Tracy Alloway: Hello and welcome to another episode of the Odd Thoughts podcast. I'm Tracy Alloway.

Joe Weisenthal: and I'm Joe Wisenthal.

Tracy Alloway: Joe, you know I have some uh prepper tendencies. Yeah, this

Joe Weisenthal: slightly prepper tendencies. Prepper adjacent.

Tracy Alloway: I know you do because my plan for when everything when everything goes bad is to bring my family over to your place. So, I'm relying on you actually.

Joe Weisenthal: That's fine. I actually figured that and I've been building an extra store of supplies.

Tracy Alloway: I'm going to like send you a whole list of things my kids like to eat and stuff like that just so that we're

Joe Weisenthal: just so we're all ready. Yeah.

Tracy Alloway: Okay. Well, one of the things I saw on a bunch of the prepper boards that I sometimes look at, I don't want people to think that I'm crazy about it, but I find it interesting. I I find it interesting seeing how people's like insecurities manifest in physical stuff. But anyway,

Joe Weisenthal: one of the things everyone was saying was you need to start taking cash out because of the situation in Iran because we're all expecting a big cyber attack that's going to absolutely destroy the US financial infrastructure. By the way, if I told you my idea for a business of like I've looked at prepper meals, like prepared meals and they all look terrible, like a slightly high-end version for yuppies, I think would be really good. Like something that you know, like some nice

Tracy Alloway: I think I think it's a physical limitation on how good you can actually get like dry food.

Joe Weisenthal: Science can do a lot of things these days. Anyway, freeze dry.

Tracy Alloway: Let's talk about the actual issue at hand.

Joe Weisenthal: Yeah. Well, are you taking cash out silver?

Tracy Alloway: I haven't yet. I I'm relying on my store of gold and silver. That's right. But I think this raises a legitimate and actually very interesting topic, which is what do we know about Iran's cyber, I guess, facilities, skills, what could happen in this context? And then also

Joe Weisenthal: everything that's going on with the world of AI, right? Like cyber security, cyber hacking,

Matt Suiche: Totally. I mean, also just within the context of the war itself, setting aside hypothetical doom scenarios, there's a really interesting report in the Financial Times about Israel having been able to hack into all of the traffic lights in Tehran. Yeah. Almost unbelievable and shocking, but there's already within the war itself or even over the last couple of years, there was the pager attack that uh Israel had executed. And so, yes, cyber is part of it. And the timing is wild here because speaking of AI, it was just on Friday, we're recording this March 5th. I'm not exactly sure the date is coming out, but a week ago basically there was the news, the complete collapse of the Anthropic relationship with the Department of Defense or the Department of War. And so it's all in the mix right now. And how is AI actually going to change warfare? And what are the national security implications of AI and AI and hacking? There is a lot in this sort of mix that's all happening right now.

Joe Weisenthal: Absolutely. The thing that really caught my eye was the story about a hacker using Claude to hack into like the Mexican government system. Did you see that?

Matt Suiche: That was really interesting because it seemed like the hacker extracted a bunch of information from Claude itself. You know, I'm pretty sure you cannot go to Claude Code and say like, I want to break into the Mexican government website. Help me like build this app. It won't do that. It's trained to avoid malicious uses. But people find a way to jailbreak them. people find a way to sort of extract information from the AI itself that has in its training and so forth. And there's been examples of leaks where you know people upload data to the AI and somehow other people see it. Anyway, there's a lot here that we have to learn more about.

Tracy Alloway: We should talk about all of it and we do in fact have the perfect guest, someone who's been on the podcast before, but it's been a while. We're going to be speaking with Matt Swish. He is the founder of DB, which is a data infrastructure startup for Aentic AI. So honestly the perfect and a legendary hacker.

Joe Weisenthal: A legendary French hacker. I should have said that first. Matt, thank you so much for coming back on OddLots.

Matt Suiche: Thank you very much. It's been a while now. I think it's been what, four years?

Tracy Alloway: Yeah, I think it has been.

Joe Weisenthal: The last time we spoke to you, you were still in Dubai and now you're coming to us from Sweden and a very Gustavian looking background over there.

Matt Suiche: I think actually when we talked was right after Russia's invasion of Ukraine. So I guess yeah. Wow. That has been almost

Joe Weisenthal: every time there's a war we call you Matt. But because war is so intermixed with cyber espionage, cyber security, hacking and so forth, it's a natural natural time.

Tracy Alloway: So for the benefit of people who didn't listen to the episode four years ago, can you just give some context around who you actually are and your sort of history in the hacking community including you know shadow brokers and the want to cry era and all that stuff.

Matt Suiche: So I've been in enterprise software for almost like 20 years particularly cyber security and my name appeared in a few of the different leagues because of various analysis that I've done of like private information that was being leaked but also a lot of attacks that happened that happened to target critical infrastructure over the last 10 years and last time we were on the podcast uh one of the thing we talked about is does cyber really matter? Once you enter into a kinetic war, which is exactly what's happening now and the main takeaway was once you start using missiles, most of these cyber elements are not really relevant because you would use cyber mostly to gather information and intelligence to prepare an attack or to disorganize an enemy or create confusion. But as we have seen now, you can use like drones that are like $20,000 and create more chaos that you would do with any sort of exploits.

Joe Weisenthal: You know, I like how you're introduced as a legendary hacker and then you're like, "Oh, I've been 20 years in enterprise software." I feel like this is like the uh the Winnie the Poo meme. It's like, you know, that casual and then fancy dress up

Tracy Alloway: hacker and then below it 20 years in uh enterprise software. But this is a really interesting point that you made this idea between okay mostly it sounds like when people imagine cyber attacks they imagine what Tracy talked about in the beginning suddenly the entire financial infrastructure like just come to a halt people worried about or there's going to be a blackout etc. But in reality or what we've seen so far by and large is that cyber in the context of war is still much more about data collection, espionage and so forth rather than these more like you know the types of things you might see in a movie.

Matt Suiche: Yeah, exactly. I mean over the last like 1015 years we have seen some like attacks like cyber attacks against critical infrastructure. Iran targeted like Aramco around like 2012. Mhm.

Joe Weisenthal: They were just mostly like using a what we call a wiper that was like a malware that was erasing the hard drive of most of the machines. And then we obviously have the case of Stuxnet few years before where it was a joint Israeli US operation against some of the nuclear centrals in Iran where some of the PLC's were targeted. But what we have seen over the weekend is some of the drones happen to target some of the Amazon data centers,

Tracy Alloway: right?

Matt Suiche: And that created so much instability because multiple of the zones have been down and I think two out of three and the third one is still recovering days after because most of companies either private companies or public companies now relying on the cloud which is something that was not really the case before. And once you have some sort of centralization in term of dependence, you also become an easy target. And most of governments, AI companies, cloud companies do not really have $20,000 drones in their threat models, which is like something that's pretty new, but also confirms that kinetic wars can have more impact. So I take the point about cyber being perhaps more useful before a war when it comes to info gathering and things like that. But we have seen some deployment of cyber attacks in the past week or so. So we know Israel is attacking some cyber infrastructure in Iran and we know that Iran has perhaps attempted some things maybe not as successfully but walk us through what we've actually seen so far. So so far we have seen an Israeli operation where one of the prayer app has been hijacked and some message was sent to the users. So it's more like to create confusion within people

Joe Weisenthal: also the traffic light operation to understand the position of some of the target but it's more used for like reconing and in term of destruction we didn't see anything significant even the government itself of Iran shut down most of the internet for a lot of the users and a lot of what we see on social media is the usual disinformation and misinformation campaign, especially now with AI, there's so much AI slop with like the videos, the text, the bots that's becoming pretty common now, even when there is no war. So, it's not really like really impactful. So, it's more like to create confusion than being actually destructive. And now, we're definitely entering in a stage where

Matt Suiche: it's been extremely uh destructive. And I cannot remember the last time we have seen so many countries being targeted which is pretty like a first I would say in term of like a war climate.

Joe Weisenthal: Can you talk about you know people stare at their screens all day and they fool themselves into thinking that they're quote monitoring the situation etc. But mostly

Tracy Alloway: is that projection Joe?

Joe Weisenthal: I don't delude myself. No, I like actually like I sort of look at my screen and I know that I'm being inundated with contextless garbage and slop and propaganda and so forth. I'm curious how you monitor the situation actually as someone who takes these topics seriously and doesn't just sort of become a overnight expert, you know, the day after bombings begin. Like how do you pay attention to what matters? How do you actually know what's real and so forth and avoid just sort of the delusion of staring at the screen and and engaging with slot?

Matt Suiche: It's a it's a good question because there's so much of it. So I think the default reaction is to ignore most of it. Yeah.

Tracy Alloway: Unless it becomes really significant.

Matt Suiche: In this case, I think it comes down to looking at the actual damage. Many people from the military world but also the intelligence community has been underestimating Iran capabilities. Exactly like people used to do with North Korea and now North Korea has some of the best hackers in the world when we see them like targeting financial institutions whereas before they would not do like much. So there is definitely like internal capabilities that are available but there is so much noise now like you say a lot of people are monitoring the situation

Joe Weisenthal: giving the quote unquote like overnight expert opinions

Matt Suiche: and that's becoming a lot of noise but I would say that in this particular case we have heard of the imminent threat of Iran for around 40 years

Tracy Alloway: right

Matt Suiche: and that's also not really a new situation So most of people would have context around it and even for like the attack that happened last weekend many people were expecting them for weeks especially as they are continuation of what happened last summer.

Joe Weisenthal: Can you actually talk a little bit more about the data center attack and um because that's not cyber really I mean that's just that's just a physic that's kinetic warfare against a data center. I was surprised how disruptive was that. I sort of would assume that cloud service providers that it's fairly liquid. Okay, one goes down but you know it can just be the same software can be run from numerous other clouds. But I saw that there were disruptions. I saw Fortnite tweeting about the fact that some of their gameplay was impaired due to the attack on data centers. How disruptive have those attacks been because this is of course a very, you know,

Tracy Alloway: this is where kinetic meets cyber.

Joe Weisenthal: Yeah. And there's a lot, you know, in future like thinking about hardening these data centers and as you say like making them they're going to be increasingly targets for war. Like how disruptive was that?

Matt Suiche: Very good question. So I I think one of uh the the main takeaway is that it has been extremely successful. So like we said before like a shah drone is around $20,000 and they managed to shut down two of the zones of Amazon. Actually even if you look at the official report from Amazon for like 36 hours they were just saying oh some objects struck the data centers before they actually explicitly said they were drone strikes. So a lot of services that have been using them have been targeted. So from like local applications from two banks because in a data center you are taking care of multiple different services

Tracy Alloway: right

Matt Suiche: and even Versel had to reroute uh their data to Bombay and to exclude uh like Middle East as deployment. So even if you take the cost of most of like zero day exploits that can go up to like multi-million dollar attacks, if you are really aiming at destructing things, the cost uh reward of using such an attack is uh really efficient. So you really enter into some sort of asymmetric conflict where you can just spend like some really old material and have way more impact than someone who's going to be like cutting edge and just trying to impress with like capabilities because at the end of the day it does not really matter.

Joe Weisenthal: How do governments actually build up their cyber capabilities nowadays? cuz I have this image in my head of maybe 10 or 20 years ago, you know, they would recruit like a 20-year-old such as yourself at the time and they would be working in a dark room, that sort of thing. But then

Tracy Alloway: drinking Red Bull.

Joe Weisenthal: Drinking Red Bull. That's right. But then, you know, we had the boom in Silicon Valley and so you had competition from private companies. Now we have the boom in AI and again even more competition from private companies and at the same time governments seem to be I guess seeding some of their own skill set to potentially private companies like Anthropic and ChatGpt and some others. Walk us through how I guess the development of governmental cyber capacity has actually shifted.

Matt Suiche: I mean something that didn't really change over the last like years in term of capabilities like I guess we all remember like the northern leaks in 2013 when we started to see more about the inside of like capabilities from a government including like domestic m surveillance global surveillance exploitation capabilities and since then every other year we have seen an history of data being leaked that belongs to uh the government. So in a way things have been changing a lot but not really much. Like most recently there was a contractor from Elfrey Aris that was uh sentenced to 87 months uh sentence because he happened to sell uh zero day exploits to a Russian broker.

Joe Weisenthal: and that's like actual exploit that belonged to to the government because there was some sort of integrator. So we see like nation states or like governments like the US investing like enormous amount of money into offensive capabilities but they also keep being burned by insiders.

Tracy Alloway: Mhm.

Matt Suiche: A lot of those capabilities also as strong as the internal coercion.

Joe Weisenthal: But I guess what I'm asking is, you know, if you're the Department of Defense or I guess now the Department of War and you're thinking about developing in-house capabilities versus partnering with a company like Enthropic and we should talk about all the drama that's going on there. How are you balancing those decisions nowadays versus say 10 or 20 years ago? Like my understanding is that now a lot of it is also like outsourced because they cannot really develop uh as many capabilities internally. So now like we have seen with uh entropic that it had been used in the Maduro operation

Tracy Alloway: right

Matt Suiche: and then after that like there was a pull out from entropic because they said it was violating their ethical like um

Tracy Alloway: standards

Matt Suiche: like policy yeah standards. So I would say now something that's really changing like very fast is the incorporation of like AI into those decisions. But as we all know like AI can also hallucinate. So even Dario the CEO of entropics that said it's definitely not in a state where it can be used for like fully autonomous decisions like that. So I would say like the AI element would be like the main difference. Even we start to see it now for like exploit development or vulnerability discovery, but it's still too early to kind of like give a definitive like opinion about it. But overall, I would say it's very similar.

Tracy Alloway: Well, talk to us about exploit development because I know that you can't just go to claude code and say like, I'm working on a zero day malware attack. Help me figure this out. But, you know, I also know that there are some very talented people who pride themselves in being able to jailbreak AI and illicit outputs that the labs do not want their AIs to produce. So, do you have a sense how just within the pure like hacker community AI is being employed today for these purposes or what they're able to get out of these tools?

Matt Suiche: It's a good question. So, like we started to see people leveraging like AI for like bug discovery. Yeah. which actually is becoming like pretty good. I think even Entropic published an article explaining how cloud can be used for discovery into smart contracts and how it found some bugs automatically. And I think even recently they released something called uh cloud for security that was aiming at uh doing um code uh assessment. But now we're entering into this like interesting uh paradigm shift where the cost of software is going towards zero. Right? So if you're a company and if your cost of building software is becoming less and less, it's also hard to convince people that auditing software for security reasons is going to be more expensive than developing the actual software. So I think that's one of the shift uh we're going to see. But when it

Joe Weisenthal: Sorry, can you explain Sorry, pause, pause that last part. What what did you mean by that? It's going to be hard to convince

Matt Suiche: if if you're going to have like uh to allocate budget for like building a product. So you have most of the budget that's usually allocated for like your software engineers to build the software and then you do some code review afterwards to make sure there's no vulnerabilities before it gets released to the public. But security risk is usually like pretty high. You cannot just rely on like AI tools at least not at the moment. Maybe in a year from now it's going to be possible. So it's going to be like really interesting to see how it's going to like do like a market shift because now with cloud code and as a famous uh vibe coder Joe I'm sure you know that

Joe Weisenthal: the cost of building software is approaching like zero if you just look on the timeline. Are you actually on this note, are you a believer in the SAS apocalypse? Because obviously there's the argument that well now everyone can just create their own software fairly easily using natural language. But on the other hand, if you're a big corporation or presumably a government, you're going to want to have you're going to still want to buy software from an external provider given some of the security concerns, given that it might not necessarily make sense for various reasons, management reasons perhaps to recreate an entire software business inhouse.

Matt Suiche: So I'm definitely biased on that but uh as someone who thinks like the cost of software is going towards zero and as someone who is like watching the software cost like collapsing one of the thing that we uh realized is that data is the only durable asset in the AI economy. That's why we decided to work on DB the current startup uh I'm working on because like that's the only thing that's really going to have value long term if agents need something to to transact or like to take decision on because even if you look in term of like in any context you know if agents are designed to think autonomously you need to have enough information to take those decisions whenever you're going to have your reasoning loops. So software itself if you just build it is pretty static whereas like the agentic like feedback loops are more dynamic but what changes the context they take decision on. So definitely like SAS business are going to have a hard time because if anyone including the uh Shopify CEO can just rewrite an MRI software in one afternoon just to like look at his uh back MRIs, you can imagine like how disruptive it's going to be like by the end of this year. I think the only thing we haven't seen yet is like enterprise AI agent. So, so far I would say like since Christmas people are mostly like still playing around trying to find like a proper like use case uh we see a lot of like consumer like AI agent like open CL that really made like agents more mainstream but we really haven't seen yet enterprise AI agent. So as everyone is kind of scared of being replaced for their jobs, we haven't really seen like actual like AI agents replacing entire departments or full on like employees. So we have seen some disruption around like software engineering mostly to make like software engineering more efficient especially in term of like development with shorter like timelines but we haven't seen yet like proper like enterprise AI agents.

Joe Weisenthal: How do you define an AI agent?

Matt Suiche: So my view of an AI agent and I like to remind people like what an AI agent actually is. At the moment most of AI agent is just like a piece of code usually written in Python or in Typescript that's just doing a bunch of calls to like entropic open AAI and running in a loop and taking decisions and calling like third party tools like MCPS or web searches.

Joe Weisenthal: So that's mostly what an AI agent is. We tend to think of AI agent as a completely like different persona, but at the end of the day, it is just like a piece of software that's running as a service on a machine or on a server. So from a security standpoint, which is pretty interesting, it's just like another service or like software, but people really like to think of it in an other way. Well, just so like I from the security standpoint, I mean, one of the exciting aspects of AI agents is that they can work autonomously, right? You set a task and it can go out and find what it needs to do and it sees like, okay, this didn't work. I'm going to try it's going to try this thing. It's going to try this thing. Oh, I need to connect to this web service to get this information, etc. The downside of AI agents is precisely the same. The downside of AI agents is that they could do whatever they want to do and if it accidentally deletes a bunch of files because it thinks that's what's necessary to execute the task.

Tracy Alloway: So like I'm curious like from a security standpoint like I mean we've already seen examples of people getting private information exposed or as I mentioned the example of deleting a bunch of information. Is this like a new way to think about the security threat model? the fact that the capability and the downside are one and the same. It's the same. It's sort of like hallucinations, right? The ability to like create an output and it also, you know, is handinhand with the ability to create a wrong output, a false output. And so, you know, the ability to do a to act on its own is also the ability to destroy on its own. Is this sort of like a novel threat model or a novel paradigm in thinking about enterprise security?

Matt Suiche: From an enterprise security standpoint, it is pretty much the same thing in the sense of like if you're building software, you cannot just really like patch software afterwards and stuff because you never it never ends like security must be like built in and you need to have like a safe design from the beginning. What we have seen now is like whenever people do something aentic, they just give like all permissions up front.

Joe Weisenthal: Yeah.

Matt Suiche: Which is like probably the worst thing you can do. And if you're an enterprise, as you can imagine, if you just give like all permissions to an agent, it just becomes Murphy's laws. If something bad can happen because you gave it access to it will happen. So you're going to see like more data leaks for sure because there is no safety by design in those like uh architectures or like those like agents which is in term of vulnerabilities and like exposure would be like very similar to what we have seen over the last like 10 20 years. But you know if people are ignoring what has been done in term of software security for the last 20 years that's why we're going to have a lot of problems and I think we probably going to start to see like people especially in enterprise like pushing back a lot because there's compliance that needs to be like you know like answered to so you cannot just give like full access to like you know your agent

Tracy Alloway: Speaking of the long arc of history, one thing I really wonder is you've obviously been in this space for a very long time at this point. Can you describe how you think your own career and I guess coding experience would have been different if you were say starting out now in 2026 versus I guess you would have started out in like the late 90s or early 2000s maybe even before that.

Matt Suiche: Yeah, mid mid 2000. Well, I would say like what has changed is back even like even without going through like uh the 2000s like u back in the Snowden days when the global surveillance program was being exposed like a lot of people were really like scared of it and scandalized by it and pushed back and people really cared about privacy. was like now we're entering in a new arc where very few people care about privacy

Joe Weisenthal: where you see like the CEO of Entropic being asked why he refused to work with the US government

Matt Suiche: and he says well they wanted to do a domestic surveillance program so that's against or like safety like chart so this is all aspect of people relationship with data which I guess is very different in term of software obviously like now you can write more things anyone can write anything but I think we're still in this weird adequate software phase where

Joe Weisenthal: we know what AI can do but it cannot really do like anything more anything less yet we still haven't seen like the actual like uh use case for it because it's obviously like very exciting and it feels like a lot of it is very different from before

Matt Suiche: but we don't really have like any evidence of how it's really helping in national security. I know it's been helping people who have been analyzing like Epstein's emails because there's a lot of data and that makes it faster but in term of like real use case I don't think the like it feels like the current world is very different than before but there's so much noise and so much like slop all over that

Joe Weisenthal: in a way it is pretty similar

Matt Suiche: I don't know if that makes sense

Joe Weisenthal: well I mean it does

Tracy Alloway: your company is b.ai AI. So, you must think that it's going to be used or that, you know, or that there's clearly something there. Actually, tell us about I'm actually really I'm on I'm on your website right now. It looks really interesting because it's something I've been thinking about, but you must have some vision for like where it's going and that there will actually be significant demand for these services.

Matt Suiche: Sure. I mean like I was saying like I think now with like anything that's agentic building software which is like the main use case so far for AI it's going to make software like going to zero. So the cost of building software is going to

Joe Weisenthal: so that's real. So like say like there's in your view there's no question that already AI I mean talk about a use that's a pretty big use case right there bringing the cost of building software to zero.

Matt Suiche: Yeah. I mean like even if you look at the author of cloud code he said they didn't write like code since November.

Joe Weisenthal: So like a lot of people are like this you know even us internally like we definitely like use cloud code a lot but if software is going to zero like what's left in term of like the internet layers so our conclusion is that data is uh the only thing that's going to be like timeless in the AI economy. So we're building like a layer for that. Especially now there's like all those innovations around like payments and stable coin that you can use to actually like pay like anything online.

Tracy Alloway: Yeah.

Matt Suiche: So we think like okay you have like issues like entropic or like openai just scrapping internet and using like public internet. So you may as well find ways of charging like bots or agent for your data. So at least you can have this new like revenue unlock that's going to emerge in this like new economy because I think a lot of the traditional like economic model like for instance like we said like with SAS are going to be like disrupted a lot. So there is a completely like new market around how people are going to consume data and I think people people would just be ready to pay like more to have high quality data because the more noise there is the more you want to make sure that the data you have access to is going to be uh like valuable and real. So yeah, even from a security standpoint, you know, like once you build like an infrastructure layer, you can have this like built-in security to make sure that the data you give back or the access like for the interface that you define is actually secure because

Joe Weisenthal: even if you look at open claw for instance, one of the top skill was malware.

Matt Suiche: So like people just like living in this like wild west but they just run everything. So it's us anticipating that enterprise is going to look very different and they won't just run like anything they find online.

Joe Weisenthal: Yeah. You know this is one of the things that I've encountered in my vibe coding foray is that one of the annoying things is okay you want the agent to like go out and grab some information or query some database or whatever and then it's like okay like let me know when you've gotten a an API. They're like, "Okay, come back and get an API." And then you have to like go to a website and then you have to like get out your credit card and you have to set up an account and then you like get an API key and then you copy and paste it and so forth. And that's very annoying. I want what I want is for the agent to just be able to go there say, "Oh, you know, like let me just pay you with some stable coins, etc. Just go out and get the information on its own without having this human in the loop." But it also occurs to me, and this is something that I've asked about with others, which is that like once I'm just like entirely operating in the terminal and the agent is going out and scraping information for me, etc., like why do we even have a free public internet anymore? And so like I'm curious like whether the direction of the internet and information in general is just entirely like you know paying microtransaction or fees for data consumption so that the data then arrives in some usable form in the terminal that I'm operating on.

Matt Suiche: Yeah. No, I think you you're raising some uh really good points. Uh we we need to talk after you can be our new design partner, you know.

Joe Weisenthal: Yeah, I'll be I can be a consultant.

Tracy Alloway: He's going to hack all the API I can be I can be a consultant. No, I hate having to deal with all these API keys and why am I still using my

Matt Suiche: You're like completely right. See, I told you you're a famous live coder. You know, it's like but like the use case you're describing now is like makes entire sense. That's why like we position oursel kind of like as a trusted like provider where like

Joe Weisenthal: instead of having to go like everywhere

Matt Suiche: to get data, you know, you have like this single point and unified access a bit like open router for AI models but for data providers. Because if you think about it when you use like cloud now in your terminal like the level one is basically you asking the model itself for information but as we know like the model may be like a few months old and doesn't have access to all the information.

Joe Weisenthal: So like the second level is the agent like clot code or or cloud for desktop or like chat GPT doing web searches and that's just them looking on the internet.

Matt Suiche: Yeah. doing like a Google search etc. But that's still not giving you an access to like the actual like relevant data that you would need where for instance you would have like those API keys and stuff. So like the level after that is basically access to private information. So and private information usually the one that's valuable like in the case of Bloomberg for instance Bloomberg has a lot of extremely valuable information but you can only have access to it through the terminal once you have the subscription or like any SAS services if you think of like SAS platform are just like some fancy UI where you can like browse the database but the data that's valuable is just in the database directly. So if you would have access to those APIs and would not have to create like 1 million subscriptions left and right uh because there's already like someone who's doing the integration for that kind of like as a programmatic like API marketplace but the integration now is like much easier and especially if you trust it because now what we have seen also is like whenever you use all those tools they make you install like MCPS but more and more people are like moving away from MCPS they're just using like skills Because like you just said, we start to spend more time into the terminal. So like the terminal and anything that CLA is becoming like a natural interface for like agents even for humans because you don't need to try to understand those like fancy UI and UX. You just like okay I want to do ABC just do that. So I think it just makes more sense to have like this interface for it and like like you said you know like otherwise it just like it becomes too complicated.

Tracy Alloway: Just on this note can we talk a little bit about I guess institutional knowledge of code because I remember one of the things that happened in the early days of AI development. I mean not that early. I think it was like 2017 or something back when Facebook's AI lab still existed. What was that called again? like the acronym was fair or something. Yeah, they had like a little Facebook like experimental lab. Anyway, they invented a bunch of chat bots and the chat bots started talking to each other in pretty much incomprehensible language, but like they clearly understood what each other were saying.

Joe Weisenthal: And so I'm just wondering if you extrapolate that to AI generated code. Could we have a situation where the models are constantly iterating on themselves? They're constantly talking to each other and so we end up with a system that becomes very very difficult for human engineers and coders to actually understand.

Matt Suiche: Yeah. I mean I think what we're going to stop to see is like you humans you know like us are going to move towards like creating like markdown files as like a programming language. So everything is just going to be like normal like language. But for the machine itself obviously. Yeah, I remember that video uh that is like gibberish like voice like transfer thing.

Joe Weisenthal: Well, if you think about it, it's not that much different from voice to text in that sense and at the end of the day just like bits like being transferred because even like whenever you connect on a web page, you know, like you write it in text but behind is just like byes that being exchanged. So agents still need to agree on the protocol that they're going to use and not necessarily an encryption format but an encoding format. So once you know what it is, you know, it just becomes like like a reverse engineering problem or like a foreign sake problem where you're just like, okay, like this is what's being used when those packets are being sent, you know, let me just decrypt it once you know the protocol, you know, like uh that's so I don't think we're going to end up in a a situation where like we would have like no idea of it because you're always going to have like people who are like pretty good like reverse engineers, but at the same time you're also going to have your like AI assistant who's going to help you to like reverse engineer those things.

Matt Suiche: So in a way even if it happens like you are not like alone with your Red Bull and your laptop and no AI agents. We all going to have AI agents.

Joe Weisenthal: Yeah.

Matt Suiche: So that's also like the reality of things. So we're far from just like the clippy plugin that we used to have in Microsoft Office. Now you can have this like CLA interface where just give like commands.

Joe Weisenthal: Yeah. And then it translate that into like okay I understand what I need to do you know and that's it.

Tracy Alloway: No I love like interacting with just the CLI now and like every time I have to go to the web it feels like some sort of failure like I'm like oh I have to go to this bright website. Yeah. And I just like want the information right there in the black screen talking you know communicating back and forth

Joe Weisenthal: in English background you know that's like feeling familiar for you. You know, something I'm thinking about is I imagine that there's a lot of like crusty old Linux and Unix programmers who are like, "Oh, this code is this isn't highquality code that the AI that the chatbots produce. This is slop code. It needs all kinds of fixing and so forth." From your perspective, is the code itself of good quality or of improving quality? How just the lines of code itself from your standards, is it good stuff?

Matt Suiche: Yeah, it's pretty good. And even when it's bad, you just tell it, okay, like this is bad, you know, like just do it better, you know, like if you're using like negative rewards, you know, if you say, okay, like this piece of code is garbage, you know, it's going to understand better because you kind of give like a strong emotion whereas like if you say, oh, it's okay, you know, you just be like, okay, like whatever, you know, if it's okay, it means it's passing like the

Joe Weisenthal: the adequate test. Whereas if you garbage, you write it and I always do this. So if I ask a bot for something, I will always say like after the first version like do better and just see what it comes up with and then just try to iterate on that.

Tracy Alloway: This is this is tough because you're talking when you're talking in English, you the brain deludes itself into thinking that you're talking to or when you're talking in human language, any language, the brain deludes itself. You feel like you're talking to a person.

Joe Weisenthal: So you have to be nice.

Tracy Alloway: Well, then I feel like

Matt Suiche: I don't have that problem, Joe.

Tracy Alloway: And that Oh, really? But then I feel like I like I don't want to say, "Oh, this is garbage." I don't but from your perspective it's actually sort of important to to be firm with the bot and you get better results by being more sharp with it.

Matt Suiche: Yeah. Because it's the equ equivalent of a negative reward like just like positive reward. If you if it does something you say okay like that's great. That's exactly what I tried to explain. Then it's like okay like that's like a reference point was like if it starts to go on a tangent you just say like oh that's completely like out of the line. redo this like why are you doing this you know like uh the more explicit you are

Joe Weisenthal: the better it's going to understand how far it is from the requirements whereas like if you're being

Tracy Alloway: I just have to get better you know I'm sort of conflict avoidant I know

Joe Weisenthal: and I I like being nice to people so I just have

Tracy Alloway: Joe still says please and thank you

Joe Weisenthal: so yeah that's right so I have to I just have to be like no this is trash this is garbage you totally we're all dumber for having seen this code okay this is good to know

Matt Suiche: I mean it's a point you know like all those like AI companies they're recording all the prompts you know so god knows if they're keeping it if there's retention around it you know like we know the open AI is keeping them you know like they can get supony is this how you pronounce it

Joe Weisenthal: yeah so like who knows you know like in years from now if there's like full on like autonomous like

Tracy Alloway: robots managed by the department of war because they think it's lawful so

Joe Weisenthal: we all get social scores based on our AI promps treated like I I these these thoughts creep into my head unironically where it's like at some point is there going to be some am I going to regret having

Matt Suiche: I don't know they I did they they were there in my head

Tracy Alloway: truly a brave new world Matt just one last question for me but going back to cyber security and the current situation with Iran what are you on just the lookout for like what would peique your interest the most to see in that particular space.

Matt Suiche: I think now in that particular space because I'm one of those people who think it's related to the Epstein files, you know, like it's just more about like getting more Epstein files related stuff to see if there is like more connections to it, you know. So I think that would be the only thing that would kind of be like digital that will like spike my interest because now we have seen you know like um just those like old drones can do like so much damage and that Iran demonstrated that they can be like really precise uh with their attacks. So now I think it's more about seeing like which direction it's going to go to and how long it's going to last which is like the the big like question mark because there's so many other components like the energy sector you know like how is it we have seen the price of like memory like increasing like a lot. So now now if they're starting to block like the Detroit of Armus, you know, like like what's going to happen to like the cost of data centers and like memory and AI in general, you know, like we're going towards on one side, you know, we're going towards like the cost of tokens and inference going down, but that may also like bring the cost up, you know. So if you're going to use like AI for like your next generation wars, but if your enemy can just like increase your cost of token and insurance, what does that even mean? like do you even need AI in the first place? Is it even relevant? So I think there's this all like asymmetric warfare that's going to happen that we really haven't seen yet and I think that's going to be like really interesting but at the same time there's so much noise and so much so many things happening at once that it's becoming like extremely like hard to focus and just like extract like what's really relevant.

Tracy Alloway: Yeah, it definitely feels like that. Tough tough choices potentially coming for Havlock as your your token costs go up. Matt, it was so wonderful to reconnect again. Thank you so much for coming back on OddLots and uh yeah, you'll have to get back to Joe about those APIs.

Matt Suiche: Yeah, I'm I'm happy. Uh let's uh let's

Joe Weisenthal: the most famous Viper on Twitter now.

Tracy Alloway: Yeah, bring me on as a consultant.

Matt Suiche: Sounds good.

Tracy Alloway: Take care, Matt.

Joe Weisenthal: Thanks, Matt.

Matt Suiche: Bye-bye.

Tracy Alloway: Joe, always good to catch up with Matt.

Joe Weisenthal: Super interesting. It's incredible how much is happening right now at this particular nexus, especially obviously the anthropic stuff. But I, you know, it's interesting. You know, you think about cyber warfare and you think about, okay, we're going to hack into a system and take out critical infrastructure, but another thing you can do is just attack a data center directly.

Tracy Alloway: Just send a drone to a data center. I thought that was really interesting that sort of like we got very used to thinking about cyber as like this thing that exists only in in code. Um but now you have this like new front of kinetic warfare where the two like really intersect. Yeah, they do really intersect and yeah, these are like huge national security vulnerabilities and his he pointed out I mean certainly today but you know was this in anyone's threat model thinking about the risks to it you know the cheapness of drones the ability to take them out super interesting also just this idea like yes you know obviously again as your observation the point we think of like cyber attacks like we're going to take out this whole thing but in at least in the warfare context his point like most of it is like before the war etc. the sort of information gathering spy spycraft and so forth prior to the actual the actual attacks.

Joe Weisenthal: But it is interesting to see Israel in particular use some cyber attacks as a sort of sewer of chaos in Iran. I can't imagine what it's like to actually be on the ground there at the moment for many reasons, but like

Tracy Alloway: you imagine just being there worried about your physical safety and then the traffic lights aren't working as well.

Joe Weisenthal: Well, and Right. And also just think like wait there's cameras everywhere or there you know like how much is being recorded like create like a sense of like paranoia among everyone about about everything

Tracy Alloway: and also the interesting thing obviously this is very topical in markets but the SAS apocalypse idea Matt in particular seemed pretty bearish on the outlook for um existing software companies I guess and I did think his comments about what that would mean for security budgets within organizations were pretty relevant. and worrying.

Joe Weisenthal: Absolutely.

Tracy Alloway: All right. Shall we leave it there?

Joe Weisenthal: Let's leave it there.

Tracy Alloway: This has been another episode of the Odd Thoughts podcast. I'm Tracy Aloway. You can follow me at Tracy Aloway.

Joe Weisenthal: And I'm Jill Weisenthal. You can follow me at the stalwart. Follow our guest Matt Swish. She's at Mswish. Follow our producers Carmen Rodriguez at Carmen. Dash Bennett at Dashbot and Kalebrooks at Kalebrooks. And for more OddLots content, go to bloomberg.com/odlodge for the daily newsletter and all of our episodes. And you can chat about all these topics 247 in our Discord, discord.gg/odlotss.

Tracy Alloway: And if you enjoy OddLotss, if you like it when we talk about the intersection of kinetic and cyber warfare, then please leave us a positive review on your favorite podcast platform. And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely adree. All you need to do is find the Bloomberg channel on Apple Podcast and follow the instructions there. Thanks for listening.

Matt Suiche: Yeah.

Joe Weisenthal: Yeah.