37 minSeason 3, episode 4

Vibe Coding: What Could Possibly Go Wrong?

With Chris Romeo, Izar Tarandach, and Matt Coles

Secure software development

Vibe coding, or using AI to generate code by describing what you want. We critically examine the concerns surrounding AI-generated code, including code quality, security risks, and the potential for creating numerous low-quality applications.

Part of: AI-written code and secure software development

Episode title image: Vibe Coding: What Could Possibly Go Wrong?

Listen to the conversation

About this episode

Vibe coding, or using AI to generate code by describing what you want. We critically examine the concerns surrounding AI-generated code, including code quality, security risks, and the potential for creating numerous low-quality applications. Our discussion explores whether AI can truly provide foolproof, production-ready code, or if it should be limited to idea generation and prototyping. Catch our candid take on the dangers of relying on AI for software development and the importance of maintaining human expertise in the coding process.



Follow AI Security Table:

➜ Home: https://securitytable.ai/

➜ X: https://x.com/AISecurityTable

➜ LinkedIn: https://www.linkedin.com/company/ai-security-table/

➜ YouTube: https://www.youtube.com/@AISecurityTable

Transcript · 37 min conversation

0:09Chris Romeo Hey folks, welcome to another episode of the Security Rant. No, wait, I mean Security Table. My name is Chris Romeo, joined as always by Matt Coles and Izar Terandash. And as we found in the preamble, which we don't publish, Izar is in a rare mood/state of mind. So get ready. This one's gonna be—

0:33Matt Coles it's not rare if it's all the time now.

0:35Izar Tarandach You're supposed to say it's rare. It's rare in quality, not as in—

0:41Chris Romeo touché, touché. Well, I guess the topic is something to do with vibe coding, which I guess I'm gonna be the old person in the room and go, I don't really know what those two words mean. I don't know what the first word really means.

1:00Matt Coles When you're vibing. Well, since I am, since I do the, since I do the definitions on this table, I suppose I'll start. So this is where Chris says, man, why don't you walk us through what vibing is?

1:12Chris Romeo Oh, it's like, it's like AI has taken over my job.

1:16Izar Tarandach Well, I always thought that it was because of my age, the way that I hold the mouse.

1:21Matt Coles It, uh, it is very fitting you said that because that's, uh, that's what vibe coding is. So vibe coding is basically you express your thoughts and ideas to, uh, an AI system. You know, pick your favorite, ChatGPT, Claude, who knows what, uh, and let it figure out what you want, write code for it. Maybe do some debugging and at the end you end up with something functional. Uh, that can range from, I, I'm guessing here or, or actually from all the research so far, uh, it can range from code, code that you normally might've gone to Stack Overflow for, but you copy paste into your own code, uh, all the way up to and including, uh, a fully functional. Business-ready application. Uh, and for those who, those who have a, have a hard time detecting small audio, uh, small vocalizations, uh, these are just, uh, his brain just went, uh, so, so yeah, so live coding is basically relying on AI to interpret your feelings and get something functional out of it.

2:39Chris Romeo I'm sorry, I wasn't supposed to laugh there.

2:42Izar Tarandach In the immortal words of—

2:45Matt Coles what could go wrong? What are you working on? I just told you.

2:50Chris Romeo What you—

2:50Matt Coles uh, everything.

2:52Izar Tarandach Everything.

2:54Chris Romeo No, even worse.

2:56Izar Tarandach What could go wrong? I don't know.

2:59Chris Romeo So this is, this is just a fancy word to disguise the use of AI to write code for you.

3:07Izar Tarandach Uh, it goes worse, it gets worse. You don't even need to be a developer anymore. So the latest finance bro with, uh, Web3 idea can just pay me away to publish their app and backend in 3 steps easy.

3:33Matt Coles This is the natural progression of the low-code, no-code movement.

3:37Izar Tarandach It's the natural progression of stupidity.

3:42Matt Coles All right. I'm going to play devil's advocate on this one. I think there is a real value. Yeah, there is.

3:51Izar Tarandach Play with it until the machine tells you why don't you go and learn how to code, which by the way has happened.

3:56Matt Coles It has happened. It has done that. Yes. Um, stop being lazy and go figure out how to code. Um, now interestingly enough, I would say that there's a big, uh, opportunity here if you're into agent development, right? A threat modeling agent. And as we know in, in some upcoming, uh, not to preview upcoming conferences, but I'm sure we'll get a lot of that in, in upcoming, uh, presentations on threat modeling. But if you had an AI agent, there would be a benefit here. Uh, so so that, you know, you ask Claude or whatever to, to write your code and then send it to this agent to, to validate it or to threat model it.

4:33Izar Tarandach You know, for, for a couple of conferences now, I'm playing with this idea of always having a, a t-shirt with broken on yellow. And I had a candidate for Barcelona and because of this vibe thing, I can't use it anymore.

4:49Matt Coles Oh, do tell.

4:51Izar Tarandach I wanted one saying my AI has, um, my startup has so much AI, not even I work there.

4:58Matt Coles Well, it would be truthful. Why would you, can't, why can't you use it?

5:01Izar Tarandach Because now it's not funny anymore.

5:06Matt Coles Yes, but now it could actually be a moneymaker.

5:09Izar Tarandach Instead, they went with, just because you can't doesn't mean your AI should.

5:15Chris Romeo Because you can't.

5:20Matt Coles So on the flip side, I mean, truthfully, this is, is this, if we talk about democratization of coding.

5:32Izar Tarandach Oh my goodness.

5:35Matt Coles I know it's wicked early in the morning. My brain is fried already. And here I am bringing up fancy terms. But you meant the democratization of coding. This allows people in high school or grade school to start building functional applications. Cool. Whether they're shooting—

5:54Izar Tarandach Let's talk about the democratization of medicine where we have people in high school making diagnoses.

6:03Chris Romeo Sounds like a good idea.

6:04Izar Tarandach Sure. Right? What could possibly go wrong?

6:07Chris Romeo It'll be cheaper. They're not going to charge us as much. They have WebMD too. They probably had—

6:16Izar Tarandach Look, I want to invite you guys to step back for a bit.

6:19Chris Romeo Okay.

6:20Izar Tarandach In a perfect AI world, this would make absolute sense. This would be Scotty in Star Trek IV picking up the mouse and saying, computer, do this for me. Okay. And everything would work. But ever since this LLM thing started coming out, we have been pointing at this thing writes buggy code. This thing doesn't understand me. This thing thinks it understands me. This thing is explaining myself to myself and I don't understand it. Why all of a sudden we declare that the models are so good that we feel good enough to just casually chat with it and come up with, uh, scary air quotes, but not really functional application. Aren't we getting like, well, so too much ahead of ourselves?

7:18Matt Coles Think, think about who's tell, who's, so first off, who understands what an AI hallucination is?

7:28Izar Tarandach Last week we had some advances in AI explaining itself. Okay. So there is a movement towards the chain of thought, which is supposed to be able to at some point filter out and figure out hallucinations better. But I'm going to go even to the basics. This thing is trained with some code. Have we reached peak secure code that we now have a corpus of code that we can use to train on AI and say, listen, this code is secure. You can use it as an example of things to do for other people.

8:10Matt Coles So isn't, isn't that supposed to be the, uh, the promise of, um, well, during the training step where you have a human go and actually validate You know, this code is good and this code is bad as part of the training data. Isn't that what it was supposed to do? Uh, I'm, hey, I'm being honest, you know, just, just if you have that as a, as a training step, right, where you could say this code is good and this code is bad.

8:36Chris Romeo Yeah, but they couldn't, there's too many derivatives that exist within, within code.

8:42Matt Coles That is true. That is true. How many humans?

8:44Izar Tarandach How many humans? Know that you wouldn't trust a piece of code and say, this is bad.

8:48Matt Coles Right. So that goes back to my earlier question for people who this opens the door to software development. How, and, and can afford, and, and by the way, can afford a, a license to, or subscription to use an AI that can generate code in a reasonable way. How many of them know this? How many know what hallucinations are? How many of them know and understand?

9:11Izar Tarandach That's the opposite, complete opposite of what you mentioned before of the democratization of computing.

9:16Matt Coles I wasn't saying it was a good or bad thing. I was saying—

9:19Izar Tarandach No, because these people are going to do this thing and not look at the code because there is vibe debugging and vibe deploying and vibe vibing.

9:28Chris Romeo I mean, in, in the use case Matt's describing, that audience doesn't have the ability to look at the code and determine if it's good. Or in your, your example too, a finance person. I'm not, I'm not making fun of finance people, but in this example, if you're, if you're, if you're a business analyst or something, you, you probably don't know how to code. If you knew how to code, you'd probably write the app yourself. You probably wouldn't use AI to generate it. And so the challenge that we haven't really explored here is trust, but verify. How do you trust what's coming out of this? And to make matters worse, how do you trust what's coming out of it if you don't have any way to check what's coming out of it?

10:10Izar Tarandach And now let's consider that you can't trust but verify if you don't know what it is that you verify. So 3 months down the road, first big thing happens. What they're going to look to, I didn't write the app.

10:25Matt Coles Well, I wrote the app, you know, and that brings up a whole host of problems, right? I mean, again, if you don't, if you don't know enough to debug it, you don't know enough that you need to debug it until something bad happens. And then when something bad happens, who's at fault?

10:44Chris Romeo It's not the AI, 'cause you can't sue the AI. There's no—

10:49Matt Coles but what's the, what's the, I guess what's the liability will have to, will have to, uh, have to change, right? Who owns the code? So I think that if there's already been legal precedence that the, If you, if somebody takes code from AI, it's not theirs. If somebody takes code from AI and modifies it, it's theirs. Interesting. And so here, if somebody fine codes an application and makes no modifications to it on their own, I think by legal precedent, by current legal precedent, they're not the owner of that code.

11:26Chris Romeo That would be a heck of a loophole if that was truly the case.

11:32Matt Coles I mean, they may be the owner of the sites, but are they the owner of the code?

11:37Chris Romeo Ownership in the age of AI is a very sticky issue for anything that you get. Anything you take out of an AI, do you, is it really yours? Is it intellectual property that belongs to your company if unless you generate your own LLM and run it within your own environment where it's your own models that you're, you know, you're not using a commercial model at all.

12:01Matt Coles Right. And if you're doing that, you're not inviting, you're not, you're, you're a company or an individual with a lot of skill who should know enough to write their own code or to debug or to understand the notion of having to do QA, et cetera.

12:16Chris Romeo Yeah. When did we lose the idea that software development was an art, a craft, and something that there was a lot of skill? Like in this example, we're describing a world where we're just basically saying anybody can generate an app in any capacity without taking into consideration the amount of learning and experience that goes into building a solid application. That doesn't fall over, that doesn't, isn't riddled with security problems.

12:51Matt Coles You know, we may have lost this. Oh, go ahead. Sorry.

12:54Izar Tarandach No, I think that we just reached peak in shitification of development, right? I said before and I'll say it again, it started back in Visual Basic.

13:06Chris Romeo I love Visual Basic. Come on, stop making fun of Visual Basic.

13:10Izar Tarandach When you start saying you don't have to write a line of code anymore, you can just draw some forms and it's going to write the code for you and everything. You know what? Back in the day, I tried it. I tried Visual Basic. I tried Visual C++. I was that much of a masochist. But then I went and I looked at the code that was written in there. And my goodness, I walked away from there. I run very, very fast away from that. As you know, I, I like what you get is what you see or what you see is what you get editors as much as the next guy. But I already knew back then that, oh my God, something is broken here. Or, or I want to go one step further and customize this thing in a way that I want so that my app doesn't look like the 1 billion other apps that this thing generates. That's when everything goes out of the window. If you don't understand the code that's being generated, what, what are we doing here?

14:07Matt Coles I'll go back even further, or maybe not, not even further, but to a more fundamental case. We, and we've talked about this in the past, right? It's when you, when we stopped, when, when, when the or when the environment, the, you know, the business environment, whatever, um, decided that quality was no longer the factor that resulted in production. Time to market. And getting it out there faster, getting out there first, getting it out there, uh, you know, and let the crowdsource funding, crowdsource testing ensue, right? That's, this is a continuation of that because now you don't even need to be a software engineer. You don't need to understand how to program in a language. You don't either need to understand how to use all the IDEs and tooling and have your own CI/CD pipeline or anything like that. Now you can have something else. Now you can, now you can hire a contract developer or team of developers. That's, I mean, effectively that's what you're doing is you get a subscription to an AI tool, you're getting a contract developer.

15:15Chris Romeo Yeah, but they're, they're nowhere near the ability to reason.

15:21Matt Coles The person who's, yeah, absolutely. The person who's hiring the, the contract developer can't assess the skills of the contract developer. To know enough that they're not—

15:30Chris Romeo Well, you look at, I mean, I disagree because I've hired a number of contract developers. You look at their ratings, you look at their references, you look at what other people have said about them. I've never hired a contractor on Upwork I didn't background investigate on my own to some degree. Now, how do I do that in an AI world? How do I, how do I look for people's reviews on this AI?

15:54Matt Coles You don't.

15:55Chris Romeo I don't. I have to trust it. I have to trust it at a level that I'm not comfortable trusting anything without understanding how it works.

16:02Matt Coles But you're not, you're not the high school kid or the 20-something who's trying to, you know, make a quick buck or, you know, has this great fancy idea for a trading app, wants to get in on the ground floor.

16:13Chris Romeo One of the, one of the big challenges that's just coming to my mind here, that's one of the negative impacts of this whole thing we're discussing in this vibe coding. Is one, there's just gonna be a plethora of crap apps that come out and don't, and, but then also if I want to knock somebody else, somebody's app off, imagine how easy it'll be to say, look at this website and build something that does almost the same thing, but just a little bit different. And so then all of a sudden you'll have 100 options in a category that are slightly different from each other that were all crafted by AI, it's gonna, it's gonna create a, just a toilet bowl of—

17:00Matt Coles and the smart prompter—

17:03Chris Romeo agree, thank you for agreeing with me, I appreciate it.

17:06Matt Coles The smart prompter would say, do this only better, make it look sharper, make it cleaner, because the models evolve over time.

17:13Izar Tarandach Oh, smart prompter, that, that's one that you have go back to. I want to see how long it's going to take until this turns into pejorative and people start telling the other one, oh, we were vibing hard when he wrote this one. But, uh, to the smart prompter, back when we were kids, getting something out of the computer meant basically writing your program, giving it to a guy usually in a white coat. Who would translate that into IBM C. It would go into a mainframe and a couple of days later you would have a stack of output that came from the computer and good luck if you had to debug something.

17:55Chris Romeo How old are you?

17:57Matt Coles Too old.

17:57Izar Tarandach I'm ancient. I'm ancient. Then the, uh, the, uh, the, again, that word democratization came in with micro and mini computers and all of a sudden we could stand in front of those green and black terminals all the day.

18:13Matt Coles Oh, I love those.

18:15Izar Tarandach Machine.

18:16Matt Coles Nothing like a VT100 terminal with the little fake key press sounds coming out of your keyboard.

18:22Izar Tarandach VT22, uh, 220 with the orange and black. That was an epic. But anyway, so now all of a sudden we are going back to the IBM days. but now with something in the middle that on one side understands us and on the other side understands code and it's a bridge between those. So just to go to the, uh, contractor analogy, do you see yourself like standing in front of a white field, bringing in a contractor, standing there looking at the field and saying, build me a house. It has to have 3 doors, 6 windows, 2 floors, and a garage. And go away, you come back a couple of months later, and you have 3 doors on one side, 4 windows on the top, the garage is smack in the middle.

19:17Matt Coles There's an XKCD or a, uh, um, frame for this.

19:22Izar Tarandach Of course there is.

19:23Matt Coles What a, what a product, what a customer wants with the product owner, product manager. Interprets it as what the, you know, the middle, put another LLM, right?

19:33Izar Tarandach That's going to understand whatever it's going to understand. And now it's not a matter of refining code. It's a matter of gaining with engine that under— that tries to understand your problem and has its own. It doesn't, but acts as if it has its own mental image of what you what you try to achieve.

19:55Matt Coles And for those who don't speak Hazar yet, bargaining.

19:59Izar Tarandach Yeah, thank you. And imagine like trying to do a smart contract that explains yourself and now you're not achieving the result that you wanted.

20:14Matt Coles Oh, it gets even worse.

20:14Izar Tarandach And you can't even debug that thing.

20:16Matt Coles It gets even worse.

20:16Izar Tarandach It's only my pain.

20:18Matt Coles So to Chris's earlier point, right, about, you know, this plethora of, you know, oh, I want to up one up. I wanna one-up the, the app that's out there, or I wanna create a flood of these things. Well, imagine the savvy individual or organization, I think probably organization more so, that poisons the training set, tricking it maybe into including, oh, every piece of everything that you generate should include this snippet of code, aka a backdoor, or, um, you know, have, uh, poison pills, you logic bomb kind of thing. Like, oh, you know, you gotta have a failsafe just in case. So have a trigger on, you know, packet X being sent over the wire. I mean, there's a host of bad things that can happen here.

21:06Izar Tarandach I think that my biggest beef with the whole thing, that when you come to the what could go wrong phase, very quick we are going to get to It either works or it doesn't work because that's all the kinds of what could go wrong that I can see here. And when you get to the stage of what can we do about it, it's basically nothing. I have to ask the model. I'm not reading all that.

21:36Chris Romeo Just FYI, I asked the model to create your—

21:40Matt Coles oh my God—

21:41Chris Romeo your example. And so anybody who's afraid that AI is gonna take their job, imagine handing that to the architect. And just for those people listening to the audio version of this, uh, uh, ChatGPT created me an ASCII art style picture of a two-story house with three doors. It literally says upper floor, lower floor with window, two front side doors, third door back or side, which looks like it's underneath the house.

22:10Matt Coles I don't even know what style of house that is.

22:14Chris Romeo I just, I just thought it would be like an Art Deco house. Relevant. No.

22:19Izar Tarandach And that's why we can't have my—

22:22Matt Coles they weren't even the same.

22:26Izar Tarandach They weren't even the same size floors there.

22:28Chris Romeo Different heights and stuff. It's all good. But we can, we can trust this thing to build production applications for us. It's going to work with the same amount of, the same amount of, of belief that you have in being able to build a house based on that blueprint that I just showed in ASCII art should be how much confidence you have in building an application using Vibe coding.

22:56Izar Tarandach It's even worse. I'm going to sit now and I'm going to watch 10,000 hours of DIY, uh, uh, channel. Of YouTube, How to Build a House, of the Amish channel, How to Build a Barn.

23:07Matt Coles Yep, those 30-minute elapsed time videos of like 9 months of building into 30 minutes.

23:12Izar Tarandach And now anybody who explains to me what they want, hey, I watched all that stuff. I'm able to do it.

23:19Matt Coles You slept in a Holiday Inn last night, right?

23:23Chris Romeo It was a Holiday Express, just FYI.

23:27Matt Coles So, so, you know, it's interesting, a couple of months ago, there was a, there was an article and about AI and its capability level. So thinking about AI as a contract developer, uh, and, and with AI agents, you could think of it like a contract development company. They're still operating today at primarily a junior engineer level.

23:52Izar Tarandach That's the point. That's exactly the point.

23:54Chris Romeo Yeah.

23:54Matt Coles And, and people have forgotten this, right? That, that AI is not a fully capable functioning system yet, right? It's, it's early days, but it's rapid advancing, but still overall, and you know, in part because of the code that you get, that it gets trained on, is operating at junior level and it can only reproduce code based on description. It can't think up new things because it doesn't know new things. And so, so Yeah, that is right.

24:24Izar Tarandach So we, we, we ranted about, about it for a while. Now let's get on the positive side of the coin. What do we tell people about, about vibe coding?

24:33Chris Romeo Oh, sorry. I thought he was joking that we don't really have a positive side. Sorry. Go ahead.

24:37Matt Coles Go ahead.

24:38Chris Romeo What do you mean?

24:38Matt Coles Don't rely on, don't rely on vibe coding for production ready. Use it for idea generation. That's what it's really good at.

24:46Izar Tarandach For the love of God, do not fire your developers.

24:49Chris Romeo Does it work for prototyping? Like, would it be useful for an existing developer that's trying to prototype something fast and kind of prove a concept? Like, it's not code they're going to commit to product to, uh, the source code repository, but like they're— sometimes when you're built, taking a user story, you don't just start building the code for the story. You kind of mock something up so that other people can give you feedback. Is that a use case?

25:13Matt Coles Yeah, I think you can. I think if you're going to brainstorm, it absolutely works for that, right? It may come up with 10 different ways of doing something. You go, oh, none of those are good, but at least now I have an idea of what's possible or what's, what might be something, right? To start with.

25:28Chris Romeo Yeah.

25:29Izar Tarandach But you need to have a hard policy in place and enforced and policy that mockup code does not make it in. Okay. You don't emerge things out of that. Take ideas, but you don't just get the thing and start riffing on it. And again, the basic idea is the AI is probably going to be using a lot of borrowed code from a lot of borrowed places. And without, I won't even say deep understanding, without an understanding of what's in each one of those black boxes, you may have something, especially from a security point of view, that's going to bite you down the road.

26:13Matt Coles And, and also think about this, just an another follow-on effect. We've been talking about secure by design and, you know, uh, memory safety is a big, I'm a, I'm a, not a, I won't say a huge fan. The idea of memory safety is correct, but the implementations are not always perfect, but the body of code that AI is mostly trained on is C, C++, Java, Python, maybe some Go in there somewhere. Um, maybe Visual Basic code, right? So I guess when you're vibe coding, do you say, hey, write this application in C, or do you not care and you get whatever language it produces? Which by the way, isn't guaranteed to be the same language consistently. That's a separate problem.

27:06Izar Tarandach Um, on that point, I'm going to go back to the RAS mode. I am not going to trust a thing that cannot count how many Rs are in the word strawberry to count the size of my buffers. Okay. So take it with a ton of salt. Right.

27:24Matt Coles But it means, it means all of those, it means all of those things that we're trying to put in place from us, secure by design, secure by default. Standpoint becomes challenging unless you, unless the, either the AI agent has already been programmed that way or you know enough to ask, right? Give me a, give me an application or, you know, based on this description, give me an application that is free of memory defects, that is free of an entire common weakness enumeration, which by the way, if you could figure out how to do that, that might actually be a function.

27:57Izar Tarandach It'd be awesome. While, while we are in the, the domain of figure out how to do that. So in the beginning, in the early days of 2023, you were prompting an LLM and each prompt would have like all of the previous prompts and results so that the thing could like have context. And very, very quickly would get to that window of maximum context and all have broken. Now we have LLMs with memory so that you can over the, the dialogue, the thing knows where it's been and where it's going and bad things happen. So I would like to see the first of our many loving, lovable security tools that is actually able to speak with the same LLM context, generate the code and say, and talk to the LLM in a language of prompts that it actually understands so that it fixes the code that it just wrote. So I think that the first tools that actually manage to close that loop, then I'm going to start sleeping a bit better about white coding. From a security point of view, from a quality point of view, I'm going to be really— no. We have tools out there that can help. Let's say they got money. So you are moving away from cheap side of the fast, cheap, good. Live coding is not the thing that's going to bring us fast, cheap, and good. You still will have to do some investment in there. And if you want to do responsible about it, that investment is going to be somewhat large because you're going to have not only one tool, you're going to have a chain of tools. And those tools will have to talk to your LLM and they'll have to understand the results that are coming out and they'll have to talk to it in a, in a language that it understands. So to go back to my IBM thing, don't think that you are now, that, that you now have a perfect layer of understanding between you the computer and you're going to get some code that does perfectly what you want. You'll have to, not all batteries are included, is what I'm trying to say.

30:23Matt Coles Guardrails and paved roads, do those go out the window?

30:26Izar Tarandach Yeah, no, but to your point, you were talking, no, Chris was talking about the thing, or at least, no, you were talking about the organization now including stuff into machines. I can already see next year's conferences telling us all about how they got paved roads and guardrails into the training of the machine. So the machine basically cannot do wrong because it goes against what's been trained, right? And I'm going to invite anybody thinking about that to go and read Asimov's Anything that goes to the three laws of robotics. I mean, are we moving away from being developers? Well, I've never been a developer. I was a guy who wrote code. But are we moving into philosophy here now? Are we finally finding how prompt injection is a security thing?

31:28Chris Romeo We're definitely moving in. We are definitely moving into philosophy just cuz it's fun to discuss and debate.

31:34Izar Tarandach Definitely. And it's basically what we do, what we should do when we retire, right?

31:39Chris Romeo Do you know when Asimov wrote the Three Laws of Robotics?

31:43Izar Tarandach Mm-hmm.

31:44Chris Romeo What year he wrote the Three Laws of Robotics?

31:47Izar Tarandach 50-something?

31:48Chris Romeo '53.

31:49Izar Tarandach '53.

31:50Chris Romeo 1953.

31:52Matt Coles So, so he had, he had the Enigma machine. And a handful of supercomputers and some early mainframe.

32:03Izar Tarandach I think that back then he had the, the ENIAC.

32:07Chris Romeo Yeah.

32:07Izar Tarandach Probably a couple of IBM typewriters or something. Mm-hmm.

32:11Matt Coles But, uh, but he, but robots were not really a thing. I mean, they were sci-fi at that time.

32:17Chris Romeo I mean, yeah, it was, yeah, he, he was, I, I don't know that he created the idea, but he, Was the person who pushed it forward enough in science fiction to make it part of the, yeah, he was before, before Clarke and, um, see, like I think Arthur C. Clarke was in the '60s, wasn't he?

32:37Matt Coles Yeah.

32:39Izar Tarandach No, but I, I think we, we, we saw robots the way that today we, we consider humanoid robots back in, uh, Metropolis was the name of the movie, 1920-something.

32:53Matt Coles Yep.

32:54Izar Tarandach But Asimov was the one that came up with the idea of the electronic brain, positronic brain, and that it was so developed that you actually had to put some firmware in there saying, under no conditions are you going to harm a human ever, ever. Which completely kills the idea of Star Wars Clone Wars.

33:19Chris Romeo But, uh, and Terminator.

33:20Matt Coles Well, maybe if they had, if they had thought about what could go wrong, they would have added the, the failsafe code. But, uh, that's the thing. He was, Asimov, enough to write a number of, uh, uh, books and tales about robots who had ethical conundrums that would move them to actually break the three laws and harm a human because it could save somebody else, or it's really, this is the, this is the classic problem of how do you do a, how do you do a self-driving vehicle if your choice is kill a kid versus kill a crowd or kill a grandmother versus a, versus a kid or one object versus another, right? How do you make a decision if your choices are you must? You must choose one of them because you cannot choose none of them.

34:13Izar Tarandach And introduce, uh, games.

34:19Matt Coles Well, the only way to win is to not play the game.

34:21Chris Romeo Not to play. Yeah.

34:23Izar Tarandach And that's why you'll never see me in the taxi that self-drives or in a car that self-drives.

34:29Matt Coles You'll never see you in a taxi in San Francisco.

34:32Izar Tarandach Nope. We were there last year and I was like, No, not only I'm not writing one of those, I'm letting it pass before I even think about crossing the street. True. Okay. Why? Because I know the people who write these things and they were not—

34:50Matt Coles I got one more before you go though. So, uh, if you don't watch Mark Rober on YouTube, you should, because his last episode was about testing auto brake behavior in cars, which included, uh, whether they use visual, visual detection versus LiDAR. And he did a Wile E. Coyote, uh, uh, version of it. So one of the tests was he created a fake wall and, uh, and tested whether or not auto braking vehicles would make the right decision. You should go watch it. It is hilarious.

35:34Izar Tarandach Dear NLM, please write an automotive operating system.

35:43Chris Romeo All right. Well, with that, I think we are out of time for this episode of Security Table. So go out there and vibe, code something.

35:53Izar Tarandach Yeah. Only good vibes.

35:54Chris Romeo Only good. I don't even know what that means either.

35:56Izar Tarandach Yeah.

35:57Matt Coles Where's the, where's the psychedelic lava lamp?

36:01Chris Romeo What?

36:01Izar Tarandach You're not seeing it? It's only me. Oh God.

36:05Matt Coles It is only you.

36:06Izar Tarandach Oh God.

36:06Chris Romeo I gotta call, call one of my kids to have 'em explain this whole vibe thing to me.

36:10Matt Coles So I'll try to get, have them vibe in him. Have a vibe them, vibe that conversation for you.

36:16Chris Romeo I, I'm so confused right now. It's, it's like Buddies, thanks folks!

Transcript supplied by AssemblyAI. It may contain transcription errors.

Originally published as The Security Table. Part of the AI Security Table archive.