Well, there are some pretty big headlines today in AI news. More and more people are becoming concerned with Anthropics mythos model. Treasury Secretary Scott Bessant and Federal Reserve Chair Jerome Powell are among the people that are extremely concerned. They had an emergency meeting with the leaders of Wall Street to talk about this new era of greater cyber security risk. They're saying this might be one of the biggest risks facing the financial industry. There's also been a story floating around that OpenAI has a very similar model to Mythos that they will not be releasing. That just like Mythos, it'll be released to a small group of companies that are supposed to test it because of its advanced cyber security capabilities. A lot of people assumed that this was Spud, the much awaited model from OpenAI. There's tons of misinformation about this thing, so we'll clear that up. It's mostly not what it appears to be. It's largely false. Amazon's doing well today because the AWS CEOs of the Enthropic's latest models including Mythos are being trained on the Tranium chips. So is relying on AWS. They're using the

trainium chips. Sounds like both for training and inference. It's doing a lot with Google's TPUs. It's supposed to receive a shipment for their own use later in 2026. And this is the most recent news. It's seeming like they're considering designing their own chip. So, still not a lot of news on that, but as AWS reported, they are learning a lot from Anthropic about how to design chips specifically for models like Claude. All right, so let's break some of this down. So, first and foremost, the whole OpenAI not releasing their latest models because it's as powerful as mythos. That does not seem to be true. Some of those things are technically correct, but the the story as it was presented, we can say it's false. So, here's Dan Chipper. Thanks to him for correcting the story. He's saying he just spoke to OpenAI and this is actually false. OpenAI is working on a cyber product with a trusted tester group. But this is not related to Spud, their newest model. Unfortunately, seems like the Axios story conflated the two and now has been updated. So, as far as I can tell, the Spud release is imminent. That's what all of them are kind of hinting on. They dropped the Chad PT, the $100 plan

today. That was the small bit of news that they were going to release. And there's also the big bit of news that is coming soon. I assume that's going to be the Spud model. And if you can't tell what this is a picture of, that's that's the opening eye logo with potatoes for the eyes. Those are potatoes. This this took me a second to understand what I was looking at. But let's talk a little bit more about mythos. There's still quite a few people out there that are saying that it's complete nonsense that Anthropic is just trying to drum up investor interest. And I'm not surprised because we've been seeing this sort of push back on any AI progress since the very early days. I've been coming this for a few years now. And every single time there's a new breakthrough, there's some subsection of people. They're like, "It's just nonsense. It didn't happen. It doesn't exist. AI can't do anything. It's just a next token predictor. It's just a stocastic parrot. It's just blah blah blah." Whatever. Whatever the case is, it seemed to have really spooked Wall Street, specifically leaders of these banks. and regulators consider the possibility of a new breed of cyber attacks as one of the biggest risks facing the financial industry with Anthropics mythos being a more powerful system capable of identifying and

exploiting vulnerabilities in major operating systems and web browsers. Okay, but what do they know, right? They're just bankers. What do they know about cyber security anyway? I'm being sarcastic, by the way, for people that can't tell. But here's somebody that might know a thing or two about cyber security. By the way, big thanks to AI explained a channel here on YouTube for flagging this in the video. I kind of missed it. So, this is Nicholas Carlini. He is in the cyber security computer security field and he knows a lot about it. This guy might be one of the top researchers in that field. Massive amounts of works published, massive amounts of awards. He previously worked with Google at Deepmind and now is in anthropic. Okay, so he's like near the top of his field. This is him talking about his experience with this mythical model mythos. >> It has the ability to chain together vulnerabilities. So what this means is you find two vulnerabilities, either of which doesn't really get you very much independently. But this model is able to create exploits out of three, four, sometimes five vulnerabilities that in

sequence give you some kind of very sophisticated end outcome. And we think that this model can do this really well because we notice that this model is very autonomous. It's just generally better at pursuing really long range tasks that are kind of like the tasks that a human security researcher would do throughout the course of an entire day. Obviously, capabilities in a model like this could do harm if in the wrong hands. And so, we won't be releasing this model widely. Working with our partners, we've been finding vulnerabilities across essentially every major platform. I found more bugs in the last couple of weeks than I found in the rest of my life combined. So just to beat this dead horse just a little bit more this is one of the worldclass top of the field experts slash researchers in cyber security. So I believe he graduated Berkeley in 2013. So a decade plus experience behind them. So this person is saying that he's found more bugs in the last couple of weeks than he did before in his entire life. So it seems like these AI models have crossed the best in the world humans ability to find these vulnerabilities. And they do so autonomously. They just sit there and

they crank out these things, these vulnerabilities, and then they create exploits, somehow apparently chaining multiple exploits together to just really stick it in there. They've described various exploits in the Linux systems, for example, where you're able to get any user and upgrade them to the the super user, the administrator. One of the other bugs on an open- source operating system that's been around for 27 years is considered very secure. And apparently, you can send it some data, some information, and just crash the whole thing. Hey, quick aside. You know what just happened to a lot of the AI dev workflows? Claude apocalypse. If you've been relying on your cloud subscription to run OpenClaw, you've probably noticed that Enthropic is now blocking those workflows. They are actively shutting the door and suddenly everybody on that setup is scrambling to figure out how to keep their AI agents running, you know, without completely ripping out the infrastructure and just starting from scratch. That is why this is actually interesting. Thanks to Kilo for sponsoring in this segment of the video. If you're locked out of OpenClaw right now, then Kilo Claw is hands down

the fastest and most secure way to get your OpenC environment back up and running. But it's not just a band-aid. It's a fundamentally better way to work. Instead of being bottlenecked by one provider's subscription tier, Kilo Code hooks you into their gateway that gives you instant access to over 500 underlying models. We're talking Frontier models, budget models, and even completely free models. They have a lot of models. There are no rate limits. There's zero markup on AI tokens. You pay exactly what the compute costs. That's it. And you can access it all across VS Code, Jet Brains, or your CLI using the exact same account. But I know what you're thinking. You don't want 500 models. I just want one that actually works. This This is where it gets really interesting. Kilo built a benchmark called Pinchbench specifically to test how different models handle actual open claw tasks, not just generic leaderboard trivia. Notice right now RC's latest model, Trinity large thinking is

performing nearly on par with Anthropics Claude Opus 4.6, their latest and greatest model. You just swap it in through Kilo and just keep moving. And if you don't want to manually switch models between planning and coding, Kilo's automodel selection will dynamically route your tasks to the right model for you, including free options so you're not burning cash on simple tasks. Giving a semi-autonomous software, open access to your files, your API keys, and your shell just hoping it behaves. Maybe not a great strategy. Kiloclaw is actually safer than running agents locally. They recently went through an independent 10-day security assessment and published the white paper to prove it. When you're using Kilocclaw, you're not exposing your local machine. There's a number of security layers protecting you. First of all, each user gets their own firecracker virtual machine with five independent layers of isolation. If an agent goes sideways or hallucinates a disastrous shell command, the blast radius is contained securely within that

virtual machine. That is why Kilo is worth paying attention to, not because it's another shiny coding agent, but because it gives you a great way to run OpenClaw again to break out of that single model dependency and do it in a way that's a lot more flexible and security aware. So check out Kilo and Kilo Claw at the link down below. I'll have the link posted in the description and pinned comment. So claw on and let's get back to the content. before releasing this model internally anthropic I believe February 23rd I'll double check that date but that was the first time that this model sort of appeared internally within anthrop even the researchers at anthropic were a little bit spooked by it there was discussion whether or not it was safe to even release it internally and the reason for that is because this does seem to be a step change the jump between the previous models and mythos is a lot more sharp than everybody anticipated so this is the epoch capabilities index ECI So this basically synthesizes a lot of different benchmarks into one. The idea is to kind of be able to look at one chart to see

the relative strength of these models by kind of taking all their abilities and and just picking it into one metric. And so Anthropic has their version of this, the Anthropic ECI. So this is powered by their own internal benchmarks. So depending on how you kind of draw the slope here, depending on what your baseline is, but let's say it kind of goes like this. The jump to Claude Mythos Preview is noticeable kind of the slope, the trajectory of capabilities of these models before Cloud Mythos, it seemed to have shifted upwards. Enthropic is not sure why exactly. So the slope measurement tells us that Anthropic's capability trajectory bent upward in the period leading to Cloud Mythos preview. It does not on its own tell us why. Interestingly, they mentioned here that they surveyed technical staff on the productivity uplift they experienced from Claude Mythos Preview relative to not using AI for work at all. They're saying the distribution is wide and the geometric mean is on the order of 4x. So there's a 4x uplift in productivity when you use Claude Mythos Preview. Interesting. In the initial weeks of internal use, several specific claims were made that

claims preview had independently delivered a major research contribution. It's kind of a weird thing to hear, right? So they released this model internally to anthropic researchers, people that are very familiar with AI. And a number of those researchers are like, "This mythos model made a massive major research contribution, some some breakthrough." And they continue. When we followed up on each claim, it appeared that the contribution was real, but smaller or differently shaped than initially understood. So I thought that was kind of an interesting thing. I'm not sure exactly what to make of it because again, these are people that work with AI every single day. So the fact that their researchers anthropic worldclass people in terms of working with AI, doing AI research. So they each kind of independently raised their hand and went this new model delivered a major research contribution independently from them. So I'd be very curious to kind of to see what exactly happened there. Is it possible that just these big jumps in in capabilities kind of stir this sense of awe in us? We we feel like it's almost magic. One very interesting thing that popped up in the alignment risk update for this model was

this little line that seemed like there was a technical error that allowed a reward code to see the chains of thought. So these were certain training environments for mythos for this model. Now OpenAI a while back we covered a paper that they did and did a whole video about it talking about basically where they're looking at the chains of thought of these models. So you can think of it as, you know, well, either kind of their thoughts, but I think it's more accurate to think of it as like a little scratch paper or notebook in which they write stuff out. They do their planning in there before actually saying out loud what they're going to do. So they kind of think at length in their chains of thought and then actually say out loud what they're going to do and take that action. And often times we can kind of spot various infarious activity that they're planning to do in their chains of thought. Right? So they're thinking about some particular hard problem. They're like, "Oh, I wonder if I can fudge the numbers here. I wonder if I can to hack this thing somehow or cheat to get a better score. Right? You kind of see him thinking through it and they're like, "Yeah, I'm going to go ahead and and do that. That's the right way to do it." And then they go ahead and they they cheat. They take that bad action, so to speak. And one of the things that the opening eye researchers did was they

thought, "Okay, what if we penalize it every single time those bad thoughts appear." So every time things, I wonder if I should cheat in this situation or I wonder if I should do this bad thing to get around some sort of a guard rail or some sort of rules. Like every time that happened, we kind of smack it in the head and go, "No, don't do that." Well, an interesting thing happens because those bad thoughts, well, they do go away. However, the bad behavior does not necessarily go away. So now, instead of seeing kind of having a a warning that they're going to do something bad before they do it, that disappears, but the bad behavior still happens. So that might be because it gets pushed deeper into the laten space. That thinking, that planning takes place elsewhere. Whatever the point is, OpenAI kind of warned the other labs not to do this. And of course, everybody knows this. Anthropic, I'm sure, has their own research finding exactly this. It seems like it's kind of accepted, understood that you don't want to train on the chains of thought. You don't have any positive or negative reinforcement for that. It's kind of like having a kid that constantly gets punished for telling the truth, right? So, every time he comes up to the parents and says, "I did something bad."

And he gets punished for it. Well, eventually he's going to stop coming up to the parents and saying, "I did something bad." Is that because he stopped doing bad things period or just because he realized that saying that out loud is not a good idea for him? So it seems like anthropic they don't train against any chains of thought. So they don't do that and they don't train against activations based monitoring. So if a certain region of the brain lights up that's associated with deception and by brain I mean these large language models they're neural networks certain clusters of neurons that light up for a particular thing that we learned. Well that tends to correlate with them doing something negative something deceptive. Well, again, like training against that. The question is, well, what happens if those signals go away? Because for us, that's kind of a signal that they might do something bad. It's a way for us to have interpretability into what they're doing, some sort of an insight in what they're doing. If those things go away, then we might not know what's happening. Just like that kid that no longer ever says they did anything wrong. But apparently here, because of a technical error for Mythos, they did do that with this model. So this issue affected 8% of

reinforcement alerting episodes and was isolated to three specific subdomains of our environment. So guey computer use. So we'll we'll come back to that. So that's kind of like the visual elements of Windows or Mac OS. It's like the stuff that you click on also office related tasks and a small set of STEM environments. We are uncertain about the extent to which this issue has affected the reasoning behavior of the final model. But it is plausible they had some impact on opaque reasoning or secret keeping abilities. This has always been so fascinating to me because from kind of early on it seemed like the less understanding and visibility we have into how these models work, the more effective the training is, the better the training potentially could be. It was kind of this theme that was developed. So if you do a lot of RL training, reinforcement learning training, you might see the models develop kind of their own shortorthhand, their own language for thinking through stuff. So they're able to do more with kind of less tokens. But us humans don't quite understand as much what they're saying. So there does seem to be this kind of a trade-off between visibility to what they're doing and how quickly

and effectively you can train them. It's always curious to me like what happens if let's say some lab finds that doing reinforcement learning training against these chain of thoughts or activation based monitoring like it's really effective at improving the model's ability and at the same time you realize that the model stopped coming up to you and telling you hey I did something wrong or you lost any visibility to detect any wrongdoing what would that look like I'm not saying that it happened here I'm not saying that but they they did this thing here where they train against the chain of thoughts at least for 8% of the RL episodes and it does seem like there was this massive improvements like the slope went upwards like it's dramatically better than before that rate of progress improved. So again, you know, we have this massive massive leap up in abilities and at the same time Claude Mythos Preview is the best aligned model that we have released to date by a significant margin. We have just won the lottery. We have the most best line model ever. It's also a

massive massive leap on on everything. We we nailed capability and alignment just in one model. This is just fantastic. This is just fantastic. We're just just winning. We're winning on on every front. But they they also had this oopsy moment, this technical error, and they're uncertain about the extent to which this issue affected the reasoning behavior of the final model. Like that that big jump. Did this have an effect on it? So, they didn't know if that's the case, but it is plausible that it had some impact on opaque reasoning or secret keeping abilities. So, just so I'm clear, let me kind of spell it out. If something like this created a model that was able to keep secrets, reason opaquely, meaning it would be able to reason in a way that we couldn't follow, we couldn't see its thoughts. Like, it got really good at not letting us know what it's thinking. And at the same time, it became kind of like super smart. What would that look like? What would the researchers see? How could they tell that it's a secret keeping abilities just improved and so did it intelligence in in capability? Well, one, there would be a massive capability

jump that slope inflecting upwards kind of a more more sharply than before. And at the same time, this would be also coincidentally the best aligned model that they have released to date by a significant margin. So, I'm not saying that that's what happened here. I am saying that if it did happen, this is what it would look like. By the way, it seems like Anthropic is going to great lengths to make sure that these models don't have some hidden messages in their chain of thoughts, that there aren't some hidden meaning that humans can't read. In this section called ruling out encoded content in extended thinking, they're saying that especially for very powerful models or models trained with extremely large amounts of reinforcement learning, we may expect to see scratch pads start to encode information steographically. Meaning they'll encode additional reasoning in the syntax of the text. Like if they're using a bulleted list versus a numbered list or using some weird punctuation to basically pass along some information that if you were to read it, you would never pick up on. So, Enthropic goes to great lengths to make sure that there

isn't extra data or extra information, extra context that isn't being passed along that wouldn't be obvious on the surface to us if we were reading it. So, some very interesting stuff happening there, but apparently this is only the beginning. So, here's Posh. He's working on Codeex at OpenAI. He's saying that things are about to get wild and then an arrow pointing to April 9th, which is today, that's saying you are here. So, are you ready for things to get wild? I was pretty sure that things are already pretty wild, but apparently we've seen nothing yet. Let me know what you think about this whole Federal Reserve meeting about Mythos. What do you think that makes mythos so special? What was the secret sauce? Why was there such a big change in capability as well as the fact that it's the most aligned model ever? Let me know in the comments if you made this far. Thank you so much for watching. My name is Wes Rolf. See you in the next