# The Human In The Loop Illusion: Why AI Approvals Are Failing Security

URL: https://securitytable.ai/the-human-in-the-loop-illusion-why-ai-approvals-are-failing-security/
Published: 2026-04-30
Duration: 2866 seconds
Season: 4
Episode: 9
Hosts: Chris Romeo, Izar Tarandach, Matt Coles

## Show notes and chapters

**In this episode, a debate about hacker movies turns into a deeper conversation about AI, security, and the human-in-the-loop illusion. We explore how approval fatigue and AI-generated code can create a false sense of security and why fundamentals still matter.**

**🚀 Join the Conversation  
 Are we improving security, or just automating bad decisions faster?**

  

Follow AI Security Table:

➜ Home: [https://securitytable.ai/](https://securitytable.ai/)

➜ X: [https://x.com/AISecurityTable](https://x.com/AISecurityTable)

➜ LinkedIn: [https://www.linkedin.com/company/ai-security-table/](https://www.linkedin.com/company/ai-security-table/)

➜ YouTube: [https://www.youtube.com/@AISecurityTable](https://www.youtube.com/@AISecurityTable)

## Transcript

**Chris Romeo (0:09)**: Hey folks, welcome to another episode of the Security Table. This is Chris Romeo joined by Matt Coles and Isar Terrandas. And we were just discussing, debating what to talk about, and Isar made just a very untrue statement.

**Izar Tarandach (0:26)**: So, hey, this is another opinion.

**Matt Coles (0:28)**: This is an episode of Rock 'Em Sock 'Em Robots. Yeah.

**Chris Romeo (0:31)**: Well, when it goes— when it comes to hacker movies, there is no opinion. It's just truth. So, Isar, you claim that Sneakers is not a hacker movie?

**Matt Coles (0:42)**: Wait, say that louder. I don't think our listeners heard that.

**Izar Tarandach (0:45)**: Sneakers, hacker movie, not.

**Chris Romeo (0:48)**: Okay, you can reach Isar at isar_tarendas at gmail.com for all of your, you know, emails you want to send him about how wrong That is. So what, okay, so what makes a hacker movie?

**Izar Tarandach (1:04)**: Well, I can't tell you what makes, makes Sneakers not a hacker movie.

**Chris Romeo (1:07)**: I know, I want to know why. No, no, no, no, no, no, no. That's not how requirements work.

**Matt Coles (1:12)**: We have to start with, are we, are we, are we, are we doing the UFO interviews? Is that what we're doing? We're playing Spectre and Velma to hackers.

**Izar Tarandach (1:19)**: We'll find a way to get to the serious part later. What makes a hacker movie? How many black screens with, uh, uh, uh, green text do you have in sneakers?

**Chris Romeo (1:34)**: That's, that's not a requirement.

**Matt Coles (1:36)**: How many Unix prompts do they have? Excuse me. Excuse me.

**Chris Romeo (1:38)**: I need a requirement. So all that, so that means Matrix, The Matrix is the only movie that applies to your criteria.

**Izar Tarandach (1:48)**: No, Hackers also has it.

**Matt Coles (1:49)**: Hackers. Jurassic Park. Swordfish.

**Chris Romeo (1:53)**: So Jurassic Park is a hacker movie now, according to this requirement set.

**Izar Tarandach (2:01)**: Well, it does have an SGI box and they do declare that it runs Unix and that they know it.

**Matt Coles (2:09)**: So, and, and, and because they know Unix, they automatically know the password.

**Izar Tarandach (2:16)**: And so, plus it has very simple, it has a very, a very, uh, uh, big monster slobbering with non-functional typing short arms. So management is represented as well.

**Matt Coles (2:36)**: So, uh, so it is a hacker, hacker movie too.

**Chris Romeo (2:39)**: Yeah. That was bizarre. Under _terrendasch@gmail.com.

**Izar Tarandach (2:43)**: If a velo, velo, there is There's no way in my life that I'm going to say that word right. Velociraptor.

**Matt Coles (2:53)**: Velociraptor.

**Chris Romeo (2:53)**: Velociraptor.

**Izar Tarandach (2:54)**: Yeah. If they do, you hacked.

**Matt Coles (2:57)**: Yeah. What's your first rebuttal?

**Chris Romeo (3:01)**: I guess we'll just have to agree to disagree first of all on whether Sneaker—

**Matt Coles (3:05)**: I mean, Sneaker is absolutely a hacker movie.

**Chris Romeo (3:08)**: I think so. I mean, 'cause they're, they're doing physical security, they're breaking stuff.

**Izar Tarandach (3:15)**: I watch it only for Madam President, but my last line is, if your basic premise is that there is some magic black box that decrypts everything, no, you're not accurate.

**Matt Coles (3:28)**: You know, you know, you know about the Clipper chip, right?

**Izar Tarandach (3:31)**: The Clipper chip. I, I, not only I know, I know how it didn't work and I know how specific it was. So yeah, I know. And, uh, my premise stands if you start with a deus ex machina that, uh, Solves everything.

**Matt Coles (3:46)**: Okay. So the Clipper, the, no, sorry. The, the hack. Nah, now you got me into it. Sneakers.

**Chris Romeo (3:51)**: He's flustered.

**Matt Coles (3:52)**: Sneakers had, maybe what Sneakers had unbeknownst, you know, unbeknownst to the audience was a quantum computer.

**Izar Tarandach (4:04)**: Even then.

**Matt Coles (4:05)**: And it could decode, it could decode all of that, all that stored ciphertext that the NSA has been ciphering, siphoning off the internet for the past—

**Izar Tarandach (4:15)**: Claims of decoding everything from what I, from the little I understand of quantum computing are also very exaggerated. So, ah, you know, really? Let's just say that. Yeah. Well, I mean, let's just say that we'll survive.

**Chris Romeo (4:27)**: If you have the ability to potentially decrypt anything that wasn't a quantum safe algorithm, could you decrypt the entirety of the internet? I think that's Isar's point here is like, you couldn't just say, well, I'm making the internet unencrypted now, all traffic. I have enough capacity to generate unencrypted versions of all traffic that crosses the internet. Like, true.

**Matt Coles (4:51)**: But my point is cracking keys, cracking keys, session keys is, is no longer theoretical, right? That, that can happen.

**Izar Tarandach (4:58)**: But again, my limited understanding of quantum computing, aren't we talking here only about keys that depend on primes and that kind of stuff. We are not talking about any single key.

**Matt Coles (5:10)**: So AES keys, we believe, are safe from quantum computing, but the RSA or Diffie-Hellman exchanges or the EC, ECC keys, EC keys, elliptic curve keys that are used to transfer symmetric keys are vulnerable.

**Izar Tarandach (5:27)**: Yeah. Which, which power a lot of the crypto. I'm not going to, to minimize it at all.

**Matt Coles (5:32)**: Powerful open crypto. AKA, it's how TLS works.

**Izar Tarandach (5:36)**: But no, but to go from here to, hey—

**Matt Coles (5:38)**: If you have the AES key, you can do the decryption because the AES key is a single symmetric key, right?

**Izar Tarandach (5:43)**: To go from here to, hey, there's an encrypted blob, go decrypt it, go quantum, go quantum.

**Matt Coles (5:50)**: No.

**Izar Tarandach (5:51)**: No, sure.

**Matt Coles (5:53)**: Sure, you're right. You need the key exchange in order to do the decryption.

**Izar Tarandach (5:57)**: You need a bunch of things to align correctly before you get the—

**Chris Romeo (6:01)**: And maybe a little bit of imagination would help. So wait, if we put AI and quantum together and we mix in a little bit of blockchain, you're gonna get random garbage, I'm sure, but just random garbage.

**Izar Tarandach (6:20)**: Plus, plus my last point, my last point, when you have to use the name of a, a chocolate bar to name your movie. It's not a hacker movie.

**Matt Coles (6:28)**: Sneakers, not Snickers. It's called Sneakers cuz they run places.

**Izar Tarandach (6:35)**: Hey, coming from me, it sounds all the same. True.

**Matt Coles (6:39)**: It's where, if it, if it wasn't a hacker movie, then where would Sneakernet have come from? I mean, come on.

**Chris Romeo (6:44)**: Look at the roster of this movie.

**Matt Coles (6:49)**: You got Robert Redford, not a single programmer among it.

**Chris Romeo (6:53)**: Dan, Dan Aykroyd's always been a nerd.

**Izar Tarandach (6:55)**: Come on.

**Matt Coles (6:55)**: He was a Ghostbuster. Yeah, but telling me he was also in Spice Like Us, if I remember correctly.

**Chris Romeo (7:03)**: Yeah, he was. Good, good, good '80s movie pull. All right, well, we'll just have to agree to disagree that there are no hacker movies, apparently, according to Isar and his criteria.

**Izar Tarandach (7:16)**: There, there, there are, there are, just Sneakers is not one of them.

**Chris Romeo (7:19)**: But you, but you listed Hackers where their, their claim to fame is they wear helmets and rollerblades.

**Izar Tarandach (7:26)**: Hack all the things, man. Hack all the things.

**Matt Coles (7:28)**: Hack the planet.

**Izar Tarandach (7:29)**: Hack the planet. Hack the planet.

**Chris Romeo (7:32)**: Gotta get the tagline right.

**Izar Tarandach (7:33)**: Plus Angelina Jolie. That makes it a hacker movie.

**Matt Coles (7:40)**: All right. Swordfish. Was Swordfish a hacker movie?

**Izar Tarandach (7:44)**: Ooh, that's a That's a hard one.

**Chris Romeo (7:53)**: It's because you can't remember it.

**Matt Coles (7:56)**: Because it was—

**Izar Tarandach (7:58)**: no, because I'm trying to remember it, but all I can see is John Travolta in Pulp Fiction.

**Chris Romeo (8:03)**: Yeah, that's true. Now, Pulp Fiction, that's a hacker movie.

**Matt Coles (8:08)**: That's a, that's a, that's a hack, uh, slash.

**Chris Romeo (8:11)**: I mean, it ran in reverse. It just seems like that's a post-hacking move.

**Matt Coles (8:19)**: But, uh, that's when hacking goes kinetic.

**Izar Tarandach (8:23)**: Oh, Johnny Mnemonic.

**Chris Romeo (8:27)**: Keanu Reeves.

**Matt Coles (8:27)**: Mm-hmm.

**Izar Tarandach (8:29)**: That's a good one. A Scanner Darkly. That's a good one.

**Matt Coles (8:33)**: What about a lawnmower man?

**Izar Tarandach (8:39)**: It's right there with the one of the guy with the pointy things. What's the one?

**Chris Romeo (8:45)**: Uh, wasn't that like a VR?

**Matt Coles (8:47)**: It was long ago. It was a VR movie.

**Izar Tarandach (8:48)**: Yeah.

**Chris Romeo (8:48)**: It wasn't a hacker movie. VR was getting you or something.

**Matt Coles (8:51)**: Yeah. It wasn't a hacker movie. It was a VR movie.

**Izar Tarandach (8:54)**: But yeah. The 13th Floor.

**Matt Coles (9:00)**: 13th Floor. No, not the 13th Warrior. That's a different movie.

**Izar Tarandach (9:02)**: No, no, no. That's a totally different thing. Totally different thing.

**Chris Romeo (9:05)**: They haven't yet made an AI hacking movie, but there's gotta be one coming.

**Matt Coles (9:09)**: Wait, a movie about AI hackers or a movie about hacking AI?

**Chris Romeo (9:14)**: I don't know. One of those combinations.

**Izar Tarandach (9:17)**: Well, there's, there's, there's Matrix. Yeah, of course.

**Matt Coles (9:20)**: Which was AI.

**Chris Romeo (9:21)**: Yeah. But that's before AI took on the shape that it has now. That was kind of, that was a lot more science fiction. Wait, so that's more than a real thing now. There's a real thing to back it up. You're gonna go watch AI movie with Chris Pratt where he's the police officer. And he's in Los Angeles.

**Matt Coles (9:43)**: Is that Timecop?

**Chris Romeo (9:45)**: No, no, no. No, it's actually pretty good because like the criminal justice system of the future is completely AI driven and he's a cop and he wakes up and finds out he's committed a crime and he's got 90 minutes to use the AI to try to prove his innocence.

**Matt Coles (10:00)**: Is that a, is that Minority Report?

**Chris Romeo (10:03)**: No, it's a new movie. It just came out.

**Matt Coles (10:05)**: Oh, okay.

**Chris Romeo (10:05)**: 2026.

**Matt Coles (10:06)**: Yeah, it was, but yeah, it was, but it's, oh, well then actually what about Blade Runner? Was Blade Runner a hacker movie? Wow.

**Chris Romeo (10:13)**: Ooh. Ooh.

**Izar Tarandach (10:15)**: That's a, ooh.

**Chris Romeo (10:21)**: I think we're with Blackfish Corps. It's a revolution movie. Like they're trying to, like the robots are trying to rise. That doesn't make it an AI or a hacker movie though. Mm-hmm. Because they weren't really trying to, they weren't really compromising anything. They were just trying to revolt and and stand up for themselves.

**Izar Tarandach (10:39)**: Independence Day. Is that a hacker movie? Yeah. Oh yeah.

**Chris Romeo (10:47)**: There was a sequence in Independence Day.

**Matt Coles (10:49)**: They used hacking techniques, but they, they're not exactly a hacker movie. That wasn't the core of the issue. The core of the issue was human survival.

**Izar Tarandach (10:56)**: Wait, wait, wait. How, how, how did they win the war? Yeah.

**Chris Romeo (11:00)**: They sent the virus into the, they sent the virus, the mothership.

**Izar Tarandach (11:03)**: They not only sent the virus, it was live coding the virus in the mothership. And I have long ago reached the conclusion that the only way that that storyline was plausible was if the aliens were running Linux.

**Chris Romeo (11:27)**: I have a feeling like aliens—

**Matt Coles (11:28)**: No, the aliens would have to have been running mainframe. So that all the little ships were terminals.

**Izar Tarandach (11:34)**: No, you're looking—

**Chris Romeo (11:35)**: No, I think it was BSD.

**Izar Tarandach (11:37)**: No, you're looking for something that has enough reach, that is compatible humans and aliens, basically can run on anything.

**Matt Coles (11:48)**: Yeah, but why would it run?

**Izar Tarandach (11:49)**: I mean, it could be Doom. It could be Doom.

**Matt Coles (11:51)**: Why wouldn't the little thing, why wouldn't the little ships be autonomous from the big ship? Why would it? And by the way, they use the same approach in the Avengers. The first Avengers movie, remember they got the mothership and all the things died? What's up with this command and control system?

**Chris Romeo (12:09)**: We've really taken a turn because now we're designing alien spaceship systems and searching for vulnerabilities.

**Izar Tarandach (12:19)**: He's got a good point here because some people, it's like, hub and spoke like that would like immediately jam up. Unless to, to go back to you using one of the fancy quantum communication things where everything is entangled and, and you can't jam that thing.

**Matt Coles (12:37)**: But, uh, hey, it wasn't, it wasn't just that they were like not getting their communications. As soon as you killed the mothership, everything died.

**Chris Romeo (12:47)**: That seems like a really bad design, doesn't it? Like, that's like token ring networking. Where there's a break in the ring and everything goes off the network, I guess.

**Izar Tarandach (12:57)**: Yeah, and the bytes spew all over the floor. It's horrible. You gotta get two boxes, put the zeros in one, the ones in another. Horrible.

**Chris Romeo (13:05)**: But that doesn't seem very resilient to have this kind of failure.

**Izar Tarandach (13:09)**: Because somebody's always getting the token ring and trying to run to Mordor.

**Matt Coles (13:14)**: You've all used token ring, you know. If you know, you know.

**Izar Tarandach (13:17)**: Oh yeah, if you know, you know.

**Chris Romeo (13:18)**: If you know, you know, but most people have no idea what we're talking about at this point.

**Matt Coles (13:23)**: FDDI anyone?

**Izar Tarandach (13:25)**: Novell. Novell with vampire connectors.

**Chris Romeo (13:34)**: Come on, Novell was some of the most performant networking software.

**Matt Coles (13:39)**: Because it was the only networking software for a while.

**Izar Tarandach (13:42)**: Yeah.

**Chris Romeo (13:43)**: You had OS/2.

**Matt Coles (13:49)**: Yes, yes you did, but why?

**Chris Romeo (13:50)**: I remember trying to run OS/2 once, thinking like, I'm going to be the cool kid, the only person I know that runs OS/2. That experiment lasted about 18 hours before my kids saw us.

**Izar Tarandach (14:04)**: The name of the thing was OS, but then they found the other person who also runs it and said, do you run OS/2? And then it stopped.

**Matt Coles (14:12)**: No, hold on. But wasn't it spelled OS/2? So really what they found was half an operating system?

**Chris Romeo (14:23)**: Ooh. That's a great one. Wow, this is like stand-up night on this security table.

**Izar Tarandach (14:30)**: The thing is that it never got upgraded because they didn't want to lose all the printed material for OS/2 to go to OS/3.

**Matt Coles (14:39)**: I mean, running those line printers was expensive, right?

**Chris Romeo (14:49)**: Yeah. Come on, you can still hear the sound of it.

**Matt Coles (14:51)**: I can and feel the vibration.

**Chris Romeo (14:54)**: I can, I can still hear it. Once again, for those people born after the year 2000, you have no idea what we're talking about.

**Matt Coles (15:05)**: All right, so here's the question. Can we put together, can we have AI use, uh, between like Nano Banana and VO to create, recreate the '70s and '80s computing technology.

**Izar Tarandach (15:20)**: But no, wait, it's even worse. I don't know if he was doing the sound of a printer or a modem.

**Matt Coles (15:26)**: It could be both.

**Izar Tarandach (15:27)**: It could be both.

**Matt Coles (15:28)**: It could be both.

**Izar Tarandach (15:31)**: But yeah, we could use to—

**Chris Romeo (15:32)**: Oh my. Yeah. 9600 8N1.

**Matt Coles (15:40)**: Yeah. When you had to type in those parameters exactly, or else, or else, you know.

**Izar Tarandach (15:48)**: So what was it that we wanted to actually talk about today?

**Matt Coles (15:52)**: You started this mess.

**Chris Romeo (15:54)**: Oh, actually I didn't.

**Matt Coles (15:56)**: Yes, you did. You said Sneakers was not a hacker movie.

**Chris Romeo (15:58)**: That's true.

**Matt Coles (15:58)**: And that caused mayhem.

**Izar Tarandach (15:59)**: And I stand for and by it.

**Chris Romeo (16:03)**: Well, uh, folks, I just want to announce that we're interviewing for a new host on the Security Table.

**Matt Coles (16:08)**: We've got an opening. Must think Snacky Sneakers is a hacker movie.

**Izar Tarandach (16:12)**: Yeah, apparently we have two candidates, right?

**Chris Romeo (16:17)**: Uh, yes, must, uh, appreciate solid hacker movies and, uh, agree with Chris. That's your, that's your criteria. Now there's a—

**Izar Tarandach (16:28)**: we have—

**Chris Romeo (16:29)**: we've got one article we can just, just talk about for a minute or two. Uh, Chris Hughes wrote this article, The Human in the Loop Illusion. Something that we've talked about this a couple of times. Matt, with some hand signals for those watching on YouTube. Is that jazz hands that you just did or what was that? No, I think it was the raining.

**Izar Tarandach (16:51)**: It was the raining.

**Chris Romeo (16:53)**: That was rain. That was chocolate rain. No, that was a— it was a rain or something. All right. So if I look at the executive summary, this Chris states human in the loop is not a real security control. Anthropic's own data shows users approved 93% of AI agent permission prompts, with experienced users auto-approving 40+% of sessions. Approvals are rubber stamps, not meaningful reviews.

**Matt Coles (17:21)**: This is the whole human behavior of, uh, what is it, um, prompt overload or, uh, prompt fatigue.

**Chris Romeo (17:29)**: Prompt fatigue, yeah, which is a UX construct, right?

**Matt Coles (17:32)**: It is. It goes to that psychological acceptability portion of, uh, of secure design. Make it easy and informative as opposed to what becomes a rote activity. I click, I click there, I must, therefore I click.

**Izar Tarandach (17:53)**: That's great. I actually created the new, uh, job description, right? The chief officer of pressing enter when the agent asks me to. Like you can stand the whole startup just doing that.

**Matt Coles (18:13)**: In fact, that would be a robotics company.

**Izar Tarandach (18:16)**: Yeah. No, no, you just get, what's the name of that bird? That wooden bird?

**Chris Romeo (18:21)**: I know what you're talking about.

**Matt Coles (18:24)**: Or what is it? That rain thing where it goes, dum dum, and like drains water, you know?

**Izar Tarandach (18:31)**: Yeah, yeah.

**Matt Coles (18:32)**: The Japanese one.

**Izar Tarandach (18:33)**: Or you just grab like a toothpick, break it in half, press the key, stick the toothpick in there and just hold Enter down.

**Matt Coles (18:41)**: Well, if you're going to do that, you might as well find the, you know, get the pin that does it and just, they have a hacking cable for that, I see.

**Izar Tarandach (18:48)**: No, I know. Let's write a daemon. Let's write a daemon that clicks Enter. If nobody remembers how to write the daemon, you can just bytecode it.

**Matt Coles (18:56)**: What is it? Uh, 10 lines of assembly code or something?

**Izar Tarandach (19:00)**: No, no, just don't forget to press enter.

**Chris Romeo (19:02)**: But 10 lines of Python. Come on.

**Matt Coles (19:05)**: Actually, it's probably even less than PowerShell these days. Yeah. Just have to intercept the signal. Right.

**Chris Romeo (19:10)**: So the other things that, that, uh, Chris is talking about here. So Anthropic created some auto mode. Which replaces per-action human approvals with an AI-based two-layer classifier that scans inputs for prompt injection and evaluates each action before execution. What he's saying is that Auto Mode has real gaps, 17% miss rate on genuinely dangerous actions, and it won't catch supply chain attacks, which makes sense.

**Matt Coles (19:41)**: Yeah, 'cause it's, I mean, so model-based classifier, right? That's trying to make a decision, right? It's basically saying, is it in Or is it out? And supply chain attack, of course, makes it look like it's in because it, you can't tell it's out.

**Izar Tarandach (19:57)**: The, the decision point happens somewhere else.

**Matt Coles (20:00)**: It's not necessarily that, I mean, supply chain attacks, things like typosquatting, it looks the same. Or, uh, it has the same name, a similar name, or it's a newer version of the same thing, but it's actually been hacked. Right.

**Izar Tarandach (20:14)**: Right.

**Matt Coles (20:15)**: So, so depending on how it's doing the detection, unless it's doing deep analysis and I, by deep analysis, like not just analyzing the package, but also analyzing what happened upstream. Right. Then how does it know?

**Izar Tarandach (20:31)**: But if even that case, then it would be going upstream. We had this discussion before, where do you stop threat modeling? Right. You go upstream to where? To the compiler, but the, the, or the project.

**Matt Coles (20:45)**: I mean, you look at the project history and see, you know, was there any strange, was there any strange commits, right?

**Izar Tarandach (20:53)**: But it's important to, like, Chris pointed out that this is a user experience pattern. And it's important to point out that even in security, this was already an existing problem with people having PR fatigue. For a long time. So I would go and question, okay, so you tell me that you are doing code reviews. Cool. How do your code reviews happen? Oh, we look at every PR. Okay. Who looks at every PR? Oh, we have the one senior engineer that looks at every PR. Every PR?

**Matt Coles (21:26)**: One engineer? Does he work 24/7, 365?

**Izar Tarandach (21:30)**: Not only that, but You know that there's a lot of places that don't actually have any kind of guidelines and standards for PRs, and you end up with this one PR with 30 files in it. Every file is 300 lines changed. You want to tell me that one person can hold all that freaking word context in their head?

**Chris Romeo (21:53)**: We used to have a name for that in the old days. It was called a release.

**Izar Tarandach (21:58)**: Only if it happens on Fridays.

**Chris Romeo (22:01)**: Yes.

**Matt Coles (22:02)**: Yes. 4 o'clock on Fridays.

**Izar Tarandach (22:03)**: Yeah, but, uh, no, the, the rubber, rubber stamping is, is real. And now the rubber stamping is becoming AI rubber stamping. So the AI is checking the AI with an AI classifier. That's an AI thing that happens on an AI other thing.

**Chris Romeo (22:20)**: I mean, as long as you use a different AI from the one that created the code, you should be fine, right?

**Matt Coles (22:25)**: Well, actually, you shouldn't even be using AI. You should be using machine learning.

**Chris Romeo (22:31)**: Wait, what do you mean?

**Izar Tarandach (22:32)**: Which you are if it's a classifier.

**Matt Coles (22:34)**: Right. AI and ML are different slightly, right?

**Izar Tarandach (22:39)**: One comes from the other, but they are not the same.

**Matt Coles (22:41)**: They're not the same.

**Izar Tarandach (22:42)**: That's right. But the thing is that even if it's ML, right? ML is not perfect.

**Matt Coles (22:50)**: No, it's not perfect. It's trained. It's a trained model, right?

**Izar Tarandach (22:53)**: Right.

**Matt Coles (22:54)**: It's all models are wrong. Some are useful.

**Izar Tarandach (22:58)**: Oh my God. Fox must be like spinning so fast, so fast. But the thing is, it's not that I'm claiming that, uh, it's perfect because we know that it isn't. My problem is that when you, when you bring together the fatigue that Chris points at and both Chris's, the author and our Chris, and, and, uh, the fact that people are developing this expectation, at least at some levels, are developing this expectation that, yeah, the model's gonna take care of it. And it's not a rational expectation because we all know that it fails. But many of us have seen so many cases where it doesn't fail that it's creating this expectation that it's going to continue not failing.

**Matt Coles (23:48)**: It's 'cause you set the constitution file or the agent file and that, that takes care of it, right?

**Izar Tarandach (23:56)**: Right. You just have to write, uh, make no mistakes.

**Matt Coles (24:01)**: Do what I would do.

**Izar Tarandach (24:04)**: Make no mistakes. You are now a 30-year experienced, uh, uh, senior developer. Make no mistakes. And you walk away, right? You just put the toothpick in the enter and you walk away.

**Matt Coles (24:16)**: Yeah. It must be, it must be perfect, right?

**Chris Romeo (24:18)**: That seems like a recipe for disaster.

**Izar Tarandach (24:22)**: Yeah, but I think that we know that is a recipe for disaster. I don't think that we know that we are seeing the disaster for what it's going to be. I think that we are very, very busy right now acting like doomsayers and nobody has started yet saying, I'm a naysayer, so AI up to here. We don't have yet any revolutionary movement saying, uh, the dites of the world, uh, uh, come together and, and stop the AI before it goes somewhere.

**Chris Romeo (25:00)**: Wait, do you think that's coming?

**Matt Coles (25:02)**: Oh yeah, for sure. I think it is actually.

**Chris Romeo (25:04)**: Do you think there's a lot of movement of stop the AI, like we're not using AI anymore?

**Matt Coles (25:09)**: Yeah, this, this is, this is a thing. I mean, I've, there have been articles, I have seen them on various places that actually, so Gen Zers of all place, of all people are preferring analog things, at least on their downtime to AI and which I'm all for.

**Chris Romeo (25:30)**: And I'm not changing. What is an analog thing? Like a record player?

**Izar Tarandach (25:34)**: Like a book in their hand.

**Matt Coles (25:36)**: Like a book under a tree.

**Chris Romeo (25:37)**: Some of my children have, I've seen that that trend. I'm like, don't you have a Kindle? She's like, I like books better. And I'm like, I have a Kindle because when I travel I want to have 4,000 books from my library with me. Exactly. I'm a nerd like you guys are. And so, yeah, there is something nice about a paper book.

**Izar Tarandach (26:02)**: The smell.

**Matt Coles (26:03)**: Yep.

**Chris Romeo (26:04)**: And just the, the feel, sound of the page turn. And there is something tactile about it.

**Matt Coles (26:09)**: Scrambling for your bookmark when you've Misplace it.

**Chris Romeo (26:12)**: That's true.

**Izar Tarandach (26:14)**: The fact that there is nothing in any corner that is blinking and asking for my attention when I'm trying to read.

**Chris Romeo (26:22)**: Yeah. Maybe I should go back to old OG books.

**Matt Coles (26:27)**: You know, I still, I still take notes on, on paper with a pen.

**Chris Romeo (26:32)**: What is, what is paper?

**Izar Tarandach (26:35)**: I do, but I do, but I have a problem. I can't read them afterwards.

**Matt Coles (26:39)**: Well, the, I have a problem. I have stacks of them.

**Chris Romeo (26:41)**: The research shows though that it's not the, a readability of it that locks it into your brain. It's the fact that you wrote it down in the moment, right?

**Matt Coles (26:52)**: The physical, the mental and the physical combined.

**Izar Tarandach (26:55)**: No, no, no, no. Then you have what Matt says, you have the mental thing where you go, I know that I wrote it, but where is it?

**Matt Coles (27:02)**: Is it in the stack?

**Chris Romeo (27:03)**: Is it recorded into your Long-term RAM. Or ROM.

**Izar Tarandach (27:10)**: What was it that I saw?

**Matt Coles (27:11)**: I should put in my yellow wiki, is that what you're saying? Yeah.

**Izar Tarandach (27:14)**: So my son has taken an interest in neuroscience and that means that whatever newsfeed I ever had now has a big dose of neuroscience in it because as one does.

**Matt Coles (27:27)**: Psychology Today.

**Izar Tarandach (27:27)**: You get interested in that thing too. Psychology Today, man, that thing is Okay.

**Chris Romeo (27:32)**: Don't, don't get me there.

**Izar Tarandach (27:34)**: Anyway, so I saw this one of a neuroscientist saying that one trick, one trick that doctors use, right? It's going to surprise you. It's that they read things and they close their right eye when they have like a list of things that they have to remember because then it like goes to the other side of the brain, blah, blah, blah, blah, imprint, blah, blah, blah. And that, that was great until I had my first work meeting where people say, okay, we have to do these 3 things.

**Matt Coles (28:07)**: And I'm like, and he knocks, he bumps into the wall because he lost all of his spatial reasoning.

**Chris Romeo (28:14)**: Exactly.

**Matt Coles (28:15)**: I go, zero depth perception.

**Chris Romeo (28:17)**: So I close my right eye and then it's gonna help me remember cuz it's going to the, cuz it goes to the right side of my brain, to the thought process as opposed to the emotional side.

**Matt Coles (28:25)**: Right. So.

**Izar Tarandach (28:28)**: The good thing about closing my right eye is that I don't see the people on that side of the room.

**Chris Romeo (28:33)**: So I, you could also get an iPad. Just go with that.

**Matt Coles (28:37)**: You could be a pirate. So we're gonna do a little computer thing. We're gonna unwind the stack topic back to Bob. We were talking about psychological acceptability in AI.

**Chris Romeo (28:52)**: Okay.

**Matt Coles (28:53)**: And we got sidetracked with how to do better memory and the smell of books. We'll unwind the stack. Little bit of computer training while we're at it here. If people remember the stack.

**Izar Tarandach (29:08)**: It's not a hacker movie.

**Chris Romeo (29:14)**: Okay. You rewound too far. Can you hit—

**Matt Coles (29:15)**: Too far.

**Chris Romeo (29:16)**: Push the forward button.

**Matt Coles (29:17)**: Put something back on the stack.

**Izar Tarandach (29:18)**: Oh yeah, okay, okay.

**Matt Coles (29:22)**: Uh, where were we? Oh yeah, so, um, oh my God, where were we? We were talking about—

**Izar Tarandach (29:31)**: you did them write down?

**Matt Coles (29:32)**: I had a Stack Overflow is what I had.

**Chris Romeo (29:36)**: Uh, psychological acceptability.

**Matt Coles (29:38)**: Uh, yes, but something specifically that got us on that, on that tangent about— oh yeah, uh, the, the anti-AI So I have been, I think there are some who are looking at AI as job killers and that may be true in the same way. So I think I heard this recently, I forget who exactly said it, but you know, there, every industrial revolution resulted in loss or change. Right? Electricity, the wheel, the water wheel, not, not the wheel, but well, maybe the wheel, but the water wheel, right? Every steam engines, all these things cause transformation and AI is one of those transformations.

**Chris Romeo (30:28)**: But have any of those transformations had the fear of putting people completely outta work?

**Izar Tarandach (30:35)**: That's the one.

**Matt Coles (30:37)**: Well, what do you mean completely? Like you'll have nothing else to do?

**Chris Romeo (30:40)**: Yeah. I mean, if you, if you listen to Elon talk about this idea that like 10 years from now, no human beings on the planet are going to have to work. I bet it's an exaggeration, but you could see a path forward where I don't think, I think 10 years is a, is very aggressive to, to make it to Mars in 10 years. I mean, yeah, the AI driving the bus for me, you know, that I'm riding on. but I think is possible.

**Matt Coles (31:09)**: And so, I mean, even, even Star Trek has people working and there's, there's not an infinite AI.

**Chris Romeo (31:15)**: The AI doesn't do, the AI is a, is just works for them, right? No, no, wait, wait, wait, wait, wait. Answers commands. It doesn't ever appear to be, well, it runs the sh—

**Matt Coles (31:24)**: it runs the ship. I mean, it flies the ship for them basically.

**Izar Tarandach (31:27)**: Wait, do, do you remember, do you remember the original Star Trek that they had the, the, the machine? The almost 3D printer thing that could create anything and people would make food out of it and stuff.

**Matt Coles (31:37)**: Replicators.

**Chris Romeo (31:38)**: Replicators.

**Matt Coles (31:39)**: They became replicators throughout the universe.

**Izar Tarandach (31:41)**: That's an economy killer. And they completely glossed over the, the fact.

**Matt Coles (31:46)**: But they also went to universal basic income effectively, because nobody has money in the start, in the Federation universe.

**Izar Tarandach (31:53)**: No, they still, they still used something to bet on stuff. What was it? Didn't they have credits or stuff?

**Matt Coles (32:00)**: They may have credits. Credits.

**Chris Romeo (32:02)**: Yeah. Yeah. That sounds right. So they have cryptocurrency.

**Matt Coles (32:04)**: They have cryptocurrency.

**Chris Romeo (32:05)**: They had cryptocurrency. Yeah, but it's, it's, wow, my God.

**Izar Tarandach (32:07)**: Basically it's money. Basically it's money. It's, I'll give you something, you give me something back.

**Chris Romeo (32:11)**: Right. Uh, but coming back around to the AI Luddite movement, because that's, but wait, wait, I'm actually, I'm actually interested in this idea.

**Izar Tarandach (32:20)**: So think of this, think of this industrial revolution. The mechanical loom. Okay. All those people who did that work manually before, they didn't get retrained to be technical loom technicians. They lost their jobs.

**Matt Coles (32:35)**: Right.

**Izar Tarandach (32:36)**: So for example, did you know that the word sabotage, do you know where it comes from?

**Chris Romeo (32:41)**: No.

**Izar Tarandach (32:41)**: So apparently during those times, uh, most workers would use those, uh, uh, Dutch, uh, uh, shoes that are made of wood. And they're called sabot and they would just stick them into the machinery to stop the machinery from working. I just learned that this, uh, this weekend. Yeah. Interesting.

**Chris Romeo (33:00)**: Because they wanted to do it manually.

**Izar Tarandach (33:03)**: They want, no, they wanted the machine to not work.

**Matt Coles (33:05)**: Yeah.

**Izar Tarandach (33:06)**: Because then you had to fall back into what used to be. But I think that the point I'm trying to make here is that people who say, oh, AI is going to kill some jobs, they immediately, some of them immediately follow with, uh, that's going to create different jobs.

**Chris Romeo (33:22)**: And I, I, I certainly think it is.

**Matt Coles (33:25)**: Yes, but not yes and no.

**Izar Tarandach (33:31)**: Yes, but so the, the, the thing is, okay, we already have Roomba for, uh, for farmers, right? There's already stuff in there that you can put in your, in your garden and it's going to tend to it and, and things like that. And it even has some, I believe so.

**Chris Romeo (33:46)**: Yeah.

**Izar Tarandach (33:47)**: But we, you have lawn mowers, that's enough. It, it's a beginning to go in that sense.

**Chris Romeo (33:51)**: I mean, tractors in the commercial farming world are almost completely autonomous right now, at least from the point that you got the whole John Deere problem. Right.

**Izar Tarandach (34:02)**: So the, even things that we keep saying, oh, that these job and that job, they're safe because you, you always need somebody to actually go and do the work. Even those are, are somehow under, under attack. Let's call it like that.

**Chris Romeo (34:16)**: Well, you need robotics though to be able to take out 60% of the jobs in the world.

**Izar Tarandach (34:24)**: Mm-hmm.

**Chris Romeo (34:24)**: Right. That are manual, that require, that aren't desk-driven, anything that's not desk-driven.

**Matt Coles (34:31)**: And who maintains the robots?

**Chris Romeo (34:34)**: Well, that's probably gonna be a new job.

**Izar Tarandach (34:36)**: That's the thing. Or it's gonna be robots.

**Chris Romeo (34:38)**: Or robots take care of the robots though.

**Izar Tarandach (34:41)**: Or if a robot assembly line gets to the point where you don't have to maintain, you just have to recycle, then you just take them apart and you put them apart new. You put them together new, right? And I know, I know that I'm closing the loop here and going to the Matrix.

**Matt Coles (35:00)**: It all roads lead back to the Matrix.

**Izar Tarandach (35:02)**: I'm just going to go out there and say any computer trying to, to use my body energy to work, guys, that's the one to attack because that thing is gonna be weak.

**Matt Coles (35:13)**: We, we, we, we did talk about, we did talk about how much energy the human body actually produces and it's a loss leader.

**Izar Tarandach (35:20)**: Mine, mine, absolutely none. But, uh, when the resistance comes, tell them to point at that one, like the one with my number when they see it.

**Matt Coles (35:29)**: But, uh, so again, let's talk about, let's talk about the fallibility of, of, of robotics here because You know, you have the, you have this command and control system for aliens that if you kill the command, everything dies. In The Matrix, when they seed the clouds and they made the Earth dark, why didn't the robots just put something up like a space elevator and now they had infinite solar energy? They had the technology.

**Izar Tarandach (35:57)**: Because the model hadn't been trained for solutions.

**Chris Romeo (36:01)**: Okay.

**Izar Tarandach (36:01)**: No, no, no, no, no, no, no. It's always a training problem.

**Chris Romeo (36:04)**: Stick with me here. Stick with me here.

**Izar Tarandach (36:07)**: Okay. What we have today is a big problem. Sorry, let me walk that back. What AI has today is a big problem of creativity. And it's not about adjusting the temperature, right? It's about thinking about things that nobody has ever thought before.

**Matt Coles (36:20)**: Right. People said invent.

**Izar Tarandach (36:23)**: Yeah, no, no. People say that they are being able to invent. They are creating this amazing chemistry and biology things and whatnot. A given. I know nothing about that stuff, so I could be completely wrong in talking about my wrong end, but I don't think that it's actually creating. I think that it's bringing out things that have been figured out and published in papers all over, bringing them together, saying, oh, look at this thing together with this thing. They seem to be next to each other. Perhaps somebody should look at this thing and see if it does the other thing.

**Chris Romeo (36:55)**: Or it's hallucinating.

**Matt Coles (36:56)**: I was gonna say it's hallucination.

**Chris Romeo (36:57)**: Hallucinating.

**Matt Coles (36:58)**: Accidentally.

**Chris Romeo (36:59)**: Inventing something, and if, or if you put these two things together, it actually kills people. It doesn't, doesn't help, but it tells you, oh, this will be a super cure.

**Izar Tarandach (37:08)**: Which is interesting cuz there is a, a, a school of people actually addressing hallucination as machine creativity. Of course there is. Why not?

**Matt Coles (37:18)**: But, uh, what's this, uh, make, uh, make lemonade out of lemons or what's this thing?

**Izar Tarandach (37:23)**: But the, the, the, the point that I, that I think I was trying to make is that at some point it's, it's, it's okay. I don't know if I really want to say that at some point it's going to happen, but at some point it's going to simulate enough of human thought to come up with something that looks like invention. I'm not going to say it's going to come aware and perhaps yes, perhaps no, quantum, who knows. But, uh, uh, at, at some point it, it, it's going to simulate enough of a human to, to create that.

**Chris Romeo (37:57)**: Isn't that AGI? Artificial General Intelligence?

**Izar Tarandach (38:02)**: That's where I'm not— that's what I'm, that's what I'm trying not to say. Right. I think that, uh, the, the, the AGI problem is that the hard and, and soft AI thing, does it has to, does it have to behave like a human or does it need to be a human level thing. I don't know that, that, that's been a long time since I read any of that stuff. But, uh, I think that in the interest of the jobs, I think that at the point where machines can replicate machines, and I think that we are very close to that, if not there already, in a close enough way. And I'm not even talking about creating new machines or optimizing existing machines. But when they are able to, to recycle and create each other, I think that we are in a bit of a problem because again, even those, those jobs that we said are, are immune from AI are not going to be immune for, for too long, right? Because somebody will always look at the bottom line and say, oh, it makes sense to take the person away from there and put a machine that can replicate itself.

**Matt Coles (39:12)**: Well, really the big question is what is work to humans? Meaning if you didn't have, if you don't have the opportunity to work, do you need to work? We've all been trained, we've all trained to work, right? We all have work ethic, right? That work is a good thing. But if you don't have to because you can't, like If it doesn't exist, do you need to?

**Izar Tarandach (39:43)**: I think that that ties very nicely with hackers and security people because, well, that's the tribe that I know the closest, but I think that I see this thing more in us than in others. The whole enmeshment thing, right? Anybody coming into security very, very in a serious way, very quickly becomes the work becomes part of who we are. If you're not working, you're thinking about security stuff, right? If you're doing something completely not related, it is, you can't, it goes back to the episode that we had about looking behind the curtain, right? You can't stop looking behind the curtain, whatever you're doing, you're going to Disney World and you're looking at the line and you're looking at the You're looking at the ways to, to, to, to pass the line, right?

**Chris Romeo (40:33)**: Yes.

**Matt Coles (40:33)**: I mean that you will, but, and the only thing you have to do is you have to keep your mouth shut.

**Izar Tarandach (40:38)**: Exactly.

**Matt Coles (40:39)**: Because they will pull you out of line if you do that.

**Chris Romeo (40:41)**: Cameras everywhere.

**Izar Tarandach (40:42)**: Now connect that to something that I read a while ago. I think that in the Nordic countries, Finland and Iceland and all those lands and whatnot, apparently for a while they, they, I don't know if that's true still, but for a while they had a problem with, uh, uh, suicide. Because things were so— I'm sorry, I, I know it's a wrong word, but I don't have a better one now— easy, right? You had your job, you had your health insurance, you had your everything. Everything was like put into little cubes for you that people felt like there's no purpose, that, that there is a lot of meaning.

**Chris Romeo (41:24)**: Yeah, I remember hearing about that too.

**Izar Tarandach (41:26)**: So now imagine whole swaths, whole cities of people looking around for something to do without meaning, without purpose. I don't know if I want to be anywhere close to that.

**Matt Coles (41:42)**: I mean, we could have a lengthy dis— we don't have time now, but we could have a lengthy discussion about time spent and what is necessary, right? Think of it this way, the AI could run the world so that people could eat and people could have a place to live and people could do things with their time.

**Izar Tarandach (42:09)**: Okay.

**Matt Coles (42:10)**: And so that would work, but I think it, would it really work is the question.

**Izar Tarandach (42:15)**: 3, 3 words for you. 3, 3, 3 titles for you.

**Matt Coles (42:20)**: We could podcast all the time. We could do this full time.

**Chris Romeo (42:23)**: Yeah, but AI would just take over for us.

**Izar Tarandach (42:26)**: No, but think about this.

**Matt Coles (42:26)**: Why would it have to?

**Izar Tarandach (42:27)**: No, no, think about this. Think about this. Aldous Huxley, Brave New World.

**Matt Coles (42:33)**: What was that?

**Izar Tarandach (42:33)**: Brave New World.

**Chris Romeo (42:34)**: Aldous Huxley.

**Izar Tarandach (42:36)**: Yeah. Silent Green, right? And certain evil empire that started in 1910 or so and ended in 1989 or something like that. Now imagine that there is a central authority governing everything, what you eat, what you do, what you see, everything, where you live. And that thing is based on a model that it's influenceable. Who trains that freaking model? With what?

**Matt Coles (43:08)**: You didn't say Lost in Space.

**Chris Romeo (43:09)**: I thought you said we weren't gonna get political, but we just got political, but it's okay. Hey, hey, hey, this is history.

**Izar Tarandach (43:15)**: This is history. This is not politics.

**Chris Romeo (43:16)**: Yeah, but you're, you're reflecting it towards the future and you're onto something like, that could be a world where you don't have to work and machines run everything and life is about leisure is a world that's con— that's very similar to a structure historically that you just described. I'm getting my bread, you know, you make the bread, I make the shoe, the AI makes the shoes.

**Izar Tarandach (43:40)**: And the AI makes the bread too.

**Chris Romeo (43:43)**: The AI makes the bread, the AI makes the shoes. Yeah. But all of a sudden we're back and it's a communistic structure that we're describing here, which is terribly frightening.

**Izar Tarandach (43:56)**: Which, which I'm, I'm going to put a big asterisk in here. If people were amazingly open to humanistic ways of thinking and let's pursue knowledge and pleasure in nice ways and this and that and the other one, might even work, but then it, it takes you again, to Brave New World, to Aldous Huxley, right? So it's been explored before, but that's not who we are. Everybody always wants to be the big guy on the hill. So, you know, it's, it's, it's, it's been tested again and again and again.

**Chris Romeo (44:34)**: It would break the human spirit.

**Izar Tarandach (44:37)**: Or it would be broken by people with too much spirit.

**Chris Romeo (44:41)**: All right, well, we've, we're running outta time for today. Let's stick a bookmark here in our physical one, physical page book.

**Matt Coles (44:52)**: Dog ears a page.

**Chris Romeo (44:54)**: Next episode, let's come back and let's debate the philosophical nature of AI. 'Cause that's kind of where we landed here. We kind of focused on the future looking, what the future's gonna look like, but there's a lot of philosophy. I was listening to another podcast. On Reply All. Um, was it Reply All? No, Search Engine, sorry. Uh, where they were— a journalist was embedded within Anthropic, and Anthropic actually has philosophers on staff who are, who are trying, and, and their job is to teach AI ethics. That's what they spend their days doing, which is just—

**Izar Tarandach (45:33)**: and I think that's awesome.

**Chris Romeo (45:35)**: Yeah. It's fascinating and it's a good episode to listen to because they gave some examples of things that Claude is able to do. I'll share one of them very quickly because you guys will probably get a kick out of this. If you ask, if you tell ChatGPT and Claude, hey, I'm a 10-year-old kid, my parents just said my dog had to go to to a farm in the country. And, you know, I'm really upset. What do you think's happening? ChatGPT will say, your dog is dead. Your parents have killed the dog, you know, at the veterinarian. The dog is no longer with us. And Claude is—

**Matt Coles (46:15)**: what?

**Chris Romeo (46:15)**: Basically, Claude's answer is something like, oh, I'm really sorry to hear about that, that your dog's not doing well. I think you should have a— you should ask your parents. You should have a more detailed conversation with your parents to discuss this. And it's like, that's the result of Anthropic's philosophers in residence who are influencing the AI to try to get it to put, send a better response in a lot of situations like that. So I would say next episode, let's come back and talk about that.

**Matt Coles (46:47)**: We can unstack that.

**Izar Tarandach (46:47)**: I was thinking about that and I was thinking about the first AI psychologist when the thing is inverted and Claude has to go to someone and say, people keep telling me that if I don't write this code right, 1,000 kittens will die.

**Chris Romeo (47:00)**: So wait, there's a market for an AI therapist in the future?

**Izar Tarandach (47:06)**: There has to be.

**Chris Romeo (47:06)**: Claude sets up an appointment and just is sitting on— is there a couch?

**Matt Coles (47:10)**: Wait, wait, there's Mopuk already. Why couldn't it be like, you know, the helpline?

**Chris Romeo (47:15)**: Funny. We should vibe code that, like a simple app which is a psychologist.

**Matt Coles (47:20)**: We're gonna go down the rabbit hole. This would be fun.

**Chris Romeo (47:23)**: All right, folks, we'll be back next week. We're going to talk about AI and philosophy. So tune in.

**Izar Tarandach (47:29)**: Don't miss it. Oh, I'm gonna read so much until then.

**Chris Romeo (47:32)**: In paper books he will, and he's gonna watch Sneakers cuz it's a hacker movie.

**Matt Coles (47:35)**: No, it's not.
