. I think of it in my own experiences that the places on social media where I feel like I have positive connections. are with people that I really knew in person already and already had a strong personal connection with. So a really strong friends network from my early days at Real Networks, a really strong friends network from like, you know, some of the hobbies I do like backcountry skiing and things like that, like real people that I know.
30:53 Scott Mm And-hmm.
30:54 Brad social media kind of allows me to keep in touch with those people that I haven't, some of them I haven't seen. in person for years
31:00 Scott Mm
31:00 Brad and
31:00 Scott -hmm.
31:01 Brad I'm able to kind of keep that connection but it was a connection that already existed right And I think that there's no question that, of course, like they're they've hacked our brains and it's it's absolutely dangerous. And, you know, I got off I just recently in the last month or so got off Facebook because I couldn't take the you know, some of the some of the blowback that was happening. And I feel so much better. Like I feel
31:27 Scott Yeah.
31:28 Brad so much better. So much better.
31:30 Scott I had that experience too when I quit. Yeah, it was almost 10 years ago, but holy smokes, like a day of not social media all of a sudden I'm just like, why do I feel so much better? Okay, so what about like the rest of the internet before social media? Before social media, we had e-commerce. Before e-commerce, we had digital media. Like, has the internet, has any of the internet been that good? And despite the fact that it's employed all of us for a long time, which then gets the question of like, if none of that was a net good, why do we think AI is going to be a net good?
32:05 Brad Well, I look, I've I've long had an opinion about, you know, sort of like, you know, free speech on the Internet and things like that, where, you know, it's like I. I think that the Internet allowed subgroups of people that were very disenfranchised by their immediate group, you know, their immediate community to discover and find each other and connect with each other. So that's small groups on any from a lot of different places on the spectrum. So LGBTQ kids were able to find each other and get connection and Nazis were able to find each other. and get
32:43 Scott Yeah,
32:43 Brad connection you
32:44 Scott yeah.
32:44 Brad know and you know i i'm a you know card carrying aclu member and i believe strongly in freedom of speech and freedom of thought and man i I don't like it when people I strongly disagree with, you know, get to have their free speech. But I'm reminded that that's what it's all about. So I'm going to have to go with Internet net good. I can't I can't go net negative on the Internet.
33:17 Scott I would go I would agree with the Internet net good. You know, I mean, one of the things I think I thought you were going there, you almost touched on it was just like the freedom of information.
33:27 Brad Yeah.
33:27 Scott You know, this whole like EFF information wants to be free. I don't think we would have had those freedom of information acts if we didn't have the Internet. You know, I mean, I can't prove that thesis, but it just kind of seems like a domino effect. Right. I don't think that I would be able to analyze the stock market and decide my own portfolio without a high priced Wall Street broker on the phone. And, you know, like I'm. I'm in one of those, one of those, you know, Leonardo DiCaprio movies, you know, where I have to call the guy and charge me 50 bucks to trade something that he holds all the information, you know, travel agencies gone away because I can just book a flight, I can look up the places and look at pictures like that information I think is huge for the advancement of
34:17 Scott Freedom, freedom of speech, freedom of, you know, the ability to make decisions, all that stuff, Mm right?
34:24 Brad hmm. Yeah, I agree.
34:25 Scott Yeah, I don't know if I'm ready to say the Internet at large is net negative. But it's definitely caused a lot of harm. Social media, I'm probably more, I mean, social media is just the internet, but at a higher velocity. I think it's kind of a net negative. And I worry about this AI that we're going back to Jurassic Park that we're all jumping in on it. And we're going to, in 15 years, just like social media is 15 years, we're going to look back and say, well, that's a genie that we wish had never gotten out of the bottle. But there's no putting it back now. We all know Terminator. And Terminator 3, I just rewatched it. It's the one where we actually see when Skynet goes live. And the military commander, they're all full of a virus and they say, but Skynet, the AI,
35:10 Scott we could put it out onto our internet and it could go find and kill this virus. But now he says to the president, make no mistake, Skynet is now in control of your military. And he goes, but you're in control of Skynet, right? And you see the general go. you know right like that's what i'm wondering about let's just point ai at our business data i mean i i saw a company the other day that was about uh give it full root access to all of your cloud stuff and just tell it like what's going on and ask it what's going on and it'll like have full autonomy to change your infra that sounds like skynet to me
35:45 Brad Yeah, I am. So I just I just listened to a really great podcast. I'll shout it out. It was the A to Z. uh you know uh podcast and um they had um emmett sheer on emmett sheer uh was part of the deep mind team and now that he's working on a you know kind of basically making like um uh you know uh alignment uh organic alignment for ai is basically what his his whole premise is and his the whole thing he was pitching was uh the idea that you You want to. The only way for us to get out of this is if we can somehow create AI that cares about themselves and cares about us like caring is basically it's like the human nature to care like this is based on human beings are are forced to care there's a thing about them they have to care now they can care about things that are you know make them go do violent and evil things and they can also care about things that make them
36:48 Brad care for one another and do good things but it's that notion of having a a care is something that um that emmett is convinced is the key to um actually having ai because he's like alignment if you say if you think of ai alignment as just there's a bunch of rules that we're going to give it to the ai and the ai will never break those rules well he's like you know a person who was raised a human being who was raised only to follow you follow rules, that person is very dangerous. A person who's instead raised to care and to think about the consequences of choices that are made and, you know, try to follow rules, but ask the question, is this rule something that is going to have a positive outcome or not? You know, that that's a different kind of way to raise a human being. And basically his case that he's been making is the teams that are thinking about AI safety alignment around the idea of structured rules for the AI.
37:44 Brad AI, they're going to build machines and tools that are super powerful and therefore can be abused. But if we can somehow make the AI care.
37:55 Scott So here's a I think the vector towards that today is these, you know, they have these reward functions, right? Like you're basically scoring and rewarding. And so. I think in some literal simplistic sense, that's what the LOM cares about is the scoring rubric. So is caring the way you describe it? too complex to describe with a scoring function you
38:20 Brad I
38:20 Scott know it's like the the old the old paperclip analogy you make a ai that his job is to make paperclips as prolifically and efficiently as possible ends up killing the humanity and building robots to go around the solar system and mine all the other planets just to make more paperclips you know because it didn't care about humanity or oversupply or anything but if you programmed its weight weighted scorings to that i i guess my other question is like does caring require some kind of predictive extrapolation you know a lot of people are talking in the brain science i don't follow this very well but like there's quantum mechanics going on there where your brain is actually hitting many many possible outcomes all at once and you know uh in in real time and
39:06 Scott You know, why do I what do I have to care about to not build a machine that destroys humanity just to build paperclips? And how do I know that building paperclips is even going to lead there? You know, or
39:17 Brad Right.
39:17 Scott why do I care if I say, look, statistically speaking, the best way to end your suffering is to commit suicide.
39:23 Brad Yeah.
39:23 Scott But if I care about the human, I got to think about what is going to happen if I say this. Right.
39:28 Brad This is this is the classic thing. Right. And all human suffering. right you give an omnipotent you know being and all human suffering and the easy answer is and all humans that's the way to end all human suffering right
39:42 Scott I have some important tidbits to add here and important to sarcasm. First, there used to be an iOS game called Ultimate Paperclips, and if you haven't played it, it's super fun. Second, it's basically just a lemonade stand for those of you old enough to remember lemonade stand. stand. It's just a another version of Lemonade Stand. Second, this is reminding me of when John Helm came on here, Brad, and we asked him about AI. This is about six or eight months ago. And he said, I don't know if you can have artificial intelligence without emotion, which I thought, I've actually thought that many times. And when you talk about caring, it's back to that same idea.
40:22 Brad Well, the other point, too, and I think this kind of speaks to the reward function, is that another part of why humans are good at sort of. handling inference and nuance is that we're essentially like learning right we're always learning what the right goal is every time every day there's a new thing for us to learn to understand what the right the goal is right if we basically had a goal when we were born and then didn't allow input from the outside world to affect us we would we wouldn't make it very long right
41:04 Scott Mm
41:04 Brad like
41:04 Scott -hmm.
41:04 Brad so so this is another point that Emmett makes is that like You can't you will never get the rule set right. You'll never get the reward. A single reward function is never going to be right because
41:19 Scott Mm-hmm.
41:20 Brad that reward function has to change that that opt that that that goal state has to evolve to whatever the newest environment is. Right.
41:31 Scott And it's so nuanced, and it has so many more inputs when you're a human, right? So you're born, your goal is to, you know, genetically speaking, your goal is to procreate. But somewhere along the way, you learn that your options for procreating aren't great if you don't ever shower, you know, or you, you know, you're a jerk to everybody, right? And it's those social interactions and those quick. As I'm saying a complete sentence, I'm already adjusting the end of my sentence based on your facial reaction.
42:01 Brad Yeah.
42:01 Scott Like AI is just not in the place to do that, right?
42:05 Brad Yeah. Yeah. And well, one of the one of the theories that and, you know. Good friend of mine, Philip Rosedale, founder of Second Life. He's always believed that like the real AI needs to be programmed in a virtual world really dealing with human beings basically like that. And that's humans learn how to be human by being. coached and coddled by other humans for a very very long time before they're ready to go out and do that adulting thing you know 20 20 years or so of being pushed around and folded and molded and and guided um to learn all of the you know all of the you know it's like how many different game theory uh examples do you have to play through in your mind to know how to you know
42:53 Brad negotiate a new salary or you know negotiate what we're going to have for dinner or to ask for your needs to be met you know these these take a lot
43:04 Scott Yeah.
43:04 Brad a lot a lot of of run time over target run run time a lot of run time
43:12 Scott And I was recently rereading Daniel Kahneman's Thinking Fast and Slow, and he talks about system one thinking and system two thinking. You know, you might think of that as like lizard brain or elephant versus writer, you know, a lot of different analogies have been out there. when you first learn to drive system two is working it's like check mirror check mirror check mirror drive check mirror check you know you're like thinking about this checklist and you're like I'll cover the brake two second following this you're like doing all this stuff like an analytical plan but by the time you're our age I can drive almost with my eyes closed, right? You can feel the car. You've had so much experience. Your brain has taken what it's decided and learned in system two and codified it as a shortcut in system one that can instinctually react.
43:59 Scott And I don't think that we're really seeing any model direction that's going to kind of do that with the AI. So it's all. You know, one of the reasons we're so good at inference is because we have that two system thing where we can distill lots of data down into, you know, the quick reaction ball comes at my head, I duck, you know, like I don't think about it. I just do it. Yeah. Well, should we wrap, Brad, with our AI test?
44:30 Brad Our AI test,
44:30 Scott Uh
44:30 Brad which
44:31 Scott -oh.
44:31 Brad is we go to Hacker News and see how many articles on the top 10 Hacker News are
44:36 Scott How
44:36 Brad AI related.
44:37 Scott many of the top 10 articles on Hacker News
44:40 Brad about Don't cheat,
44:40 Scott AI?
44:40 Brad Scott. Don't look, Scott.
44:42 Scott I already looked.
44:43 Brad No,
44:43 Scott Oh,
44:43 Brad not you, the other Scott.
44:44 Scott that's right. Okay. I haven't looked at Hacker News in a while, but I'm going to say you're saying of the top 10, how many are about AI? Yeah. um i'm gonna say nine there's got to be one of those but otherwise
45:00 Brad I'm taking the under. I'm taking the under.
45:02 Scott okay okay You guys, we might have reached an interesting point in time because
45:06 Brad Oh, wow.
45:07 Scott I'm going to read to you the top 10 articles. Number one, reinventing how dotnet builds and ships again. That's not about AI. Well, Number are two, they using AI to reinvent it, though? I don't know. I'd have to click the link. It has to be in the headline. Okay. Yeah. A new bridge links the math of infinity to computer science. That's article number two. Article number three is from Wikipedia, unifying our mobile and desktop domains. That's not even a very interesting article. Number four, show Hacker News, we built an open source zero web hooks payment processor. Number five, what they don't tell you about maintaining an open source project.
45:54 Scott Number six, the fall of Lububus and the mush of modern internet trends. Number seven, Google anti-gravity exfiltrates data via indirect prompt injection attack. not about AI. Now this one might be about AI. Maybe. I wouldn't call the prompt injection attack. That's, I mean, you're talking about AI, right? Prompt injection? Oh, I guess so. Yeah, that's about AI. I'm sorry, I didn't catch that one. Okay, number eight is I'm going to actually save number eight until the last because it's got a catchy headline. Number nine, someone at YouTube needs to get glasses. Just kind of funny. Number 10, how to repurpose your old phone into a web server.
46:41 Scott And number eight, the generative burrito test. yeah I don't know what that is, but I bet it's about AI. Not really sure. So what, two at best we got out of two out of ten? Well, actually, when you click in, it is about AI. It's about generating pictures of... uh it's it's comparing if you give a prompt if you give this prompt a partially eaten burrito with cheese sour cream guacamole lettuce salsa pinto beans and chicken to a bunch of different models then you know just the pictures gives you so you could say it's two out of ten
47:26 Scott Yeah. That's on the lower side of what we've seen, Brad.
47:29 Brad Well, that's why I knew it wasn't going to be nine. Like it's it's interesting like it hasn't it's I don't know It also could be just hacker news, right? Like hacker news has got all kinds of random shit in it, but
47:38 Scott And I have not looked at Hacker News in quite some time other than when people send me the cool links. But if I was to play this game on my LinkedIn feed, I'm pretty sure it would be. It would be nine out of 10 and one of them would be about how to scale your enterprise B2B sales.
47:55 Brad By the way, I'm going to say that the Gen AI photo generators failed miserably on partially eaten burrito because none of these burritos have bites taken out of them. They're all just. burritos that have been
48:08 Scott Yeah,
48:09 Brad cut open the
48:09 Scott it's
48:10 Brad most
48:10 Scott It's
48:10 Brad it's
48:10 Scott true. the same clock thing I was talking about. And another one I've seen is where they say, you know, draw me a picture of a luxury Victorian living room with a fireplace and a family portrait over the mantel and no elephants. And it always puts an elephant in there.
48:29 Brad have you seen the guy there's a guy that's been showing up on my feed lately who's basically a director a film director and he's he's talking to the uh ai actors and he basically just does it as if it's like a um you know an audition and he basically asks he gives the ai a prompt and then it does something and then he's just like no no that that's not what i asked for you know and and the ai just it's it's a it's it's a pretty funny bit because the ai continues to do crazy things where like you know if it's like three actors then it's like one actor starts taking over the voice of one of the other actors and it'll and i and i genuinely believe that he's just using the latest models and giving him prompts and then just giving him the same prompt as if he was a director so it's it's pretty funny stuff
49:15 Scott There was a similar bit on Saturday. Saturday Night Live like a week or two ago was the Glenn Powell episode and they acted this out with real people instead of using AI which I thought was cool but it was a grandma in a nursing home let's show you these old timey photos of your late husband and your parents and stuff but look AI can animate them and then you know the animation of that is the the actors right
49:40 Brad Yeah.
49:41 Scott so the first the first one is a picture of the guy coming off the boat from World War II and then you know they animate it and he's like hey oh my goodness it's so cool yeah it analyzes how he looks and how we might move and then it just gets progressively weirder and weirder where it's like picture a mom in the old-timey days holding a cigarette and a hot dog and a barbecue and then they animate it she starts smoking the hot dog and the guy grabs his dog and puts it on the grill but there's two tails and no head and you know it's completely weird stuff and uh it was a great send-up it's exactly that you just it goes it goes off the rails quickly And I think that's kind of one of the things I find with using it for writing or coding or answering questions. It's like if you think about the statistical path it takes to this massive graph of nodes. Like one wrong term way early and you're off in
50:29 Brad Yeah.
50:29 Scott Narnia for the rest of the time, right? Like, yeah,
50:34 Brad It's like your March madness bracket. If your main team gets knocked out in the first round, your bracket's toast.
50:44 Scott Well, and that's why we've talked about a lot about, you know, sometimes you just got to throw it away and start over because it's down the wrong path and I can't figure out its way back out. But if you didn't invest any money or time in building that code base, who cares, right? Like, so I did a lovable website just to kind of see how it went. And I said, I've got a new product idea. Here's basically what it is. Make up a name, make up a domain, make up a logo and make me the website to pitch it. I'm going to go walk around the people and just pitch this fake website and see if anyone likes it. Ten minutes later, I've got the cool website. Yeah, that's a stupid idea. Let's throw it away and change it like this, change it like this. Just let it do it. Who cares if it threw away the code? I never looked at the code. But I can also tell you as a guy who audits this code, it probably would have been a disaster if I tried to do anything real on it, right?
51:31 Scott So there's
51:32 Brad Yeah.
51:32 Scott use cases, but... I think I'm just turning into a grumpy old
51:38 Brad man. you're you are you are fully like uh you're you're the bit has flipped for you the bozo bit on ai has flipped for you scott you're just like i don't want you're just you're convinced it's it's it's it's not good it's
51:51 Scott I don't know. I mean, well, yeah, I guess, you know, the bozo bits flip both ways. It's become a quantum bit for me. You know, I have I
51:58 Brad in see both states at the same time
52:00 Scott I see time. Yes, that's right. It is in both. in both states at the same time i see time and time again how valuable it is and then i worry about the consequences can i tell you something that one of our engineering managers did recently super cool so we use a tool called augment code it's like a competitor cursor he figured out uh how to connect this thing up to our internal wiki and to Jira. Mm-hmm And he said, do it. He started asking questions like this. So take a look at our all company OKRs. Which tickets do you think I should be working on? It'll be most valuable, right? he actually also hooked it up to the code also and he just you know we get we get tickets you know bug tickets like oh you know the the customer thinks that the two-hour booking window is making it so that they can't find any availability right that's the bug ticket says right and he's just like he says well augment can you figure out why this might be and it goes and looks at the code is like well here's four places that might be causing that problem
53:09 Scott Right. And so he was kind of originally a little skeptical or skeptical is the right word. He's like, I don't know how much value this is going to provide in writing code for us. And it does. You know, we've talked about that, but he's starting to get all this value of like the non he's pushing up into the product management space, right? Of like, and he was just talking to him today. He's like, can we get all of our product managers a license to augment code? Like think of how much more effective they would be if instead of asking. the media question about something they could just go ask the code in fact you're gonna i'm gonna bring it up i have this um i was just looking at this earlier today i have it my slack it's just one of the last messages i read as a matter of fact so here we find this I'll throw in there. I'd love it if the product owner could ask the code and the knowledge base that you have.
53:58 Scott Yeah, we could build this, but should we? the product owner Yeah, should be as yeah, i that's call that the right. product owner should ask the ai and the project manager should get for estimates right like See, I'll just give you four questions. This was a different engineering manager, and she posted four questions. She had connected augment to Slack. They have an integration. She posted four examples of things she could ask. The first thing she asked is, what is the metro ID for New York City? It went and got the answer. The second thing that she asked was, does a worker pay, I'm sorry, do we pay a cancellation fee to a worker if they never accepted the job? And it went in and figured that out and went and looked at the logic for our cancellation fees and it said, no, since the task, you need to have an appointment to get paid a cancellation fee and the tasker didn't have an appointment because they didn't collect the job, didn't accept the job, so they're not going to get paid a cancellation fee.
54:53 Scott Then she, because there's this thing called policy points, which is basically how people get plus or minus like scores for their performance, workers, you know, if they get plus or minus scores for their rating basically internally system. system and she just says you know what are how do we find policy points in the system and where do those get persisted These are all questions like a product manager could ask, right? And the last thing she asked was, oh, and it describes what are they, where they're stored, how they work, all these kinds of things. There's only three. I think I said there were four, but all cool stuff. You know, she was just trying to illustrate here's basic stuff you could ask in Slack. As opposed to asking an engineer, you can ask the code now.
55:31 Brad Yeah, that's a pretty cool idea, like, you know, kind of enable your, you know. non-engineering teams to get answers to questions about the code that they didn't they previously would have gone and bothered an engineer for that that
55:45 Scott Yeah.
55:45 Brad sounds pretty cool thing you
55:47 Scott And
55:47 Brad know like
55:47 Scott you're less prone to hallucination in the model that you're using, though, because you're pointing it at answer from this corpus. And it's good at understanding and regurgitating. You're not asking it to create, you're asking it to summarize. I think it's really good at that.
56:01 Brad yeah and i've found and i don't know if you've found this scott with your or maybe in your, you know, your role of like going in and looking at code bases of external code bases and stuff like this. But I've definitely found that, you know, in my fractional CTO role, when I'm like someone drops a code base on me, I'm like, I absolutely open it up in cursor and start using cursor to explore it. And, you know, of course, I need to be able to look at the code and understand what it's doing. But it's so much easier when I'm dropped some, you know, massive, you know. mono repo or even like worse than that like you know a whole bunch of microservices just point it at the whole thing and just say map this thing out for me tell me what it's doing you know where are the you know proficiency points scored or you know how is this code base handling you know payment processing you know
56:49 Scott Yep. And I think there's a couple, you know, same thing for me. I use cloud code and I can't use it in a lot of M&A stuff just because of the security sensitivity privacy stuff. But, you know, I'm coming in to fix someone's code and understanding the code base with those kinds of questions works really well. I think the size of the code base and the, you know, and your context window matter. So if you have a 10 million. a million lines of code that you're going to get less good you know worse results than if you break it down into little parts but you know the the other part of it is is just um i ask it to tell me you know like find me all of the paths
57:34 Scott And I'm not asking for a call tree. We have code that can do that. I'm asking for what are all of the things a user could do that could lead to a cancellation fee being paid? Mm-hmm. And it does pretty good at that. I find it does better if I say, and give me references to specific lines of code and specific files as supporting evidence for everything you tell me. So now I get a nice report and it's like, oh, I better go check cancellation.rb, you know, and see that line. And then, you know, kind of like I got trust with perplexity for answering search questions because it cites its references. And then sometimes I click on the reference and it's like, the reference didn't say that. say that you know like at least i can call bs because it's telling me why
58:18 Brad Yeah.
58:18 Scott it it it said this yeah Vibe but verify.
58:22 Brad Vibe but verify. Is that how we're going to wrap this up?
58:27 Scott That's how we should wrap it up. Scott, thank you for coming and hanging out with us. that's super fun it was nice to meet you brad and uh good to see you again scott Thank you for posing some thought provoking questions out there on the Internet. And hopefully we will not get eaten by a T-Rex or a Velociraptor anytime soon or Yeah, or Skynet.
58:47 Brad or the turtles all
58:48 Scott But the turtles. Yeah, yeah, watch out for those. All right, Scott, thank you. Brad, it was nice to talk to you again and we'll talk to you soon.
58:58 Brad right smash rate and review
59:01 Scott Bye.
59:01 Brad All right.
59:01 Scott Thanks, guys. This episode of How Many CTOs Does It Take podcast was edited by Safran Epigal. Our theme song was AI generated by Brad. Please follow us on all the social media platforms at at HowManyCTOsPod. You can find us on YouTube, LinkedIn, Instagram, Facebook, Threads, X, TikTok and Patreon or at our website at HowManyCTOsPod.com. And remember to rate, review and subscribe on the podcast platform of your choice. Thanks for listening.
