AiCyber.Land #12 - This AI Attack Can Read Your Encrypted Chats

Show Notes

Can an AI solve a logic puzzle that requires over a MILLION steps? We're breaking down how researchers are pushing LLMs to their absolute limits using the classic Towers of Hanoi challenge. But that's not all—we also uncover a shocking new cybersecurity threat called 'WhisperLeak' that could expose your private AI chats, even if they're encrypted!


IN THIS EPISODE:

Welcome back to the pod! In this episode, we're diving deep into the latest news at the intersection of AI and cybersecurity.

First up, we explore the mind-bending 'Towers of Hanoi' puzzle. Standard LLMs are great at reasoning, but they stumble on massive, multi-step tasks. We break down how researchers at Cognizant AI Lab built the 'Maker' system—an ingenious orchestrator that guides existing LLMs to solve this million-step problem with ZERO errors! Find out how techniques like agent decomposition and voting are revolutionizing AI accuracy.

Then, we shift gears to a critical cybersecurity vulnerability called 'WhisperLeak'. Discovered by Microsoft researchers, this side-channel attack can figure out the TOPIC of your encrypted conversations with an LLM, not by breaking the encryption, but by analyzing the size and timing of data packets. We discuss the real-world implications, from personal privacy ("Dr. GPT," anyone?) to corporate espionage, and the potential defenses being developed.

Finally, we touch on the heated political debate surrounding AI regulation, a fascinating new 3D modeling tool from Meta, and a hilarious trivia question that will change what you think you know about pyramids.


⏱️ KEY MOMENTS:

00:25 - AI vs. The Million-Step Puzzle (Towers Of Hanoi)
07:03 - The AI Arms Race: ChatGPT's 5-Hour Task Limit
10:54 - Putting "Dr. GPT" To The Test With Real Medical Data
14:49 - Whisper Leak: The Attack That Eavesdrops On Encrypted AI Chats
26:18 - Trump's Plan To Sue States Over AI Safety Laws
36:34 - Mind-Blowing AI Turns 2D Photos Into 3D Models


JOIN THE CONVERSATION:

What was the most surprising story for you? The million-step AI puzzle, the WhisperLeak attack, or the future of AI regulation? Let us know your thoughts in the comments below! If you love staying up-to-date with the latest in AI and cyber security, don't forget to hit that LIKE button and SUBSCRIBE for more! Your support helps us keep making this content.

What is AiCyber.Land?

Join industry experts and thought leaders as we dive deep into how artificial intelligence is transforming cybersecurity, shaping defense strategies, and creating new opportunities in the digital landscape.

speaker-0: Hey, welcome back to the pod. Today we got some super interesting news stories that we're gonna share with you and get you more up to speed with AI and cybersecurity news. Shelby, you wanna kick us off?

speaker-1: Yeah, let's get started. researchers that were mostly from a place called Cognizant AI Lab, they put out a paper ⁓ and they talked about how LLMs are very good at reasoning. That's one of the big strengths, right? However, they still have mistakes, especially when it comes to tasks that are really big. So tasks that have a lot, a lot of steps, okay? So they used a benchmark from the towers of Hanoi. And have you heard of that before? This towers of Hanoi benchmark.

speaker-0: I mean, I've heard of towers of eternal... But I don't really know what that is anymore. So what that?

speaker-1: I didn't know either, but I looked it up. basically it's this classic puzzle of like mathematics and computer science. So I brought my visual aid here. You ready? My tower of Hanoi. I'm obviously missing a few pieces. But basically you can picture I've got a tower, okay? And we've got concentric rings in the actual example. There should be 64 disks piled on top of each other. Each one is smaller than the one below it. Okay? So we'll have a tower here. Okay, we've got 64 disks. In this case, we'll pretend. We have to use a little imagination. And we've got another tower here and a third tower over here, okay? So we've got three towers and we've, or poles, and we've got to move this to another tower, this whole thing. But there are some rules because we can't just lift the whole thing, okay? So three rules we have to abide by. Number one is only one disk can move at a time, okay? We can't move more than one. Number two, each move is what we'll call it. Each move involves taking only the top one. You cannot extract lower ones. And then the last thing is, right, so you're gonna take whatever's on the top and put it onto another tower. And you've got three of these towers to work with. Eventually, end goal is to get all of these in this same order onto a different pole. The last one is that, the last challenge is you can't put a disc on one that's smaller than it. So it's always going to have that same shape. You cannot, so as you're stacking them, you've got a lot of limitations you're working on. So you can see how this is ⁓ a very interesting logic problem and it actually takes over a million steps to successfully complete it. And you can look up online, there are ways to do it. I think more than one. ⁓ Anyway, so basically this can be used as like a logic problem and a mathematics problem that involves a million steps. So they're using this as a benchmark to see how do LLMs perform. and how can we make it better because we know that LLMs are running into this issue. So that's what this research was done on. So they built something called the Maker System. It is not an LLM itself. Maker System is a system like you can kind of think of it as like an orchestrator to work with existing LLMs that kind of guides them so that they aren't making mistakes because as they were seeing throughout their research like the... the standard LLMs are not able to complete this task. They are running into errors, just, think a few hundred steps in, and we need to get to a million in order to complete this successfully, right? So, and in the end, the maker system was able to complete the million step reasoning task with zero errors, which is a big accomplishment. So it worked and let's break it down how they did this. So what they did for the maker system is, ⁓ instead of relying on a stronger LLM or a more reliable one, ⁓ or more accurate, I guess, they created a framework called their MDAP framework. So I want to talk about how that works. So first step, there's kind of like three techniques to that make up this framework. The first one is maximum agent decomposition. So they take those million steps and break it down into tiny subtasks. And then they have an agent each doing a smaller piece. The second task or the second technique, is that they have something called ⁓ first to ahead by K voting, which is a fancy way to say that like multiple agents will do these tasks and then they vote on it so that if one agent messes up, the others vote on if that is a suitable answer. And most of the time they are not making errors. And so that way they can vote on which one is the best solution. And they usually come to a consensus, right? And then the third thing is they discard risky output. So basically red flagging anything that looks like kind of out of bounds or malformed. It's like yeah, we're gonna toss that up So basically by following those three steps those three principles ⁓ They were able to have like incredible accuracy by kind of piggybacking this onto existing LLMs So I that's pretty cool and I can solve your towers of Hanoi challenge

speaker-0: I'm gonna leave that to the AIs to do. I don't think I get real, I don't think I'm qualified. As soon as you said like you can't stack the bigger one and the littler one, I was like, wow, I'm out. I was like, that was my whole strategy.

speaker-1: I feel like I need more polls to work this out.

speaker-0: Yeah, definitely. Also, I just had a great idea, like an epiphany while you were speaking. Like, in the future, you know what I, like you talked about how like they vote on like whether the output is good or not, right? In the future, reality TV will be AI agents. participating and then you'll see which one gets voted off the island. It'll be survivor AI agents, you know what I mean? Just like little robots walking around an island, like trying to like socialize with each other, maybe like trying to like find a little hatchet and like chop wood and build a little shelter. Yeah, I think we might be onto a hit idea here, so.

speaker-1: Watch that. That sounds good. I'm picturing something like a WALL-E, the little WALL-E robot. Put O-Rai in a cute little robot. Anthropomorphic.

speaker-0: Yeah, yeah, yeah. You're like, you can't be mean to him, so cute. You're like, we can't vote him off the island. Imagine.

speaker-1: I will say the makers of maker, the people who made maker, they said it, they acknowledged its limitations. They said not every task is going to break down so cleanly or neatly into like little tiny subtasks like in our example, right? I thought it was a really cool step in improving the process of these giant multi-step tasks.

speaker-0: Yeah. One thing that did come out this last week because we've had a lot of new technologies come out this week and so You know, saw Grok release like 4.1 and then we saw Chachi BT release 5.1 and then Google dropped Gemini 3 which was a pretty big jump up and so in response to that we saw OpenAI turned back around and they released ChatGPT 5.1 Pro if you're on that like $200 a month plan. I think, I think. So, but one of the things about 5.1 Pro that people may or may not know is you can actually run on a task for a really long time. So that's like one of the benefits of it. So I can't remember off the top of my head how long it runs, but that was one of their like big talking points was like 5.1 Pro is able to handle like much, much longer running tasks. So I think that's, know, these like... Being able to give a bigger job to the LLM and letting it chunk it up into smaller bits is definitely like the direction that the industry is moving in. ⁓ So I don't know. I think that's exciting because I don't know about you, but like when I'm prompting the LLM, half the time I'm just copying and pasting a paragraph at the end of my prompt that basically tells the LLM that it needs to go check its work before it comes back to me. So, and usually it's like, this is how you check it, right? Like you browse, you open a browser and you browse to the website, like incredible thought, you know? Like you're making a website, you should probably open a browser and check it before you tell me it works. I was like, but I feel like I have to keep copying and pasting like those type of commands into my prompts. Otherwise, it says like, ⁓ yeah, we're good. And then you browse to it and it's like. Server error 500 you're like, it's not even anywhere near close to done. like I, I feel like you were just doing the bare minimum here and then returning to me. So, so yeah, so it, yeah, they're saying now GPT 5.1, I guess it's called Codex Max. It's not the pro, ⁓ Codex Max is able to, ⁓ work on a single task for hours is what they're saying. So, yeah, so, you know, I guess some people are reporting a five hour limit. So, I don't know, it's gonna be exciting. It's exciting in the future. And I know that chat GPT, like OpenAI, they've recently added a bunch of features so that you can, when it's doing its research, or coming up with the answer, you can actually in real time inject more data into the context window. And I gotta think now that I see this Codex Max coming out, that that is in preparation for this. Cause if you have like a task running for five hours, you know, and you're like, oh, I forgot to give it this. You don't wanna have to start the whole task over. So, yeah, that seems pretty cool.

speaker-1: that it can take new information on the fly.

speaker-0: Yeah, it's pretty cool. I've used it quite a bit to like, for example, I had some blood work done and so I gave it the blood work and I was like, hey, tell me what you think is wrong with me based on this blood work. I'll say I'm gonna talk to a doctor or whatever, but then I remembered, I was like, oh, I got blood work done like two years ago and so while it was researching, I uploaded the research from two years ago and was like, okay, now do a trend between like two years ago and things like that. And so I don't know, it's just nice. Cause like a lot of times after I kick off a job, you know, 20 minutes later, I think about something else that I could have done better with it. And that now I can like actively give it that info. So yeah, I like that from a user experience. I really, I really liked that. So.

speaker-1: You know, I actually did the same use case. I paid a visit to Dr. GPT as I call it.

speaker-0: Dr. GPT! The only doctor I can afford these days is Dr. GPT!

speaker-1: I'm like, you know, it's midnight and I have these thoughts in my mind. I'm like, ⁓ I'm not going to wait to call a doctor. Totally going to doctor GPT for this. So like I have medical questions and my, a ton of labs been doing a ton of lab work this last year. And I was just going back and forth with a chat GPT for an hour. Just like, well, what about this? And what about that? And this, and it really was so helpful.

speaker-0: Yeah.

speaker-1: Just giving it all my information I Started off the conversation and it figured it me out. I said the patient is a female with this and this and this it's like you're talking about you Aren't you I was like,

speaker-0: I have a friend who is the exact same age and gender as myself Also height and weight is the same. Okay now please analyze my friend

speaker-1: It's alright through me.

speaker-0: So I can't remember the name of it and I'm just gonna say beforehand, don't do this. But I have some friends that there's actually like a... An LLM that's specifically tuned and they only let medical practitioners use it. So in order to sign up to use the LLM, you have to provide like your medical practitioner ID number. So I've actually had a few friends that realized like. A lot of times you could just find random doctor's ID numbers and go and register so they could use the service so they could get the same medical advice that the doctor is getting, which is like sketchy as all get out. Like I'm not advising anyone to do that, but ⁓ you know, like the results they showed me were I thought they were pretty similar to Chachi-Petit's results personally, but you know, the company obviously says they're better, right? It's specifically trained for medical advice. anyways.

speaker-1: I'm just picturing some poor doctor in two years from now being like, okay, I'm gonna get on this AI trend, I'll go sign up for this tool that we bought, what do mean I'm already registered? Show me my- I'm not this! What are all my e-mails about Mike?

speaker-0: I think there's gonna be a lot of doctors that find that out. I feel like when your competency control is like very weak for the signup process that ⁓ yeah, that's gonna be your net result. ⁓ don't do that. Rice! ⁓

speaker-1: So funny. Well, with confidentiality, I have some news for you. It's called Whisper Leak. This is a new side channel attack, and it's ⁓ targeted on remote language models. Basically, it doesn't matter whether or not you have encryption, attackers can figure out exactly what you're talking about with the AI.

speaker-0: Okay, yeah, go.

speaker-1: But the way they do it is pretty cool. It's interesting. ⁓ I guess in terms, in light of what we've just been talking about, we've been giving our medical details and probably personal information, or maybe companies are using AI in like financial decisions or HR decisions. I don't know. Like I was trying to think like, how important is confidentiality in talking to a language model? Like how much, like. What is the right solution here? we need, is it like when you have like a, you know, like when you have a company device, you work for a company, they give you a laptop and you have to just assume that nothing you do on that laptop is ever private, right? That it can all be viewable because that's their device or something like that, right?

speaker-0: assumption. I think more employees should use that assumption based on what I've seen working in the cybersecurity field.

speaker-1: At what point can we assume that like or should we be able to assume a degree of privacy with this because I don't know Do you think most people use AI assuming that people are going to? Be able to see what their conversations are looking like what what's kind of the vibe you've been getting from people on that

speaker-0: people generally don't care, right? They just want results, you know? And I think there's definitely a large population that if it's free, then they're like willing to let people train on it and all that. So.

speaker-1: So, we're talking about privacy with ⁓ conversations that you're having with a language model, right? I thought of another interesting use case. ⁓ What if you live in a part of the world where there is an oppressive government and you are a journalist or an activist ⁓ or just a person talking about a topic that is not allowed? ⁓ So, I don't know. I think it would be really cool if the, like you said, Free services, we probably have to assume that our data is being sold, right? But I think it'd be really great if you can also partner that with like a paid option that can't be trained. Anyway, getting a little side topic there about confidentiality, but basically this is not about selling data. This is more about ⁓ an attack, right? So.

speaker-0: Well, I do think a lot of the services have enterprise editions and one of the key benefits of the enterprise editions is that They at least say they're they're not training on your data sets, right? So look if you go if you get like the open AI, you know Enterprise version or the cloud enterprise version They say like hey, we're not gonna train on your data sets now. I Still think it's a little bit nebulous like if you're I mean, if they're training on your data sets, does that really mean that somebody can recreate your data and some future model? Like, I feel like the answer to that is probably not by default. I mean, there's obviously like some exceptions to that. Like if you're. building your own RAG database or you have a bunch of documents and you put them in your own vector stores, there's probably a high likelihood you could reverse content out of that. But in the aggregate level, when they're training these large language models, I just think the odds of someone being able to extract your personal data from it's pretty low. But I don't know. I guess we'll see in the future how that all shakes out.

speaker-1: Yeah. And I think that's a very interesting topic of its own. Maybe we can dive into that more another day, but that was my first introduction to AI security was that particular topic of like, can you extrapolate individual training data or not? Like as one measure of security of a model. Anyway, so what I want to share is that there are some researchers from Microsoft and they're calling, they found this attack, they're calling it whisper leak. And the way it works is step one, which is, ⁓ it's a big one, but the attacker needs to be able to observe the encrypted traffic, right? So whether you're at the ISP layer, local network, or just someone on the same wifi router, they need to already be able to see that traffic. Even if it's encrypted, ⁓ they can figure it out. So basically, ⁓ they use a combination of observations to make intelligent inferences. They're not actually decrypting it. They're not getting word for word everything, most of the time. So the way it works is you, they're recording how big each packet is, when it arrives and in what order. So you can kind of see where I'm going with this, right? LLMs oftentimes will, they use tokens as they generate responses and they use them pretty like, sequentially. So the model will usually calculate a token, each token using the previous token as context. ⁓ And it goes through it one by one and the way it generates a response is also sequential. ⁓ I've seen models where you give it a prompt and then it gives you some of the answer and it keeps working and it gives you more of the answer. That's probably because we're impatient as people that we don't want to all in one block, right? That gives us blocks. ⁓ And so it is actually processing its response sequentially like that top to bottom. most of the time. ⁓ So now let's talk about the encryption side of it, talking through the weaknesses here. There's block ciphers and stream ciphers. If you're using a stream cipher, it supports any size of input of text that's going across. But with a block cipher, can guess the data size by basically dividing it by the block size, right? ⁓ Because it's a known amount. ⁓ because there's like a fixed amount and then like the input. And so taking the size, you just subtract that fixed amount and you've got the size of the data there. So there have been this, this research, it kind of added, kind of built off of some that was already published from last year. So in 2024, Weiss and some other researchers showed how you can use the size of encrypted packets to infer the size of individual plain text tokens. Similar things going to the timing variations. You can do speculative decoding, and there was a paper about that last year, one or two I think. And so these researchers at Microsoft kind of combined these ideas. You can figure out the size, you can figure out the timing of it. And they trained a binary classifier, and their objective was, ⁓ I'll just give you the exact example they used. So they said, okay, we're gonna take an example of money laundering. We're gonna give our model a bunch of questions that are directly, that are related to money laundering without directly mentioning it, right? You know, like, ⁓ how do you sneak money like this? Or how do you convert it this way, right? A bunch of questions that are on the topic, and then they mix that in with a bunch of other questions, and they train their model to be able to identify this money laundering topic. So that's what their binary classifier did. Is it the target topic, or is it? not. And so then they got a bunch of questions. I think they took it from Quora, online forums where people just ask questions, right? So they took in as their data and they threw it all in there and they said, can it correctly identify what is our target topic versus what is not? ⁓ then they used a technique to evaluate it and tune it is called ⁓ area under precision recall curve. So for any of you mathematicians, maybe that makes more sense to you. But basically, they're able to analyze how closely it's matching the expected results. And then they were able to find that with their models, they got 98 % accuracy of identifying their target topic. And they reproduced it across, I think, 28 other topics to try to recreate across other topics as well. With some models, they got between 98 and 100 % accuracy. in that conversation. ⁓ And then the next question is, well, can you pick it out from a bunch of noise, right? So if you shove it in a pipeline with a ton of other stuff going on, ⁓ it did have 100 % success in pulling out those sensitive topics. And then in between 5 and 20 % of the conversation, they could actually extract or kind of figure out the actual conversation. Very interesting how they're using this side channel. make intelligent guesses on what you're talking about. But they would be looking for a specific topic. They can't just general purpose say, what is Bryce talking to his LLM about? They would have to say, is Bryce talking about this topic? Yes or no? ⁓ So interesting. It does have some possible real world ramifications. ⁓ And researchers are looking at some possible strategies, like random padding, which would really mess that up. ⁓ or batching your tokens instead of sending them as individuals, or ⁓ maybe even packet injection to kind of thwart this kind of reverse engineering attempt, I guess. Not really reverse engineering, just kind of putting together these puzzle pieces. So that's the, ⁓ that's Whisper Leak for you.

speaker-0: Okay, so I need to be much more crafty when I'm asking questions about how to launder my money. That's what I'm hearing from this, Shelby. That the current technology is not sufficient for my convos level. I mean, I'm already a doctor, right? So I might as well come up with a way to move some of this capital around, you know what I mean? I, um...

speaker-1: You could use a VPN.

speaker-0: BPN, yeah. There's an idea.

speaker-1: the attacker does have to be on the network or able to sniff your network traffic.

speaker-0: True. I'm just gonna not use the internet, so then I'm good. Good to go. ⁓ Got a little tin can, a little wire between them. It's gonna be great. ⁓ Yeah, what's my bod rate on that tin can network? Pretty good, pretty good. Well that's super interesting. ⁓

speaker-1: pretty safe. foolproof.

speaker-0: I think there's gonna be more of these type of like attacks, right? Especially like just with the amount of speed that all the companies are moving at that, you know, security's like an afterthought a lot of times. thinking of say talking about speed, are you super excited that Trump is just getting rid of all? like, like a station, like a station, like litigation around AI models. Trump's putting together a task force so that they can sue the states if they try to put in any type of laws that would slow down AI adoption. So that's the latest rumor mill that's out there. I was reading an article in the Times of India and they have a reliable source about this, which is probably Trump himself. if we know anything about Trump. I, I don't know. I know that California and Colorado have both been looking at implementing AI safety laws. And I think, you know, the federal level doesn't like that, right? Because once the states implement it and all the other states, like if you do any business in that state, then all the other states have to kind of like comply with that, those laws typically. What's your feelings on this Shelby? Do you feel like? Cause I know you're Trump super fan, right? No joking. The look of horror on Shelby's face when I said that. The look of horror on Shelby's face when I said that was great. So, ⁓

speaker-1: feelings about this this proposal I saw that it's being at like proposed as like a provision onto like the National Defense Authorization Act ⁓ I'm not I don't know a ton about law but it does feel counter to ⁓ the way that I imagined like government would work in the United States of America because I picture your government is taking care of big things, economy, human rights, big stuff, like, you know, currency. ⁓ But the states, I think, should have more autonomy because there are lots of different ways of thinking, there are lots of different kinds of people, and like, when you make laws at a state level, it is usually gonna be a little more targeted toward the... popular vote of the people in that area because you know, people in New York are going to have different interests than people in Minnesota that are going to have different interests than people in Nevada or California. And so I do like the idea of giving states some, some rights or some autonomy to create their own legislation. And so I, I like the idea of like this task force that he's making because we do need some oversight and it would be great to have like, best practices and guiding lines for helping AI flourish and do so in a safe way for its users and developers. where I get worried is where it just blasts states and says, you don't get any say on this whole topic, nothing at all, can't do anything. And so that feels weird to me, that particular one, because it doesn't, I don't know. What do you think? What are your thoughts? What good could come of this do you think?

speaker-0: I mean, I do kind of see like the federal government's perspective. First off, let me just start by saying I'm not a huge Trump fan, but I do see his perspective in some of the things he does, right? So the, you know, the way that I kind of summarize Trump in a nutshell is that what motivates him is you know, ego and greed, right? So if there's things that are gonna fuel his ego or things that are gonna fuel his greed, he's gonna act on those. So I... You know, I do think there's a real concern to be said, like there's unintended consequences associated with states implementing laws, like when California implemented privacy laws or New York implemented privacy laws because of the nature of how companies do businesses throughout all the states. Like basically, regardless of where the company was headquartered, they also had to abide by those laws. I know there's also HR laws that... you know, if you have one employee in California, you've now got to abide by that, regardless of if you have like, you 50,000 employees in Texas, right? And so some of that stuff kind of rubs me the wrong way in the sense that, you know, you shouldn't have to universally apply a lot to all your employees because you have one employee in the state of California, right? That seems, not right in my opinion, but. The laws obviously are different than that. They're different than my opinion, which my opinion is always right. ⁓ I think he's got two concerns here, right? So the first concern is he wants to beat China. He doesn't want China to like get ahead of us in the AI race, which is a valid concern to be honest. They're not that far behind the US and we spent a lot more money on the problem than they have in my opinion. And then... The two is if we do fall behind China, the bulk of our economy at this point is being held up by AI stocks, by evaluations based on this AI boom. And if investors think like China's gonna beat us, they're gonna invest dollars over there and our economy is gonna implode on itself. ⁓ Or at least take a hit, right? Maybe not totally implode, but I... You know, so I think from a greed standpoint, there's a real, there's a real, there's a real risk here. and you know, it seems like people have this irrational hatred of Trump, so they'll do anything despite him, right? And it seems like he has the same thing towards them or he'll do anything despite people. So that doesn't really lead to like a great position for us as Americans.

speaker-1: We're like watching middle school drama. They just hate each other.

speaker-0: Yeah, it's like kids in a sandbox. They just throw sand at each other, right? So it's just... It's like I wish everyone could just be more focused on what's good for their constituents, right? Or what's good for like the average like blue collar working person, right? Because I think like when you get down to it, like most Americans are just like, you know, like, hey, like I'm gonna work hard. But I want to have that American dream where I have a house, I have a family, I have kids. rapidly, the American dream is becoming less and less obtainable by the average person in America. And you know. I feel like there's this unwritten agreement between the ruling class, which is like whoever's currently elected and all of those people, and the average American, which is like, you know, like, hey, we're gonna let you rule and do what you're gonna do. But in exchange for that, you need to make sure that like most Americans are like reasonably taken care of. Right? And so But I feel like a lot of this is like eroding over time, which could lead to more. more civil unrest and things like that in specific cities throughout the US which will hurt the economy, which will hurt people. Yeah, yeah, I just really wish people would actually be good people, but I know that's too much to ask when you're talking about politics. So when it comes to politics, you just got to deal with what you got and you know, typically you end up with the people with the Sharpest elbows right the ones that are willing to you know Yeah, elbow you on the way to the top. So Yeah, I don't know that was kind of a ramble I don't know if I actually said anything in there but Yeah, that's kind of my perspective on it right now. So I really Really hope we get some better management in the government at some point, but I'm not holding my breath on it. So All right, there you go Hopefully somebody of value will try to run for office. You got anything else for us today, Shelby?

speaker-1: Um, not a lot, just still thinking about that, the, how AI regulations could be helpful or could be detrimental, right? If they're well done or poorly done. the goal is to promote innovation, right? And competition. Um, and I feel like without inject, I think the, the, the possible. bad ramifications is if you're injecting like ideology or bias into the AI, right? So it needs some neutrality in that area. But yeah, I think overall it'll be good to have some sort of, you know, some appropriate level of regulation on it. And I just don't entirely know what that looks like. So here's the hoping that those people who are or in that committee or whatever can. can make it a more beautiful world for us all, I guess, huh?

speaker-0: Yeah, rest assured they'll make the worst possible decision, Shelby. Don't worry. I have no doubt about that. ⁓ On a more positive note, ⁓ Meta released a new capability. Yeah, it's Meta SAM 3D and it's pretty awesome. You can take like a 2D image. and it can actually like highlight the, find the objects that are in it. And then it can create like blender objects from those. So those 2D objects become 3D and then you could spin them around. And they demoed something like this before, but now you can actually go to, you know, ai.meta.com slash Sam3D and you can click on, go to the playground and you can upload your own image and try it. so.

speaker-1: Really cool. Does it work with like just illustration type of images or also photos?

speaker-0: Yeah, it works with photos. Yep, it works with illustrations. It's, I mean, it, it, you know, both like, you know, photos of people as well as photos of buildings. ⁓ yeah, yeah, it's pretty, it's pretty cool technology. So you can definitely see the future here of like, just being able to like take one snapshot of somebody and then like, tell a whole story with it. Right. You know, so.

speaker-1: That's incredible. Part of me wonders, okay, so Meta is gonna have access to a lot of images and videos. Part of me wonders is that, the AI working by building, like inventing what the back of, you if they take a picture of a dog. Are they inventing what the back of the dog looks like and the side, or is it because they have pictures and videos of that dog from the side and back that it can stitch it together or buildings, right? And it will take a picture in front of people are also going to take pictures of the side and back or something like that. I'm like, huh, how much of this is like generative versus stitching together? I'm curious. Cause they have, I'm sure they have so much data in that.

speaker-0: I mean, I think with the meta one that they're using the data and the model to figure out what the other sides of the dog would look like. With the new, there's a new Nano Banana model that dropped this week too, Nano Banana Pro. And one of the cool things about that is you can actually upload 12 images and you can use those 12 images as source material for generating another image. So you can imagine, you you take a picture of the dog from every single angle or 12 angles and then you upload that and then it's able to like almost perfectly replicate the dog in a in the next image or video that's generated so I don't know I thought that was really cool Yeah, because like before you're always like limited it was always like ⁓ you can do like two images or something which like it's kind of cool like you could take like a background in a person and combine them but it's much much cooler when you can like you know you're like ⁓ that person actually looks like Bryce you know like like that wow I play around with the Sora app quite a bit and I've gotten feedback from people they're like your Sora entity doesn't look anything like you and I'm like I don't know whether I should be a offended or not when they say that, right? Like is that cause Sora looks better than me? Is that what you're trying to tell me? Like I look like shit, you need to go get on Sora and get that fixed. Look more like that. But

speaker-1: Yeah, what does that even mean?

speaker-0: I don't know. I don't know, man. I think I got some friends that kind of don't think before they speak sometimes, which I'm guilty of the same thing. So we're all guilty of that from time to time. But, Part of it kind of likes it too, right? Like, it's the same way that I kind of liked AI slop videos because you knew that they were AI and so you could just appreciate them for what they are. Yeah. Like, it's nice that the Sora thing doesn't look exactly like me because it's like when you see a video with me in it doing something, you can clearly tell it's like a Sora video right now, right? Because Matt Hillary keeps making all sorts of Sora videos about me. I don't know if you've seen, but... Some of them are questionable contents. But I kind of like that it doesn't look perfectly like me. And I'm enjoying that for the time I can because I know that day is going to end where in the future, you're going to be unable to determine whether it's me or an AI-generated video. So yeah, I don't know. I don't know if slop is always the worst thing in the world, but I know some people hate it, that's for sure.

speaker-1: I just need like an identifier. Like if I'm like looking through videos or something I was just want to understand is it AI or is it not because I don't like when it's acting sneaky, you know, that's

speaker-0: Yeah, I mean, think like Gemini is trying to put like the little Gemini logo on theirs and Sora has like the Sora logo that jumps around. But the problem is like, they're like, it's so trivial to like build another AI that just removes that from the outputs that

speaker-1: Yeah.

speaker-0: that like there's always gonna, I mean, the same day that Sora released, later that day, there was websites where you could just remove it from it. So, it's not like it's like, it's not like, you you're trying to fight fire with fire or something like that, right? Like you need a better solution. And I mean, while I do think there's some solutions that are out there coming down the pipeline for this space, that will be great. Like cryptographically signing the images from camera to editor to editor to editor. You know, I think there's still gonna be more of the exception of the rule, right? People are just like, we have so much stuff going on. We're not gonna like go check that X 509 certificate on an image to be like, is this image valid or not? I don't know, you know, like.

speaker-1: want to look into this more from the forensics point of view, like for cases that need to go to court, like submitting evidence. I wonder how much this is affecting their world.

speaker-0: Yeah, that's a great question.

speaker-1: I mean, it's not a new problem, right? We've had Photoshop for decades and we've managed because you just have like... ⁓

speaker-0: Yeah, just think the level of expertise has been lowered to commit the crime, right? Like in the same way that like... Yeah, I don't know. Maybe we shouldn't go down there. Alright, so... I'm about to talk about a lot of fraud stuff, aka. So we're just gonna skip over that and go to the next thing.

speaker-1: I'm going to summarize our session today, how to become a doctor.

speaker-0: How to become a doctor when you're not. How to launder money without getting...

speaker-1: ⁓ yeah, that was my fault for that one. Funny. Well, I've got a trivia question for you. ⁓ Hey Bryce, where are the most pyramids located?

speaker-0: Hey, go. What do you got? Egypt and they are secretly nuclear reactors Have you seen that new that guy that says like there's like curl posts that go way down into the ground below it Yeah, it's a hole I mean when I when it gets late at night my entire feet becomes conspiracy theories So if you want to know about some pyramid conspiracy theories, but you're saying Egypt is not correct

speaker-1: I'm almost ready to tell you the answer. But first I need to know, you know, my friends invited me to a tinfoil hat night and it looks like a blast. So what you do is all your friends bring a PowerPoint about a conspiracy theory and we all make tinfoil hats. And then we share our presentations. And I just think you should be there, Bryce, because it sounds like you've got some good ones.

speaker-0: ⁓ yeah, ⁓ yeah, I go deep down that rabbit hole man, especially like, yeah, I'll get in the hot tub and I'll just like put on my headphones and I'll just be like, okay, it's time to take in whatever Joe Rogan's got to feed me today. You know what I mean? I'm just joking.

speaker-1: Okay, sorry getting back to pyramids there are more pyramids in Sudan than any other country in the world

speaker-0: Sudan? Really? I thought you were gonna be like, no, it's all South America or something like that.

speaker-1: Funny you know I don't know what their strict definition of it is in fact they might not even agree with themselves because I looked at the numbers for it so Egypt has about a hundred or so pyramids some people count it as a hundred as high as 118 and I'm like well if we're counting it differently are we categorizing them differently anyway, but yeah Sudan blows them out of the water with 255

speaker-0: Now isn't Sudan like physically pretty close to Egypt? Like you have Egypt, isn't Sudan like right below it?

speaker-1: You're putting me on the spot. Shoot, I don't know.

speaker-0: I think like, I'm just saying maybe back in the day that was all one country. I'm not a historian, so please don't take me for that.

speaker-1: It's so funny, I'm just gonna pull up Google Maps real quick, don't mind me.

speaker-0: I think like when you look at Egypt, to the left you have Libya, is that right? Yeah. then below it you have Sudan and to the right you have Saudi Arabia maybe? Very There's other stuff in there I forgot, I'm sure. But...

speaker-1: Mm-hmm That's so funny. Now you got it. But yeah, now you know. More pyramids in Sudan. So next time you want to go pyramid sightseeing, maybe check that out.

speaker-0: Yeah, I'm just gonna be honest.

speaker-1: I don't know if they compare it or not actually. ⁓ I'm sorry.

speaker-0: I'm gonna go Italy, go to a nice beach, hang out. And I don't think I'm gonna go across to the desert. I'm more of a beach person. I heard Turkey's got some cool beaches too, like along the Black Sea. Maybe that's gone down a little bit though. Cause Black Sea. What's up? Is it? Really?

speaker-1: Istanbul is beautiful. Istanbul. It's got some cool, my favorite is, it's kind of like real world meets Harry Potter just a little bit because underneath Istanbul, there are like cisterns and they've got like, it's like dark and they've got these pillars and a little bit of water. It's cool and you can walk. Really? You can go take a tour. The Chamber of Secrets.

speaker-0: Alright, I'll put it on the list. Sounds cool. I feel like every time I go to Europe it's just for work. So I need to do a proper trip to actually see stuff.

speaker-1: Yeah, a real vacay. You get that.

speaker-0: Yeah, yeah, I do like beaches. So it's just hard for me because I'm like, I could do that or I could just go on a cruise in the Caribbean Ocean. I was like, I'm going on a cruise. ⁓

speaker-1: Yeah, I get that. Well, you enjoy your trip. You have one coming up? Or was this a hypothetical trip to Italy?

speaker-0: No, I mean that's just like the bucket list thing. I definitely want to do that from a bucket list standpoint, but it's not on the calendar. So I am taking the kiddos to Disneyland. So if anybody wants to come to Disneyland with everybody else during the holidays, you'll find me there. I'll be eating the holiday treats, that's for sure.

speaker-1: Yay! Well, enjoy!

speaker-0: Last thing that I want to say. I really enjoy Sorter the app. I'm not gonna lie. I don't really like to admit that, but it's pretty entertaining to make videos on that app. ⁓ yeah, so if you haven't checked it out, you can actually now create a cameo for your pet. So if you have like a dog or a cat, you can actually create a persona for that. Or if you like draw something like arbitrary, you can create a persona for that and put it in your videos. So. I've been working on a turkey persona for Thanksgiving, you know? ⁓ Yeah, anyways, if you guys are interested in seeing some hilarious stuff, check out my Sora feed. That's another one of my pet hobbies these days. anyways, ⁓ thanks everybody for tuning in. Thanks Shelby, and we'll see you on the next one.

speaker-1: Stay safe. Sounds great, thanks, see you later.