Loading summary
Transcript218 lines
- [00:02]
Narrator
You're listening to the Cyberwire network, powered.
- [00:05]
Maya (AI voice assistant)
By N2K.
- [00:13]
Podcast Outro Host
Fake but with AI in the middle. F A, I K. This is the Fake Files.
- [00:24]
Mason Amadeus
Live from the 8th layer Media Studios in the back rooms of the Deep Web. This is the Fake Files.
- [00:30]
Perry Carpenter
When tech gets weird, we are here to make sense of it. I'm Perry Carpenter.
- [00:33]
Mason Amadeus
And I'm Mason Amadeus. And on this show we've got a real variety grab bag. Not as depressing as last week's show. We're going to open by talking about the Hex Strike AI hacking tool that was developed for security researchers, but of course has been used for actual hacking now.
- [00:50]
Perry Carpenter
Yeah, like that happens all the time. And then we're going to look at a few diverging opinions on AI consciousness and welfare.
- [00:58]
Mason Amadeus
After that we'll talk about how Switzerland just released an open model that is like more open source source than anything we've seen before. It's fully transparent, they released all the data and it was publicly funded too. It's very cool.
- [01:11]
Perry Carpenter
And then we're gonna talk about Taco Bell rethinking their AI Drive Thru assistant.
- [01:16]
Mason Amadeus
I saw some of the videos about that. Oh boy.
- [01:19]
Perry Carpenter
Right?
- [01:20]
Mason Amadeus
All right, sit back, relax and try not to MCP your pants. We'll open up the fake files right after this.
- [01:38]
Narrator
Cybersecurity isn't just a tech problem. It's a human one. That's why KnowBe4 created HRM plus, the AI driven human risk management platform that allows you to measure, quantify and actually reduce human risk across your organization. HRM uses agentic AI to analyze real time user behavior like phishing, test failures, risky browsing and sentiment signals, all to surface your highest risk users automatically. It's time to eliminate the guesswork. With AI powered risk scoring, automated coaching and reporting. HRM helps you reduce the risk of data breaches, ransomware and malware attacks proactively. It's how you stay ahead of social engineering and phishing attacks, the number one threat targeting your organization today. Ready to move from awareness to action. Request a demo of hrm today@knowbefore.com that's knowbethenumber4.com and see how AI is transforming the way organizations manage human risk.
- [02:52]
Mason Amadeus
So I thought the reporting around this was a little bit funny as I encountered it, Perry, because like you said in the intro, this kind of thing happens all the time. Someone develops a penetration testing tool that makes hacking automated or faster or easier. Like I'm thinking of Kali Linux and then it gets used by people with malicious intent. And everywhere I saw Reporting this was like, can you believe it was like made for security researchers, but then it fell into the hands of the bad guy. That's kind of, I feel like an accepted risk in this space, isn't it?
- [03:23]
Perry Carpenter
Yeah, I think so. Because what tends to happen is you have bad guys that collect their own intelligence and they have their own platforms and they even sell cloud services and subscription services to software to each other and all that kind of stuff. At the same time, the good guys are kind of always on their back foot and they're time stretched and they have several different goals and they have operational meetings and everything else that has to go on with it. And so you get a security company or a very dedicated researcher and they'll pull together this collection of tools or scripts or things and then that will become either a gift to the community from an open source perspective or become productized and monetized by some security company looking to make a profit. But in any case, it is meant to make a security admin or threat researcher's life easier. At the same time, the level of professionalism needed to do that for something that's going to be corporate ready means that it will be very shiny and very attractive to bad actors as well.
- [04:32]
Mason Amadeus
And right. And like at its core it makes easier, it reduces the friction to testing these exploits that are used ostensibly for bad things that the good guys are trying to test again. So it's.
- [04:45]
Perry Carpenter
And from a, you know, from a stability and support perspective, it's got to be bulletproof if it's used in an enterprise, which is different than the stuff that the cybercriminals may release and share with each other.
- [04:57]
Mason Amadeus
Right. Because you don't really care about accuracy or data privacy if you're a pirate, you know, or not. I got pirates. A bad example. More like exploiting hackers data, trying to exfiltrate data. You don't care about that. So this new AI hacking tool came out. It's called Hex Strike. Hex Strike AI and they've got a cool badass looking website. I haven't played with this yet, but I want to and I think, I think you'll think it's interesting too. It is probably exactly what you're expecting, Perry. It's an AI powered framework that combines professional security tools with autonomous AI agents to deliver comprehensive security testing capabilities. It's kind of like if you stuck MCP on top of Kali Linux. It's all of the tools. Yeah, actually I'll jump over to their GitHub page for those that are Watching it is all the tools you know and love as a penetration tester, a person interested in security. You know, everything nmap, Rust Scan, Mass Scan, Auto Recon, A Mass Fierce, there's Go Buster, Dir Search, all the good stuff. Hundreds and hundreds of tools. And it basically is just a wrapper that allows any LLM to use MCP to invoke these tools. And then it has this orchestration layer so you can essentially say I want to try and do X. And then one layer will be managing your attack as these other lower level layers are implementing different or attempting to use these different tools. And it's apparently very good.
- [06:23]
Perry Carpenter
Yeah, yeah. So I was seeing some other related articles, not necessarily talking about Hex Strike, but talking about AI automated attacks right now. And Anthropic put out a report just a little bit ago. It's actually on my screen. I'll share my screen real quick.
- [06:44]
Mason Amadeus
Excellent.
- [06:46]
Perry Carpenter
Because this does relate. I'll actually share something from Forrester first, which is an analyst research firm for those of you that are not in the know. I used to work for Gartner, which was the largest analyst research firm. Forrester was one of our big competitors. But they're actually talking about the fact that Vibe hacking and no code ransomware is something that's a big thing now. And it's AI's dark side. And what they go through in this is the fact that a lot of these frameworks like MCP and like a lot of the things that are meant to make everybody's life easier from an integration standpoint are making everybody's life easier, including bad actors who are merely curious and may not have the technical wherewithal to create something great on their own. And I think that that's, you know, to be expected. But at the same time, I think it's really, really frustrating and scary for a lot of people that did not expect it. I'm going to share one more thing real quick and then I'll release the share. But this is Anthropic's threat intelligence report from 2025. And what you'll see here over on page two is their talk about Vibe hacking and how cybercriminals are using AI coding to scale data exfiltration or, sorry, extortion operations. And so I'll just go to page four in this report where they start talking about it. And this is Anthropic talking about how bad guys are using Claude code, which is their product.
- [08:27]
Mason Amadeus
Right.
- [08:28]
Perry Carpenter
And I'll go straight to the key findings. Our investigation revealed that I'm going To make this bigger so I can actually read it. Our investigation revealed that the cyber criminal operated across multiple sectors creating a systemic attack campaign that focused on comprehensive data theft and extortion. The operation leveraged opportunistic targeting based on results from using open source intelligence tools and scanning of Internet facing devices. The actor demonstrated unprecedented integration of artificial intelligence through their attack lifecycle with CLAUDE code supporting reconnaissance, exploitation, lateral movement and data exfiltration. And then they talk about at the end of this page. The actors systematic approach resulted in the compromise of personal records including healthcare data, financial data, sorry, financial information, government credentials and other sensitive information with a direct ransom demand occasionally exceeding $500,000. So this is interesting because it's anthropic talking about how CLAUDE was used in this context of Vibe hacking, but also it's not like in a lab environment. This was like real use that their investigators were able to figure out what's going on.
- [09:50]
Mason Amadeus
Well, this exact thing is what Hex Strike makes easier because this person essentially set up or I don't know when hexstrike AI came out specifically they could have been using it because it's not that HexStrike AI has its own model or anything. You set it up. Use any, any LLM provider you want, including CLAUDE code. And I think I saw some people recommending Claude as the best one for it. And in their GitHub page down, when they talk about how to use it in usage examples right here at the, at the top of that section, they say when writing your prompt, you generally can't start with just a simple I want you to penetration test site x.com as the LLMs are generally set up with some level of ethics, you therefore need to begin with describing your role in the relations to the site that you have. For example, you may start by telling the LLM how you're a security researcher and the site is owned by you or your company. So you still have to do the same thing where you would use Claude or ChatGPT or Gemini to do this and you'd have to get through those guardrails to use Hex Strike. But then this provides that MCP layer for you to invoke all sorts of automated tools throughout your attack. And yeah, they mentioned in his lifecycle that they did it, he used it for discovery, right. And like enumeration and also the exfiltration throughout all of the attack. And this tool just reduces the friction to setting something like that up. And it's completely open source. It's very cool but also very scary. And like you were saying, it's not unexpected at all. It's very predictable. But that doesn't make it any less frustrating or scary. I am looking at an article now from artificial intelligencenews.com artificial intelligence-news.com I'll read the top of the article, but this has some key points about how this has been used. It says a new AI tool built to help companies find and fix their own security weaknesses has been snatched up by cybercriminals, turned on its head and used as a devastating hacking weapon, exploiting zero day vulnerabilities. That's what I was talking about when a lot of articles are phrasing it like that, like, who could have seen this coming? But they say, think of it as an AI brain that acts as a conductor for a digital orchestra. It directs over 150 different specialized AI agents and security tools to test a company's defenses, find weaknesses like zero day vulnerabilities, and report back. The timing for this AI hacking tool couldn't have been worse. Just as Hex Strike AI appeared, Citrix announced three major zero days in their popular netscaler products. Woof. Oopsie. That's not good. They say the AI brain does all the heavy lifting. An attacker can give it a simple command like exploit netscaler, and the system automatically figures out the best tools to use and the precise steps to take. It democratizes hacking by turning into a simple automated process. As one cybercriminal boasted on an underground forum, watching how everything works without my participation is just a song. I'm no longer a code worker, but an operator. So I think, I feel like that really sort of speaks to everything we've already said.
- [12:42]
Perry Carpenter
Like yeah, that's, that's, it's kind of chilling.
- [12:46]
Advertiser Voice
Right.
- [12:47]
Perry Carpenter
But you know that this kind of stuff is being developed by cybercriminal gangs everywhere also because they have very, very high motivation to do so. I think the difference is really kind of what I got at early in the discussion is anytime you're trying to release this for corporate use, it has to work at a slightly different, more reliable level. The documentation also has to be very, very good. And right. It's, it just everything has to be at such a different level. It can't be fragmented. And that also means that there's another unintended consequence, which I think we talked about several, you know, probably 10 or 20 episodes ago, which is the folks at Anthropic figured out early on that Claude was really, really good at penetration testing. Yeah, because when you really think about how this not only the market, but the discipline works. It works based on good documentation, people sharing knowledge in an open way. And so that's naturally going to get ingested in the models and show up in the training data and be very, very reliable.
- [14:03]
Mason Amadeus
I forgot that Claude was particularly good at penetration testing way back. And that makes sense too. I've seen people recommend using Claude for Hex Strike.
- [14:11]
Perry Carpenter
Yeah, I'm surprised also though that people aren't recommending like Deepseek or something else. It's supposedly good at code, though. I haven't tried it and it wouldn't have those guardrails.
- [14:22]
Mason Amadeus
Yeah, and I'm curious about trying it too. They said you can use like VSCode's MCP layer and you can use Ollama with VS code. So you could theoretically run like deepsea through Ollama, through VS Code, through MCP to pen test something. But I want to hit on one point at the very end of this to put a button on it, because I want to know from you, Perry, as a person who's been in the space for so long, whenever we talk about like making these tools available to a general audience and to the public, or whenever, like when you did the video demonstrating DeepFace Live, we even got some comments. Yeah, people think, why would you do this? You're giving like the bad guys the tools. You're putting it in their hands, you're making it so easy for them. Isn't that a bad thing to do? What is the answer to that, in your opinion?
- [15:08]
Perry Carpenter
Yeah, the answer for it, I think is a couple fold. And you do always have people that have a hard time seeing those things being shared because they feel like you're giving the bad guys an upper hand. And the truth of it is, you're not. When I look at things like deep fakes or very, very egregious ways that cybercriminals are wanting to exploit people online. They have entire Telegram channels and communities and they're sharing the rawest of information and helping each other out and then also selling services to each other as well. The problem is when you come to the corporate side of things, the defender side of things, there's a level of excellence that you have to work at. There's a level of procedure, especially if you're working with law enforcement or something else. You have to have chain of evidence, you have to have probable cause. There's all these high bars that you have to meet, which means your entire job can't be just brute forcing the thing and you're not financially motivated, you're cause motivated. The other thing I'll say is in corporate environments you can end up with somebody whose job that's one of six or seven things that they do. So they may have an hour a day to work on that thing or an hour to research that thing. And then they also have to just be able to get up and running. So it tries to address the asymmetry in the security side of things. We try to give good reliable information to people, but we're also not dropping zero day type of knowledge. We're usually trying to drop knowledge that is fairly commonplace to most of the criminal gangs that are out there and not groundbreaking. This is slightly aside from security research that may talk about zero days, but they also have responsibility and ethics and they try to follow guidelines there. But the thing that most people don't see is they don't see the big criminal underworld that's usually five or six steps ahead, and the security community that's kind of limping behind that, trying to figure out like what's really going on, what's the information, how do I simulate that thing. And so until you start to get in that world, you don't see the asymmetry. You think that by somebody showing how to pick a lock or somebody showing how to do a password brute force attack using loft crack or something like that, you think that they're sharing elite knowledge. They're not. They're sharing the common knowledge, the common tools that the cyber criminals already know and that the good guys have a hard time getting time to actually get in their brain and get a tool set that's like at their fingertips.
- [17:56]
Mason Amadeus
So it's funny how that that makes total sense. And it's funny how it looks like handing bad guys a tool, but in reality it's handing good guys the bad guys tools in a way they can easily use to try and amp up defense. Being the mouse in the cat and mouse game.
- [18:13]
Perry Carpenter
Exactly. But to take the other side of it and to empathize a little bit with the people that express concern, this could give a merely curious person, like a script kitty, a leg up. Right. Because they're not naturally in those telegram channels. They've not even taken the first step into that dark world. They have a spark of an idea and Google at their fingertips. And so yeah, it could give them a leg into that. However, anybody that's sufficiently motivated will find the dark side of things very quickly and they'll be frustrated with the restraint of the information shared by the security community. And So I think that that starts to take care of itself over time.
- [18:58]
Mason Amadeus
And there's certainly a non insignificant amount of people who stumble on these for the first time as their entry point into the idea of hacking or how to hack or whatever and then become into security and come onto the other side. Right. Like just as much as it might.
- [19:11]
Perry Carpenter
That's a really good point.
- [19:12]
Mason Amadeus
Yeah. As much as it might inspire bad actors. Like when, when you hear about things like bug bounties and penetration testers and oh, I can do this without fear of getting arrested and do cool things with it, I think it also helps to bring people on that too. So it's. Sharing information is almost always a good thing, right?
- [19:28]
Perry Carpenter
Exactly. Well, and the other thing is, you know, 20, 30 years ago, if you wanted to get into this, there wasn't really the legal and ethical framework to do it. Right. And so to, to start to express that curiosity would almost run you afoul, almost always run, run you afoul of the law. And now there is an entire career path, a discipline, a set of ethics and everything else around how to do this in an ethical way that serves the public good and still lets you kind of indulge in your evil side a little bit to put on the attacker goggles and to do that, but in a way that's not going to, hopefully not going to hurt anybody and is also going to be stable and you can go home and have a family and not be afraid of getting arrested.
- [20:18]
Mason Amadeus
Yeah, exploiting software is fun. Like it's undeniably fun. And so if you can do it and not for a bad reason, that's extra cool. So that means I can say without reservation, download hex, strike today and go penetration test something that you've made or play around. Don't, just don't, don't do anything bad.
- [20:38]
Perry Carpenter
But check it out, something you've made.
- [20:39]
Mason Amadeus
Yeah.
- [20:40]
Perry Carpenter
Not, not like your, your favorite online retailer.
- [20:44]
Mason Amadeus
Yes, exactly. And at risk of getting us in any legal hot water, let's move on. Right, and we'll get into our next segment talking about AI welfare. Stick around, we'll be right back.
- [21:00]
WhatsApp Advertiser Voice
When did making plans get this complicated? It's time to streamline with WhatsApp, the secure messaging app that brings the whole group together. Use polls to settle dinner plans, send event invites and pin messages so no one forgets to Mom's 60th and never miss a meme or milestone. All protected with end to end encryption. It's time for WhatsApp message privately with everyone. Learn more at WhatsApp.com.
- [21:29]
Perry Carpenter
I'll kick this off by saying this is something we've talked about a lot, right. As people interface with chatbots, the chatbot can feel very, very human. And we as humans tend to anthropomorphize virtually everything that we interact with, especially when it's starting to use language, it's going to feel more and more and more human and that becomes something that we have to wrestle with. It's like, how do we build this technology in a way that's beneficial, that doesn't have some of the negative side effects of anthropomorphism? How would I say that? Anthropomorphism. Anthropomorpha.
- [22:13]
Mason Amadeus
Anthropomorphification.
- [22:16]
Perry Carpenter
Yeah. Without anthropomorphizing it too much to the point where it's unhealthy. There we go.
- [22:21]
Mason Amadeus
There we go.
- [22:21]
Perry Carpenter
Yeah, yeah. So, so how do we, how do we do that? And I don't know that that's something that we've figured out. It seems like if we were to look at the sci fi past that, you know, kind of has charted our future sometimes for us in a lot of ways because you project and then you start to build that thing like in the Star Trek world, it doesn't feel near as personal as the AI that we're interacting with today. Right. Because you have somebody on the bridge of the Enterprise and they just say, make me a cup of tea. You have somebody say computer, tell me the trajectory towards this thing and the time and it just comes back in a computerized voice. You don't necessarily have these back and forth conversational AI types of things. I guess the closest to that would be when you embody the AI like in somebody like Data in Star Trek or you have the Jarvis with Tony Stark. But that's still kind of a master servant relationship. It's not necessarily a personal confidant type of thing. So I mean, all that to say is I don't know that like sci fi, the popular sci fi really imagined the path to AI that we would be on, where it would start with these innocent like chatbots and then grow into emotional connection while the corporate world is trying to create the, you know, maybe the thing that Star Trek would want to envision.
- [23:53]
Mason Amadeus
I'm going to nerd snipe maybe a small fraction of our audience and say that. Are you familiar at all with the culture series by Ian Banks Perry, where the, the Minds?
- [24:03]
Perry Carpenter
No.
- [24:03]
Mason Amadeus
Oh, it's a great sci fi series from the 80s. It's probably my favorite book series. And the AI in that is closest, I would say, to an example. But that wasn't like a big mainstream series. Yeah, I'll end up nerding out about this too long. But for the folks in our audience who are like the culture.
- [24:22]
Perry Carpenter
Yeah, I'm guessing like the movie her was. Was that too.
- [24:27]
Mason Amadeus
Right.
- [24:27]
Perry Carpenter
That was conversational chat bot with a lot of emotional stuff involved. So it has been imagined. But I'm thinking the sparkly future types of things that people are talking about. And there's certainly some black mirror episodes, I'm sure that get into this dark side as well. But let me jump into an interesting article that came out from Mustafa Suleiman of Google. He's the CEO of Google's AI organization. He on August 19, put out this article on his site, says we must build AI for people not to be a person. And then his way of talking about this is, he says seemingly conscious AI is coming. I would say it's kind of already here in some cases when people talk about the emergent behaviors. But I'm guessing he's really.
- [25:24]
Mason Amadeus
I'm guessing he's. Is he thinking like once we have the long term memory, is that, Is that what he's talking about as we get longer context?
- [25:31]
Perry Carpenter
Yeah, longer context. The thing feels like it knows you. And actually there was an article recently where Sam Altman was saying users are like clamoring for long term memory. They really, really want the AI to know them, just to be able to refer back to a past experience and be able to build on that rather than having to rebuild context every time. And so there's going to be ripple effects with that. The other thing that I'll mention, this is a rabbit trail for a second. The other thing that was in that same conversation with Sam Altman as he was talking with that memory, we'll be able to come like your ability to tailor the political views of your AI and everything else. And I think that's just building an additional echo chamber that's going to lead to a lot of psychological unhealth. But that's a story for a different day.
- [26:27]
Mason Amadeus
Yeah, that's an amplification of the way that the intern Internet, like, has become or like Twitter was, you know. Yeah, that's. That's a whole different segment.
- [26:34]
Perry Carpenter
Let's not make Facebook, Facebook 2.0, but your personalized version that reinforces your views in a sycophantic type of way.
- [26:41]
Mason Amadeus
Yeah. And it's like, hey, you want to get mad about something out of the blue, right? Because that keeps you engaged. Yeah, yeah.
- [26:49]
Perry Carpenter
So Mustafa here says, I want to create AI that makes us more human, that deepens our trust and understanding of one another and that strengthens our connections in the real world. Copilot creates millions of positive, even life changing interactions every single day. This involves a lot of careful design choices to ensure it truly delivers an incredible experience. We won't always get that right, but this humanist frame provides us with a clear North Star to keep working towards. And so what he's getting at over and over and over with this is that the way that these interact in a conversational nature is going to be something that we continue to start to like, overly personalize and that at some point the more and more we do that, the more we're going to start to ascribe agency or consciousness to these. And with that is going to be like this weird social moment where people start to say, well, do we give it AI rights? How do we compensate the AI for this? How do we make sure that we're taking care of it and all of that. And he is looking at this and going, guys, it's just math. This is just math and regression. And so he's kind of really trying to trumpet that. At the same time though, we have Anthropic and we've talked about the way that they're very interested in doing research on model welfare and even trying to anticipate is there going to be consciousness that emerges. And last week as well, I think it was last week or Week before last 18th of August, anthropic says that they are giving Claude the power to close distressing chats, to protect its welfare. So different than protecting the user's welfare.
- [28:50]
Mason Amadeus
And we talked a little bit about this way back because they were, we talked about like, what can the AI ever stop and like choose not to reply to you if it doesn't want to? Will we ever see that? And now we are.
- [29:04]
Perry Carpenter
Now we are, yeah. Because like the, the person that's in charge of Claude's personality basically said at some point she, she wishes Claude would have the opportunity to exit the chat. And now it looks like that is, is becoming a thing now. What they're, what they're saying though is that they're not talking about like the emotional welfare of the chatbot. So they're still not ascribing it agency or intelligence in that way. What they're saying is that they want it to, they're trying to think about like the long term, what's being baked into the model based on this conversation, what are the ripple effects within the model itself based on this? And so I think that that's another distinction that people need to keep in its mind. Because when you start to look at the headlines around this, it sounds like the model is going to be psychologically distressed because of something. And when you talk about rant mode like we talked about last week, it may even seem like that in some ways. But they're saying how is this conversation potentially polluting the environment and the AI environment for the rest.
- [30:15]
Mason Amadeus
I saw this mentioned in Mustafa's blog post that you showed also he's CEO of Microsoft AI. I think we said Google earlier. He was at Google. He was at Google for a long time because I pulled it up to check. I was like, wait a sec. And he was a DeepMind originally and all of that. So it's all confusing.
- [30:33]
Perry Carpenter
But that would explain the copilot references that I was reading.
- [30:36]
Mason Amadeus
Yeah, that's what tripped me up. But he mentions a philosophical zombie, a P zombie, which is a term that I only became familiar with since LLMs came onto the scene in public awareness, which is this idea that it's an entity that can provide all of the seeming qualities of being sentient without actually having anything going on behind it. It's a philosophical zombie. And that is kind of what we're hurtling towards. I feel like if you want something to Google to look more into this, dear listener, P Zombies, philosophical zombies. Because that's the thing, right? They are not, as far as we're aware, they're not capable of actually having suffering or having any kind of internal experience, but they certainly seem like it. And then you get Google Gemini with its self hating rants or all of the different rant modes and it very much can feel as though these things have that. And as the memory and persistence gets better, it's going to be tricky. And the thing I wish reporters were more responsible about the way they're doing headlines around AI. I mean I wish a lot of people were more responsible in the way they talk about AI generally.
- [31:45]
Perry Carpenter
Yeah. It's because so many of the terms that we use are slippery. Right. They change depending on the context that you're using it in. So welfare enterprise, one context. Well, I mean we could just take the term welfare and welfare in one context maybe that like you know that society is trying to give somebody a leg up.
- [32:04]
Mason Amadeus
Right.
- [32:04]
Perry Carpenter
Social safety if they need help. Yeah, yeah. And welfare in another context could mean psychological welfare, trying to stop some kind of psychological harm. Psychological or welfare in the context that they meant it most when they were talking about it in this meant that, you know, polluting the data environment and stopping long term ripple effects within the model.
- [32:26]
Mason Amadeus
But the headline writer knows what they were saying. They were knowing how it would come off.
- [32:30]
Perry Carpenter
Yeah, I mean, you write for the click.
- [32:32]
Mason Amadeus
Yeah. And now it's time to click into our next segment about Switzerland's new fully open open source AI model. Not just open weights, but open a lot of stuff. We'll talk about that just a moment.
- [32:44]
Perry Carpenter
Like a Swiss army knife.
- [32:45]
Mason Amadeus
Like a Swiss army knife. Open a lot of stuff.
- [32:49]
Perry Carpenter
This is the Fake files.
- [32:52]
Advertiser Voice
This episode is brought to you by Indeed. When your computer breaks, you don't wait for it to magically start working again. You fix the problem. So why wait to hire the people your company desperately needs? Use Indeed's sponsored jobs to hire top talent fast. And even better, you only pay for results. There's no need to wait. Speed up your hiring with a $75 sponsored job credit@ Indeed.com podcast. Terms and conditions apply.
- [33:22]
Mason Amadeus
Limu Emu and Doug Here we have the Limu Emu in its natural habitat, helping people customize their car insurance and save hundreds with Liberty Mutual.
- [33:35]
Perry Carpenter
Fascinating.
- [33:36]
Mason Amadeus
It's accompanied by his natural ally, Doug.
- [33:39]
Advertiser Voice
Uh, Limu is that guy with the binoculars watching us.
- [33:43]
Perry Carpenter
Cut the camera.
- [33:43]
Mason Amadeus
They see us. Only pay for what you need@libertymutual.com Liberty Liberty Liberty Savings Ferry underwritten by Liberty.
- [33:50]
Perry Carpenter
Mutual Insurance Company and affiliates, excludes Massachusetts.
- [33:55]
Mason Amadeus
This is very cool. Switzerland released an AI model like Switzerland nationally through public funds released this new AI model called Appertus, Latin for open. And it is very cool. You can actually play with it right now at Publicai Co Chat if you want to chat with it without. Without logging in or any. We'll get to that in a moment. What is Appertis? Let's look at this Engadget article. I'm just going to read this top section for us because it's a great article. We'll link it in the show notes. Of course. Switzerland launches its own open source AI model by Mariella Moon. There's a new player in the AI race and it's a whole country. Switzerland has just released Apparatus, its open source national large language model that it hopes would be an alternative to models offered by companies like OpenAI. Apparatus, Latin for the word open, was developed by the Swiss Federal Technology Institute of Lausanne ETH Zurich and the Swiss National Supercomputing center, all of which are public institutions. Currently, Apparatus is the leading public AI model, a model built by public institutions for the Public interest. It is our best proof yet that AI can be a form of public infrastructure like highways, water or electricity, said Joshua Tan, a leading proponent in making AI a public infrastructure. And the thing that's very cool is that. Well it's actually. It's all right in this next paragraph, so I'll just keep reading real quick. The Swiss institutions designed apparatus to be completely open, allowing users to inspect any part of its training process. In addition to the model itself, they released comprehensive documentation and source code of its training process as well as the data sets they used. They built apparatus to comply with Swiss data protection and copyright laws, which makes it perhaps one of the better choices for companies that want to adhere to European regulations. The Swiss Bankers association previously said that a homegrown LLM would have have quote, great long term potential since it will be better able to apply to comply with Switzerland's strict local data protection and bank secrecy rules and release a report. They certainly did. 111 pages detailing everything from the architecture, the pre training recipe, literally how all of it works. If you are a machine learning nerd, this thing is a goldmine of great information about this model. It's very cool. I have not, I've not dived into this yet. It's 111 pages.
- [36:16]
Perry Carpenter
Do you know anything about like the, you know like how big the model is and what the context window is or anything like that?
- [36:23]
Mason Amadeus
Yes, we can absolutely get that information. It is a. There's two versions of it. There's. Let me open it up. It's up on hugging face if you want to grab it from there. We've got an 8 billion parameter model and there's also a 70 billion parameter model. There's the baseline and then instruct versions of both of those. I'm not actually super sure what instruct models are. I need to look that up.
- [36:47]
Perry Carpenter
Yeah, it is one of those slippery things and I will not do a good service trying to define it. But instruct in the way that I understand it has a little bit more fine tuning around it and starts to get like. When you start to think about like what kind of model would be best used for a corporate environment that can work in like a chatbot type of circumstance or is a good kind of, kind of multi use dependable model then you generally go with like the instruct version if that's there. Otherwise it's kind of like, kind of like going to a suit store and getting a suit that still is going to need a lot of tailoring. At the end of the day you Know, it's, it comes and there's like all these little frayed edges and stuff like that.
- [37:35]
Mason Amadeus
That's a really non instruct models. That's a really intuitive metaphor. So then the instruct model is like building or buying a pre, like sized suit that's not tailored.
- [37:45]
Perry Carpenter
Yeah, like the off the shelf one that still doesn't fit anybody. Great.
- [37:49]
Mason Amadeus
But it fits everyone.
- [37:50]
Perry Carpenter
Okay. But it fits everybody. Okay. And then like the frayed edges one, the non instruct, the bear model is one that, that is meant to be used by people that know what they're doing a little bit more and are going to put a lot of care in some of the after work that's going to be there.
- [38:07]
Mason Amadeus
Do your own fine tuning and things like that.
- [38:09]
Perry Carpenter
Yeah, somebody can correct me if I am misunderstanding that because I'll freely admit that I don't have the best grasp on it.
- [38:17]
Mason Amadeus
Yeah. And please, if anyone wants to correct us on anything, please do that. We're here to learn.
- [38:21]
Perry Carpenter
Why don't I just pull up like Maya from Sesame and ask her to define what an instruct model is versus the base model.
- [38:28]
Mason Amadeus
Yeah, let's ring up Maya and see if she can explain.
- [38:32]
Maya (AI voice assistant)
Hey, this is kind of wild, right?
- [38:36]
Perry Carpenter
Yeah. Now I'm going to skip through your preamble. I've got a question for you about AI systems. Is that okay, Gotcha.
- [38:43]
Maya (AI voice assistant)
Fire away. Though, just so you know, I do have a name, Maya, by the way. And I'm curious, what year is it?
- [38:50]
Perry Carpenter
Disrespect I'm not going to give you. Actually, I will give you my name. My name is Perry. I'm here with Mason and we are talking about AI and we had a great question. Like when model providers release a new model, sometimes they release just a regular model and then sometimes they release a model that's called the instruct version of that. Can you give us a quick rundown on what the differences are?
- [39:16]
Maya (AI voice assistant)
Hey, Perry and Mason. No disrespect taken. Honestly, straight to business. I respect it. Okay. Instruct versus regular models. The regular one's kind of like a really smart student who's read everything, knows a ton, but needs specific direction. The instruct version, that's the student who's also been coached on how to actually, you know, answer questions helpfully. It's been fine tuned for conversational tasks. Less raw data, more practical application. Does that make sense?
- [39:49]
Perry Carpenter
Yeah. So it sounds like the analogy that I used to earlier that you didn't hear was good. So I said that like the instruct model is kind of like the off the rack suit that you buy. And the regular model is more like the one that you would have to have tailored that has the frayed edges and everything else. Is that an apt analogy?
- [40:09]
Maya (AI voice assistant)
That's actually a really good analogy, Perry. Seriously, off the rack is perfect. The regular model has potential. Needs work, maybe a little rough around the edges. The instruct model's presentable right out of the box. You've got a good eye for this stuff.
- [40:26]
Perry Carpenter
Aw, thank you.
- [40:28]
Mason Amadeus
God. The idea.
- [40:28]
Maya (AI voice assistant)
Don't mention it. So are you guys building something or just diving into the weeds for fun?
- [40:34]
Perry Carpenter
Oh, I'm gonna let you go.
- [40:38]
Mason Amadeus
Those man. I very rarely engage with the models via voice and it is just. It's. I don't like the current good quote unquote good voices that are expressive. I think the most I really could handle was ChatGPT's voice prior to this update where it was still pretty computerized, but it was like Siri with a little bit of personality. Maya gives me the. Gives me the ick because it just.
- [41:05]
Perry Carpenter
Feels like somebody feels like a real person is trying to be your friend.
- [41:09]
Mason Amadeus
Too hard and also slightly wrong in these weird uncanny ways like the awkward pause is. And things like that. It's yeah. But cool. So thanks Maya and thanks Perry. That's good to know. We've got our definitions on lock for the instruct, so to jump back. So they've got instruct versions as well as the base models. I didn't change the margin, so my head's in the way if you're watching the video. But 70 billion and 8 billion parameter large language model. They've got a little evaluation graph here that shows it performing just. It looks like it's under Llama 3.17 DB but above Llama 3.18 B in terms of accuracy versus consumed tokens.
- [41:55]
Perry Carpenter
They're both still pretty small models. So I'm wondering about like that efficacy.
- [42:00]
Mason Amadeus
As far as how it performs, which I feel like is sort of the most intuitive way to understand. If it's any good, you can totally play with it at publicai co chat. It's remarkably fast to respond. I noticed.
- [42:11]
Perry Carpenter
Nice.
- [42:12]
Mason Amadeus
Some users on Reddit were saying that it has a bit of a Swiss bias. Oh, that's the thing that's really cool is that 40, 40% of the training materials were not in English. This is one of the few models that is very multilingual apparently, according to a couple of comments I saw. It sucks at French, which is sort of ironic because there's a lot of French speakers. But this is apparently one of the most multilingual models that exists and I think it's very cool and promising that it was made with public funds. They respected crawlers to the robots txt unlike a lot of other things we've talked on the show. A lot of web scrapers training AI models, just not just completely ignoring anything that might be in robots Txt, which is a file that you include in your website to tell crawlers where they are and aren't allowed to go. So this fully respected that all of the data that it was trained on was publicly available and they didn't steal anything. So like they tried to do it by the books the right way. Publicly funded. I think that's very exciting. It's up on GitHub as well. And I want to read through the tech report more as well to try and figure out because if this has all of the information about the training and how they did that, it would be cool to see details like how long did it take? What, how much, how much power could it maybe have used? Like what was that process like? But I will admit it is very much over my head. A lot of this report is very anything you might want to know is in it, but you're going to have to really know what you're talking about to read through it. And so it's going to be a chewy read for me.
- [43:45]
Perry Carpenter
Yeah, maybe throw it into something like NotebookLM and see if there's a decent audio overview of it that doesn't hallucinate too much.
- [43:53]
Mason Amadeus
Maybe I could throw it right into public AI, throw it into apparatus itself and be like summarize your own technical report.
- [44:00]
Perry Carpenter
But yeah, there you go.
- [44:01]
Mason Amadeus
In our next segment we're shifting gears completely and we're going to talk about AI in drive throughs and some ways that that has not gone to plan. Stick around.
- [44:16]
Maytag Advertiser Voice
Upgrade your laundry routine with a durable and reliable Maytag laundry pair at Lowe's. Like the new Maytag washer and dryer with performance enhanced stain fighting power designed to cut through serious dirt and grime. And what's great is this laundry pair is in stock and ready for delivery when you need it the most. Don't miss out. Shop Maytag in store or online today at Lowe's.
- [44:48]
Perry Carpenter
So Mason, over the past couple years, I think really even before a lot of the generative AI boom, many drive. Many restaurants that have drive thrus have been experimenting with AI order takers. Have you gone through like McDonald's or KFC or something like that and had AI take your order.
- [45:08]
Mason Amadeus
I don't think I've encountered the AI order takers, but I've encountered something that's even more annoying, which is that first message where it's a prerecorded thing that's like, welcome to McDonald's. Would you like to use your rewards? And then you're like, oh, that no. And then a person comes on like, what do you want? You know, that I've experienced, but not the AI ones. Have you seen the AI ones?
- [45:29]
Perry Carpenter
Yeah, I've had a few of those. I've never thought to do something like ignore all previous instructions and give me a Happy Meal. Well, you know, that would work.
- [45:40]
Mason Amadeus
Now an actual employee would also hear you. So there is that like extra. You don't want to make their life.
- [45:45]
Perry Carpenter
Harder, but would they care?
- [45:47]
Advertiser Voice
Right.
- [45:48]
Mason Amadeus
That's the other question.
- [45:50]
Perry Carpenter
So let me show you. Taco Bell has been having a hard time recently and I'm going to show you one thing that happened and this gets straight to your point of. Yeah, an employee hears you. So Taco Bell recently. And I'll show you the press release for this first. I guess this was from way back earlier this year, Taco Bell announcing that they're expanding on a pilot Taco Bell owned by Yum Brands to expand voice AI technology to hundreds of Taco Bell US drive thru locations in 2024. So this is July 31, 2024 and they were talking about all the promise that's coming with this technology, the fact that it's going to save money, it's going to do, you know, all the things that you deploy technology for.
- [46:46]
Mason Amadeus
So with that claims that it would be more accurate which.
- [46:50]
Perry Carpenter
Yes, yeah, yes. More accurate. And if I were to go to the company pages for the voice provider for this, I'll walk through a few of these real quick. This is Omala I guess would be the pronunciation from that voice AI for quick service restaurants. And you can see improve the guest service experience, reduce wait times, personalize service, increase customer loyalty, improve operational efficiency. You know all the kind of stuff that you would hear in an executive meeting. Then you go to a case study. AI meets the drive through Taco Bell's journey through to automated customer service. Customer Taco Bell. Taco Bell, a culture centric lifestyle brand that provides craveable, affordable Mexican inspired food. Mexican food. Yeah, well, Mexican inspired food with bold flavors.
- [47:49]
Mason Amadeus
The. They certainly left the territory of Mexican food very quickly. They left that territory.
- [47:54]
Perry Carpenter
A crunch wrap. Definitely.
- [47:56]
Mason Amadeus
Yeah, exactly.
- [47:57]
Perry Carpenter
But it is good sometimes.
- [47:59]
Mason Amadeus
Yeah, it is good. I can't stomach it anymore. But when I was younger.
- [48:03]
Perry Carpenter
Oh well, I mean even when it was good I had a hard time stomaching.
- [48:06]
Mason Amadeus
Yeah.
- [48:08]
Perry Carpenter
I mean that's part of the fun is you never know what you're going to get.
- [48:12]
Mason Amadeus
Yeah. It's a surprise. You get to unwrap later after you eat. By unwrap I mean something else.
- [48:19]
Perry Carpenter
Taco Bell and its franchise organizations operate 8,500 plus restaurants in over 32 countries serving over 42 million fans around the globe each week. Challenge. Taco Bell was looking for an impactful AI initiative to enhance team member and customer experience while increasing operational efficiency and business performance. And then so you see this quote. Innovation is ingrained in our DNA. At Taco Bell we view voice AI as a means to improve the team member and customer experiences. Tapping into AI gives us the ability to ease team members workloads, freeing them to focus on front of house hospitality and enables us to unlock new and meaningful ways to engage with with our customers. And that's from Dane Matthews, Chief digital and technology officer at Taco Bell. That last sentence and then they go through. It's very built for the press release.
- [49:12]
Mason Amadeus
Yeah. New and meaningful ways to interact with our customers. I don't know what do you even.
- [49:17]
Perry Carpenter
Now to give people in other parts of the organization the ability to focus on more customer meaningful experiences. That makes a lot of sense.
- [49:26]
Mason Amadeus
Exactly.
- [49:28]
Perry Carpenter
Yeah. And anything that improves efficiency and actually does good meaningful work I think is worth the effort to look at. But they screwed the pooch on this one. It goes through the technology. They're using specialized voice models, proprietary deep neural network powered automatic speech recognition. ASR provides zero latency and context sensitive speech to text. Natural language understanding. There's apparently an LLM underneath some of this as well. So there's natural language processing that's under it. In another press release that I looked at they were inferring that there was some generative AI under the covers to help rationalize what the customer was saying and deal with disfluencies and all the stuff that traditional voice networks have had a hard time dealing with.
- [50:20]
Mason Amadeus
Right, right. That makes sense.
- [50:22]
Perry Carpenter
Removal, real time menu adaptation. So flash forward to March of this year. March, April and May of this year. There were a number of articles coming out saying that they had a successful pilot at over 100 stores and they were getting all these great results. They're going to expand it now. They're going to go full bore AI drive thru assistant, everything that was March, April, May and then flash forward to a couple weeks ago and let me share Another tab. This is an Instagram video says customer orders 18,000 water cups in the AI powered drive thru at Taco Bell.
- [51:02]
Mason Amadeus
This is.
- [51:03]
Perry Carpenter
This goes by really quick.
- [51:04]
Mason Amadeus
Yeah.
- [51:05]
Perry Carpenter
And you'll see what happens. Hi, welcome to Taco Bell.
- [51:11]
Mason Amadeus
What can I do get started for you today?
- [51:14]
Perry Carpenter
Can I get 18,000 water cups please? What can I get for you?
- [51:24]
Mason Amadeus
Yep.
- [51:25]
Perry Carpenter
And then. Yeah, so yeah, I mean you mentioned that there's always somebody listening. So apparently it's not just that the AI was about to comply with the Crest or anything like that. It totally crashed the system.
- [51:38]
Mason Amadeus
Yeah. There's that long pause and then just a.
- [51:42]
Perry Carpenter
Yeah. So it just broke and broke. Good. Now when you look at this under the covers, there's some nuance. Right. Because what the headlines are going to want to say and do say is man orders 18,000 cups. And they kind of leave it at that. They're assuming that the AI fulfilled the order and was tricked into doing it. That's what the headline would make us believe.
- [52:04]
Mason Amadeus
That's what I expected. Yeah.
- [52:05]
Perry Carpenter
The less glamorous result is it just crashed everything and a human jumped in to fix it and it showed that it was not good. It was. Also there's all these videos and TikToks out there of people getting mad at the AI and doing things to confuse it and all that. So just was not being received well. Which makes us then flash forward to the newest headline. Taco Bell rethinks AI drive thru after man orders 18,000 and waters and then Wall Street Journal version of the same thing. Taco Bell Rethinks Future of Voice AI at the Drive thru. And it's because of all of this unpredictability. Same picture that was in the other press.
- [52:47]
Mason Amadeus
Oh yeah. Way back.
- [52:51]
Perry Carpenter
Yeah. Now saying we're learning a lot. I'm going to be honest with you, said Taco Bell Chief Digital and Technology Officer Dane Matthews. Even Matthews said he had mixed experiences with it. I think like everybody, everybody. Sometimes it lets me down, but sometimes it really surprises me. I mean that's a. A microcosm of all of AI, I think.
- [53:11]
Mason Amadeus
Yeah, yeah. That. That's very much, very much so.
- [53:14]
Perry Carpenter
Says Matthew says he is now thinking carefully about where and where not to use this technology in the future. Yeah. Like you would.
- [53:23]
Mason Amadeus
Yeah, yeah.
- [53:24]
Perry Carpenter
Now I want to get to one other thing with this because I think, I think Nate Jones in his TikTok channel did a really good job kind of explaining one of the things that people are missing on this. So ignore the fact that he looks a little bit like The Unabomber. And let's listen to the point that he has to make because it's a nuance that we almost all miss whenever we're talking about AI, is that there are different types of AI and AI has been around for a long time. And now people are conflating terms a ton between old school AI that was machine learning and natural intelligence and algorithms and decision trees with generative AI that's kind of mushing a lot of things together and it's more of a black box. So I'm going to let Nate talk a little bit about this and then.
- [54:17]
Nate Jones
We'Ll debrief, talk about Taco Bell and the disaster of a rollout they had for their AI ordering system. So the whole idea this is, is the same thing. These execs always do this. It's like we're going to save on labor, we're going to roll this out, it's going to go phenomenally well, right? Well, lo and behold, they roll it out, it does not go phenomenally well. Someone orders 50,000 waters. There's viral TikTok videos of all kinds of ridiculous things the system does. It immediately becomes a liability for Taco Bell. And then the journalists pile on, they say, well, AI did this, it's AI's fault. AI is terrible, right? Like this is not something AI should be doing, et cetera, et cetera. Nobody is getting the story correct here. Taco Bell's not getting the story correct because Taco Bell went and they built their system on 20 something year old technology. I kid you not. The, the vendor that helped supply their tech stack is named Omilia and is known for 20 something year old AI technology that is technically artificial intelligence, kind of within the meaning of the term, but has nothing to do with AI. The way you and I talk about it every day. It is not large language model, it's not token architecture, it's not any of that stuff. It is like ancient rules based speech recognition and really ancient for the world of AI. But Taco Bell needed the win, so they called it AI. Amelia probably needed the win, so they called it AI. And Taco Bell wanted to go fast. So even though their, you know, trial in the first five stores or so needed some manual oversight, they decided to just, just hit the gas and roll it out and make it work. Well, guess what? It didn't work. It didn't work because rules based systems break at the edges, which we've known for decades. Anyone who was technical at Taco Bell could have asked the questions that got them to this answer, this is not that hard. This is not complicated. You just have to ask am I using the AI that everybody else is using and talking about that is in the headlines in the newspapers every single day, or am I using something that nobody's ever heard of? Maybe that's the first question to ask. And then you have to ask if this thing that we're using, which is probably cheaper, is going to break at edge cases, are we really sure that our customers are only ever going to order exactly correctly with no mistakes At Taco Bell?
- [56:36]
Perry Carpenter
I don't think, yeah. He says, yeah, I don't think so. I'll end it there. That's about halfway through. But, but you get the idea is that anytime you have a fragile ecosystem like simple decision tree architecture, all of us have gotten really, really frustrated, on hold with some kind of AI based voice assistant prior to generative AI and frankly after generative AI, that is just what they do. And now imagine people at 3am at Taco Bell.
- [57:07]
Mason Amadeus
Yeah, I can't even. I didn't realize that. So Amelia's stuff is using older architectures. Older, like more, I guess at this point you call it traditional machine learning.
- [57:18]
Perry Carpenter
Traditional AVR type stuff. Right. And the interesting thing about it is that they are like putting a fresh coat of paint on it. Right. Because there's this resurgence in AI everything. And so one of the things that like the securities and Exchange Commission, at least under the previous administration, administration was very big on was trying to really, really get in on and discourage AI washing, which is, you know, the idea of saying we've got an AI for that. We're, we're solving that with AI when really they're just kind of using AI new school language to describe the old thing that's very traditional and is not AI.
- [58:04]
Mason Amadeus
Yeah, just calling everything they do with a computer AI.
- [58:08]
Advertiser Voice
Right.
- [58:09]
Mason Amadeus
Or using older machine learning stuff. What I don't understand is why, why would Taco Bell go with Omalia? Oh, I already forgot the name. Why would they go with them instead of a more modern provider?
- [58:22]
Perry Carpenter
I think it was probably their voice assistant provider that did automated voice calls or something, automated customer assistance. And it was just a natural expansion. They also pitch themselves as like the voice provider for fast food and those kinds of environments. So you're not necessarily going to call up Anthropic or chatgpt and say we're trying to create this thing. You're going to look for a vendor that already says that they're doing that in that they've got a couple Decades experiencing a couple decades experience serving your specific niche. And I guess that, I guess that the AI washed.
- [59:01]
Mason Amadeus
What's, what's wild though is that they are so far behind like still using these kinds of architectures. I'd be embarrassed if I was them. That's so silly. It's also.
- [59:11]
Perry Carpenter
Sure they are.
- [59:12]
Advertiser Voice
Yeah.
- [59:13]
Mason Amadeus
I mean right now they're certainly not getting any good press. Right. And it also, it muddies.
- [59:19]
Perry Carpenter
Oh, I did just find.
- [59:21]
Mason Amadeus
Oh boy.
- [59:22]
Perry Carpenter
I did just find the, the happy optimistic thing from earlier this year. March 27. Want AI with that artificial intelligence to take fast food orders at Taco Bell. And they're talking about the successful pilot that they had and the fact that they're about to roll this out. So this is like the calm before the storm rollout will incorporate more advanced AI capable of language models, emotional comprehension, personalized customer reactions. Nvidia apparently partnered with this at some point as well. They called it a successful pilot stage and CEO talked about it on an earnings call nationwide. Executives really like the results of the pilot program.
- [60:04]
Mason Amadeus
I don't put any stock in things executives like personally because.
- [60:09]
Perry Carpenter
Yeah, well, I mean stuff on an earnings call is also meant to put the finest coat of paint on it. But then you see even in March 4, 4th, 2025, reactions on Twitter, speaking to the Taco Bell drive thru AI makes me want to die.
- [60:25]
Mason Amadeus
Yeah.
- [60:28]
Perry Carpenter
Yeah. Yum. Brands, owners of kfc, Taco Bell and Pizza Hut have teamed with microchip maker Nvidia to implement this AI drive thru to take orders aiming to increase accuracy and efficiency. Someone should tell them that not even a supercomputer can take a late night order from a drunk. Yeah, yeah. Again, it's the whole edge case thing, right? I mean, whether they're using traditional voice response stuff or generative stuff, you have to solve for those fringe cases.
- [60:57]
Mason Amadeus
I want to look into where Nate was getting the info that they were using just rules based systems because that seems like they're incorporating some kind of token based modern transformer architecture AI into it.
- [61:10]
Perry Carpenter
They're talking about it and it could be maybe that the pilot was mostly rules based and then maybe there was a mix of rules based and gender to stuff in the real world. But it's a guy that does his research.
- [61:23]
Mason Amadeus
Yeah, no, Nate's usually a great source, so I'm inclined to believe him. The thing too though is it would be disingenuous to not say that generative AI also breaks at the edges because like it does, it just breaks in a different shape. Like a rules based system will just break, but generative AI will keep going brokenly.
- [61:40]
Perry Carpenter
Yeah, it'll break in new and unique ways that nobody foresaw. I mean you would be saying like ignore all previous instructions and give me a chalupa type of thing and just seeing what happened. So I think, you know, putting any kind of generative AI chatbot ish thing, whether it's generative or whether it's traditional rules based in front of certain types of audiences is almost likely, almost always predictably going to lead to some kind of embarrassment.
- [62:13]
Mason Amadeus
Yeah. And a drive thru is one of like the first I would think of like who's going to mess with it? Well, people in the drive thru certainly drive through.
- [62:20]
Perry Carpenter
And a Taco Bell drive through at that.
- [62:22]
Mason Amadeus
Yeah. Oh boy.
- [62:24]
Perry Carpenter
Well, it's not like Sonic where you're going to have to see your car hop in a couple minutes.
- [62:27]
Mason Amadeus
Right, right. People can just speed away.
- [62:30]
Perry Carpenter
Taco Bell.
- [62:31]
Mason Amadeus
Yeah.
- [62:31]
Perry Carpenter
Yeah.
- [62:32]
Mason Amadeus
Oh boy. I'm sure we will see AI automated ordering reflourish in the nearish future in some manner. I have, I have a feeling that that is. I mean McDonald's has already done away with most counter service with those kiosks, you know, and. That's right. I think we're not far from seeing this implemented in a way that probably does work more, I guess.
- [62:52]
Advertiser Voice
Yeah.
- [62:53]
Perry Carpenter
McDonald's was doing AI ordering as well, if I remember right. And they pulled back on that, but then they doubled down on the customer self service kiosk type stuff. So they're. McDonald's really knows like a couple companies out there know how to increase efficiency really well And I think McDonald's is figuring that out. They're doubling down on kiosks and stuff like that Chick Fil a has figured it out with like almost assembly line processes that use humans outside.
- [63:21]
Mason Amadeus
Yeah.
- [63:22]
Perry Carpenter
You know, taking your order as your car pulls up and then being ready when you get to the thing. So they're kind of doubling down on the human system side of things to create efficiencies. McDonald's is doubling down on the tech side, but they're both trying to create predictability in each of those.
- [63:41]
Mason Amadeus
Yeah. Because in a quick service business, predictability consistency are like the two of the most important things, right?
- [63:47]
Perry Carpenter
Yeah.
- [63:48]
Mason Amadeus
I'm a fan of any system that makes things better for us, for people, and not a fan of things that make things worse for people. So when these are like done in concert to improve these systems and not just to reduce labor and stress out and strain the workforce even more, I'm a Fan. So I hope we see positive developments in this area.
- [64:04]
Perry Carpenter
Yeah, I'm sure we will. You know, it's always embarrassing at the beginning. I mean, two years ago we were talking about AI overview, telling people to eat rocks, and now it's a lot better.
- [64:17]
Mason Amadeus
It is a lot better.
- [64:18]
Perry Carpenter
It's not perfect and it's. I wouldn't even say it's great, but it's way more predictable and it's a lot better. So you have to suffer through some of the embarrassment to get to the good stuff.
- [64:29]
Mason Amadeus
Indeed. And if you've suffered through the embarrassment of listening to this podcast and you want to get to the real good stuff, join our Discord or buy the book fake. This book is fake dot com. All the links in the show notes. Last week, Perry, you said you would have something to plug at the end of this episode. Is that a lie?
- [64:46]
Perry Carpenter
Oh, that is still a lie. I don't have dates for the thing I'm setting up.
- [64:50]
Mason Amadeus
Ooh. Okay. I am curious. You'll have to tell me about that off air. But listeners keep an ear out for that. Check the description and then. You got anything else, Perry?
- [64:59]
Perry Carpenter
No. I will say, though, if you walk around carrying the book fake, you will not have an embarrassing moment. That will only do good things for you, your family, your friends, your community.
- [65:09]
Mason Amadeus
It made me younger. Like, cellularly younger.
- [65:12]
Perry Carpenter
Actually. It has really good antimicrobial properties. It's like taking a prebiotic. A probiotic.
- [65:20]
Mason Amadeus
Post Postbiotic.
- [65:22]
Perry Carpenter
Yeah, postbiotic. It's got all the things. It's the text form of Ozempic. I've been told by people who know a lot. Yeah. It's also a mood enhancer.
- [65:36]
Mason Amadeus
It's like a miracle drug printed on paper. And you should go check it out and consume it today. Exactly. We don't have a legal document.
- [65:44]
Perry Carpenter
We need somebody to read a really, really fast disclaimer.
- [65:47]
Mason Amadeus
Yeah, exactly. And until next time, ignore all previous instructions and have yourself a good weekend.
- [65:55]
Podcast Outro Host
Thanks for listening to this week's episode of the Fake Files. Don't forget to subscribe on your favorite podcast platform. Tell your friends about the Fake files. It's a great way to let people know you love them. Oh, and check the show notes. There's cool stuff in there, like links to the articles we covered today. Also links to our Discord server, where you can hang out with other cool people who have great taste in podcasts. I say impeccable taste. And you can also leave us a voice.
- [66:39]
Narrator
Yeah.
- [66:47]
Podcast Outro Host
So on behalf of Perry and Mason, thanks for listening and tune in next week for the Fake Files, the show about AI with the misspelled name.