Витік коду Anthropic Claude: чи прискорить це появу автономних AI-агентів?

Once again, it's been a really crazy week. There is a lot happening and it is my goal to separate the actual noise from the interesting and useful stuff for most people. And this being the week of April Fools, well, there was a lot of noise this week. Let's put it that way. So, let me break down what actually matters. I'm not going to waste your time. Let's dive right in. Starting with the biggest news of the week, which was the fact that Anthropics Claude code was actually leaked onto the internet. I believe this is the original source of the discovery here. Fried rice over on X. Claude Code source code has been leaked via a map file in their npm registry. And then they actually shared a link to where you can download the source of Claude Code. Now, Anthropic, a company that typically keeps what they're doing pretty close to their chest. But now that this was leaked, essentially people are out there sort of iterating off of it and treating it like open source and making their own versions of it. Enthropic is sort of going and trying to DMCA takedown as much as possible because this spread across GitHub like crazy, but I think the most interesting thing to come out of this leak is actually what we learned from looking at the code and what was actually in the leak. It kind of gave some sneak peeks into what Anthropic has been working on and what we can expect from the future of Enthropic. I think this Venturebe article here does probably the best job of breaking down what they found in the code. The leaked source reveals a sophisticated three layer memory architecture that moves away from traditional store everything retrieval. The architecture uses a self-healing memory system. So at its core is memory MD, a lightweight index of pointers that is perpetually loaded into the context. The index does not store data. It stores locations. While raw transcripts are never fully read back into the context, but merely gpped for specific identifiers. Now, GP is a word that essentially means that it searches. GP searches text for lines that match a given pattern, often using a regular expression. So, as you're talking to Claude Code, it's saving a bunch of information about what you're talking to it about, and then it's saving a reference to all of that within a memory file. And then that memory file is what gets loaded into context. And then it tells the context where to search for the thing that you're talking about essentially. I'm sure I'm oversimplifying this, but this is what I'm gathering from this information. But what's even more interesting in my opinion was this thing they found called Chyros. Chyros represents a fundamental shift in user experience, an autonomous demon mode. Now, a demon is essentially something that runs in the background of your computer and then waits for something to happen to trigger something else. So, this Chyros will allow cloud code to operate as an always on background agent. The agent performs memory consolidation while the user is idle. This expost I found probably does the best job of explaining this. Chyros is an always on proactive claude that does things without you asking it to. It runs in the background 24/7 while you work or sleep. Every few seconds, Chyros gets a heartbeat. Basically, a prompt that says anything worth doing right now. It looks at what's happening and makes a call, do something, or stay quiet. If it acts, it can fix errors in your code, respond to messages, update files, run tasks. Basically, anything Claude code can already do. just without you telling it to. But it has three exclusive tools that regular Cloud Code doesn't get. One, push notifications. It could reach you on your phone or desktop even when you're not in the terminal. Two, file delivery, so it can send you things it created without you asking for them. And three, pull request subscriptions, so it can watch your GitHub and react to code changes on its own. Regular Cloud Code can only talk to you when you talk to it. Chyros can tap you on the shoulder, and it keeps daily logs of everything, what it noticed, what it decided, what it did. He goes on here to give some examples of what this could mean. Think about what this means in practice. You're asleep and your website goes down. Chyros detects it, restarts the server, and sends you a notification. By the time you see it, it's already back up. You get a customer complaint email at 2 a.m. Chyros reads it, sends the reply, and logs what it did. You wake up, and it's already been resolved. Your Stripe subscription page has a typo that's been live for 3 days. Chyros spots it, fixes it, and logs the change. We're heading to a postprompting era. This is something I feel like I've been trying to describe for a little while now. I did a panel at CES and I kind of talked extensively about this during that panel that I think we're moving into the era where AI is going to be more proactive. You're not sitting there prompting it asking it to do things. It's going to learn what you do, what you want to use it for, and it's going to start to learn to be more proactive and do things on your behalf because it learns what you need it to do without needing to prompt it. the LLMs over time are going to fall into the background and become more and more of the plumbing and less of the thing that you're actually interacting with. And I think this Chyros here that Anthropic is working on now, it's not rolled out into cloud code yet. This is just something that was noticed in the source code, but this is something that feels like we're making steps towards that. We're also seeing similar things with OpenClaw when it runs cron jobs and it has the various heartbeat where it'll just kind of do things autonomously. And this really feels like we're moving closer and closer to that AI falls more into the background kind of thing. People also found more evidence of Capiara, which we talked about in last week's video. I think they've been kind of using the names Capiara and Mythos interchangeably. It's like the next level of model for anthropic. They even found what they call a hidden buddy system, which is a Tomagotchi style terminal pet with stats like chaos and snark. A lot of people are speculating that this was actually a planned April Fool's joke that likely got sidelined as a result of the leak. Personally, I think the most interesting takeaway from this whole thing is the Chyros, the sort of background agent, the demon that runs on your computer and just proactively looks for things to do so that it can go and solve problems without you needing to go ask it to. And that's where I hope AI is going. Honestly, I don't want to have to sit and have conversations with it. I want it to know what it's supposed to be doing and just go do the things it's supposed to be doing. Boris Churnney here, who I believe is like the head of Claude Code, basically said that this happened because it was just a developer error. He went on to say, "Mistakes happen. As a team, the important thing is to recognize it's never an individual's fault. It's the process, the culture, or the infrastructure." In this case, there was a manual deploy step that should have been better automated. Our team has made a few improvements to the automation for next time. A couple more on the way. So, it doesn't sound like anybody was really blamed or getting fired for this. And Enthropic doesn't even seem to be making that big of a deal. There hasn't really been any sort of public statements about it. No like private information about any customers or anything like that was leaked. But I do have a feeling we're going to see some spin-offs of Claude code popping up onto the market really really soon from people that leverage the source code and then sort of spin up their own versions. We've got another visually stunning AI model update this week from ReCraft. If you haven't heard about them before, they're an AI tool built specifically for professional aesthetics and agency quality design assets. I've used them before myself, and they just recently launched the Recraft V4 family of models. So, this upgrade is really focused on delivering elevated art directed outputs for creative professionals. So, for example, instead of just generating one-off images, the new FI4 models are designed for things like cohesive brand visuals, website prototyping, product packaging, logos, and vector illustrations that are ready to use. And there are actually two new models under the V4 release. The V4 and V4 Pro models, which are better for photorealistic scenes, mockups, detailed illustrations, and more complex compositions that actually understand lighting, materials, and negative space like a real art director would. and the V4 vector models which can generate native vector graphics. So things like logos, icons, and illustrations that are actually scalable and editable SVGs instead of just faking the look of vector art, which is what usually happens when you try to use AI to generate this kind of stuff. There's strict prompt adherence, great aesthetics, and it handles text on images surprisingly well for AI. I've mentioned before how to me that's one of the biggest tests if an AI image model is good or not. Because even with the advanced models today, I still see some wonky text on AI images. And well, that won't cut it for professional-grade brand assets. You can also generate things like photorealistic product mock-ups with lighting, shadows, and packaging that looks like a real product shoot without, you know, the cost of a real product shoot. This is a really great tool if you're a designer, a marketer, a content creator, or really anybody building visual assets for a brand. It just feels more production-minded than most of the AI image tools we usually see. So, thank you to Recraft for supporting my channel and sponsoring this portion of today's video. And if you want to learn more and try it out for yourself, head to the link in the description box. This week, OpenAI raised $122 billion at a valuation of $852 billion. This is the largest raise by any company ever. And it also makes OpenAI pretty much the fastest growing company in history. They even brag a little bit about it in their post here. At this stage, we are growing revenue four times faster than the companies who defined the internet and mobile eras, including Alphabet and Meta. They are now generating $2 billion in revenue per month. There's been a lot of talk about a falling out between OpenAI and Microsoft. And well, this new round that they just raised, guess who's listed here? Microsoft is in on this raise as well. So, they're still pumping money into OpenAI despite all of the speculation that there's been a big falling out between the two. Now, this blog post that they put out, which mainly focuses on the giant rays that they just had, if we scroll all the way down to the bottom here, this is what's really interesting to me is they go into some details here about what they're doing next. So, we learned where Anthropic is going because of a leaked source code. We're kind of learning where OpenAI is going because they buried it in this fundraising article here. They say, "We're building a unified AI super app. As models become more capable, the limiting factors shift from intelligence to usability. Users do not want disconnected tools. They want a single system that can understand intent, take action, and operate across applications, data, and workflows. Our super app will bring together chat, GPT, codecs, browsing, and our broader agentic capabilities into one agent first experience. They don't want you to have to go to use the Atlas browser for browsing the internet and chatgpt for having chats and codeex for coding and jumping from app to app to app. In fact, it feels like they're kind of modeling Enthropic here a little bit cuz Enthropic has their Claude app, but their Claude app has the Claude chat in it, Claude Co-work, and Claude Code all within a single app that you can use. It kind of feels like OpenAI is sort of going in that direction. And I wouldn't be surprised if they also follow in the footsteps of Enthropic with a more proactive in the background AI agent as well. I mean, after all, they did bring on board Peter Steinberger, the guy who created OpenClaw to be part of the OpenAI team. So, they're obviously thinking in that agent sort of mindset and that let's build things that just do what we want them to on our behalf before we tell them we need it to do it. That's also the reason why they phased out Sora. They seem to be skipping the adult mode that they were going to work on. and they're really sort of focusing in on their core competencies which are really really intelligent models for both you know chatting and research but also coding and agentic use cases. As a side note, I find this really really interesting. This is an article from the Wall Street Journal. It breaks down how Sora got shut down last week, something that we went in depth in in a previous video. But one thing that I found fascinating about this article is that Sora was losing roughly a million dollars a day. With that context, it's pretty obvious why they had to shut it down. It was losing over a third of a billion dollars a year. We got some news out of Microsoft this week as well. They released a brand new transcription model called MAI Transcribe 1. It's a new speech recognition model achieving best-in-class accuracy across 25 languages, which is purpose-built for transcription workloads. We can see here on this test for WER, which stands for word error rate, that this new MAI Transcribe 1 significantly outperforms GPT Transcribe, Scribe V2, Gemini 3.1 Flash, and OpenAI's Whisper model by quite a bit. It also apparently works really well in noisy environments. And well, I did get early access, so let's test it real quick. We can see here inside the MAI playground, we have access to MAI image 2, which we did play around with in a previous video. the MAI voice model, which is Microsoft's texttospech model, and their brand new MAI transcribe one model. So, let's go ahead and select this. I can see the sea from here, but I can't hear my friend because the wind is louder than her playlist. She yelled, "Come here." And I shouted back, "I'm here. I just can't hear you." Then she said, "Our snacks cost $8." And I replied, "I ate the evidence already." A dad nearby told his son to put on sunscreen. And the kid asked if that meant he had to block the sun, too. The sea kept sparkling like it could see the whole comedy show. All right, let's go ahead and submit this. Notice I used a lot of homonyms. Is that the right word where it's like the same word but different meanings? Let's see how it did. I can see the sea from here, but I can't hear my friend because the wind is louder than her playlist. She yelled, "Come here." And I shouted back, "I'm here. I just can't hear you." Then she said, "Our snack cost $8." and I replied, "I ate the evidence already." A dad nearby told his son to put on sunscreen and the kid asked if that meant he had to block the sun, too. The sea kept sparkling like it could see the whole comedy show. So, it knew the difference in words that I was saying because it, I guess, understands all the context of what I'm saying as well. It was also announced this week that the various Microsoft tools are now available inside of Microsoft Foundry. So, developers and people that want to build on top of these models now have access to them as well. So, you can now use the MAI Transcribe 1, MAI Voice One, and MAI Image 2 models directly inside of whatever you're building Microsoft Foundry. Google this week rolled out a new version of their video model, VO3.1, now releasing VO3.1 Light. This is a more cost-effective video generation model. I'm sure the trade-off is in quality, but it is cheaper now to generate videos. We can see the pricing here. VO3.1 is about 40 cents per standard video with audio. VO3.1 fast is 15 cents and VO3.1 light is only 5 cents and that's for 720p. Now it also looks like 3.1 light will not generate in 4K. So you do need to use 3.1 fast or 3.1 if you want the 4K. But you know for most things 1080p is probably going to be fine. And then you can always just use an upscaler later if you want. Another thing I want to point out about this real quick is that Logan Kilpatrick from Google, he put out this expost which introduced VO3.1 light. But he also mentioned that on April 7th, we're also reducing the price for VO3.1 fast. So these prices are all going to come down as well on April 7th. It also makes me wonder if there's other news coming out around April 7th. I don't have any insider info on this, but who knows? Maybe there's something else coming out around VO on April 7th. And with that release, they're dropping the price of other models. Maybe. Don't quote me on that. It's just me, you know, maybe speculating out loud a little bit. And while we're on the topic of Google, they released a new feature called AI Inbox. This AI Inbox has smart prioritization and daily personalized briefings. Now, the reason I didn't talk about it earlier in this video is because unfortunately, as of right now, it's in beta and only for Google AI Ultra subscribers, which is their plan that goes for $250 a month. I mean, you can get it for $125 a month for the first 3 months, but, you know, typically $250 a month. Now, I do have a Google Ultra plan for testing purposes, and this is what it looks like. Unfortunately, I'm probably going to have to blur most of this cuz it's, you know, private. So, instead, here is their screenshot of what that looks like. You've got your suggested to-dos at the top based on your inbox, and then you've got your topics to catch up on down here. And that's exactly what my inbox currently looks like. We got a handful of new large language models this week as well. Probably the most notable being Gemma 4 from Google. Now, this is an openweight, open-source model that's available for you to run on your device, run in the cloud, fine-tune, do whatever you want with, and it's a really powerful model. We can see they released it under the Apache 2.0 license and it's designed to be run on Android devices and laptop GPUs. So, if you have a more modern device, you should be able to run this model. If you do have something like the Nvidia Spark or Apple Mac Studio with some decent memory, it looks like this is also designed to run your agent. So, if you're running like an OpenClaw or something like that and you want to run it all locally, this new Gemma model looks like it's going to be great for that sort of thing. Here's a quick peek at the benchmarks for this model. Really only comparing it to past Gemma models, but from the looks of it, a really, really solid model, especially for something that's open weight and that you can fine-tune and run locally. We got not one but two brand new models out of Alibaba this week. We got Quinn 3.5 Omni, which is an omnimodal model. So, it works with text, image, audio, and video understanding. And they describe a standout feature that you can use this for. audiovisisual vibe coding describe your vision to the camera and Quinn 3.5 Omni Plus instantly builds a functional website or game for you. They claim it outperforms Gemini 3.1 Pro in audio and matches its audio visual understanding. And here's a peek at the benchmarks for this one. Mostly just comparing it to the older Quinn models, but it does compare to Gemini 3.1 Pro here. Pretty much beating every other model in all of these benchmarks here. But then just a few days later, Alibaba also released Quinn 3.6 Plus, which appears to be a model designed for realworld agents, designed for next level agentic coding, enhanced multimodal vision, and has a 1 million token context window, which is available by default, so you don't have to turn on a special higher token limit. If we take a look at the benchmarks on this one, we can see that the dark purple is this new model here, and they're comparing it to 4.5 Opus. So, not the newest Opus model. They're also comparing it to Gemini 3 Pro and apparently not Gemini 3.1 Pro. So, they're kind of comparing this to like the last generation of models from some of the Frontier Labs. But on Terminal Bench, it beats these other models. SWEBench Pro, the software engineering bench, it's pretty much on par with Opus 4.5. The Quinn Claw Bench, which apparently is a benchmark created by Quinn, I would guess. It's really good at that one. really good at real world questions and answers, really good at video reasoning, and really good at document recognition and understanding. So, a new really, really solid model out of Alibaba. This model does not appear to be an openweight, opensource model, but they do mention in this article that they will be open sourcing more models soon. And we got a new model out of a company that, well, this is my first time actually hearing about them. They're called RC and they released a model called Trinity Large Thinking. And this is also an open-source model released under Apache 2.0. And it's another American company releasing open-source models. Taking a look at the benchmarks here, they're comparing it with Opus 4.6, Kimmy K 2.5, GLM5, and Miniax M2.7. And it seems to be pretty on par with most of those models, but not really beating out any of them except for in this TOA2 airline, which not super familiar with this benchmark myself. They do show it off in use, creating a snake game, doing what appears to be some like agentic work, working through a whole bunch of different steps. Here we see it doing some code insanely fast. I don't know if that's real time or not, but it's building some code and building out like a monthly budget planner app. And then we see a maze screen saver that it generated that uh reminds me a lot of like Wolfenstein 3D maybe. Again, what's really impressive about these is that these are openweight models. These are models that you can download, run on your own systems, run in the cloud, fine-tune, iterate off of, do what you want with, and they're quickly approaching the state-of-the-art models. Now, it feels like the closed source models are kind of always a little bit ahead, a few steps further from the open models, but the open models just keep on gaining ground, it feels like. And those were the new LLM releases this week. I've got to be honest with you, I have a really hard time sort of testing and experimenting with the new LLMs that come out because for the most part I'm able to do kind of everything I want to do with the existing models. Like I've got aentic systems running using the OpenAI and the anthropic models. I've got models running locally with my Neotron model. It helps me brainstorm. It helps me think of title and thumbnail ideas. It helps me with all of the various use cases in my life. And so I really struggle to think of new ways to really test and find ways to prove that these models are getting better. I would love to have like my own benchmark that I'm constantly testing these new models against, but I just don't really have any ideas cuz my main goal is to show off how these models are useful to the most amount of people to like everyday use cases in your business and your life. And for the most part, these models are kind of already doing what we need them to do. So I'm looking to create some sort of like everyday user benchmark. And so if there's somebody out there that wants to help me create that benchmark or work with me to help better test these large language models as they come out, hit me up over on X or LinkedIn and uh let's chat cuz I would love to figure out some sort of better way to test and sort of benchmark these models for normal everyday use for people. All right, I've covered a lot, but I do have a few more things I want to cover, but I'm going to break them down real quickly. So, let's jump into a rapid fire. Let's start with a new feature update out of Claude here. Last week we talked about Claude's computer use feature where it can actually take control of your mouse and keyboard and do things on your browser. Well, they've now added that feature into Claude Code. You can preview that feature if you're on one of Claude's Pro or Max plans. If you're not familiar with that feature, check out last week's news video because I did do a demo of it in that video. Speaking of Claude Code, this is something I found really interesting. I would not have thought Enthropic would have even allowed this, but apparently there's a codeex plugin for Claude Code. If you already use Claude Code, the Codeex plugin is a simple way to bring codeex into the same workflow. This essentially allows you to use your OpenAI key inside of Claude Code and allow Codeex to review code and actually delegate tasks. So, if you just really love the UI of Claude Code, but you also want to use codecs, well, this now allows you to sort of blend them together. Speaking of OpenAI, ChatGpt is now available inside of Apple CarPlay. So, if you've ever wanted to have conversations with Chat GPT from directly inside of your vehicle, well, apparently now you can. And in news that I would have sworn was an April Fool's joke if, well, it didn't come out on April 2nd, OpenAI acquires TBPN. TVPN is like the ESPN of Tech Ed Business. I believe it stands for Tech Business Production Network and they do a daily show that's like a three-hour a day show where they bring on founders and CEOs and thought leaders and do little quick interviews and break down all the news and they do it as a live stream over on X at YouTube. And it's just wild to me that OpenAI would be going and acquiring essentially a daily live podcast, especially after all of this talk of not doing side quests. Now, I think it's awesome for the TBPN guys. It's a great show. I actually tune in to make sure that I'm not missing anything in the news. It's just not something I would have ever seen coming. I wonder how this is going to change the show editorially. Like, are they allowed to still make jokes about OpenAI when OpenAI does something a little weird? Are they still going to be able to interview competitors at like Anthropic and Google and places like that? I don't know. Seems like a very interesting choice on the part of OpenAI. Something we'll see how it plays out. Just really wild. This was posted on April 1st. I would have guaranteed it was an April Fool's joke. Perplexity released a new use case for its computer model called computer for taxes. Perplexity computer loads tax modules for US federal tax questions. computer can help with hard parts of tax work that usually cost people the most time and money. This can mean helping you draft a federal income tax return on official IRS forms, review a professionally prepared return, build dashboards and tools for more complex parts of the tax code, or supporting any other workflow that you want to design. Now, I haven't tested this myself yet cuz, well, I'm procrastinating and I haven't done my taxes yet. But maybe I'll run my taxes through this and see how it does. I do have a CPA who I'm probably still going to trust a little bit more, but maybe I'll run my taxes through this and see how closely it compares to what my CPA came up with. If you're a fan of using Slack, Salesforce, the company that owns Slack, is making Slack more agentic. Personally, I'm a big fan of Slack. I do all of my work inside of Slack. Even my open claw is built for me to talk to it through Slack. So, this is something I'll definitely be playing with. But here's what they say about it. This week, they announced 30 new capabilities that take Slackbot from personal agent to the ultimate teammate. There's a new meeting transcription and note-taking capability, reusable AI skills, so you can define a task just once. It's a new MCP client. Businesses now have native customer management built into Slackbot. It reads your channels, understands what happened, and keeps your deals, contacts, and call notes up to date automatically. There's new skills like deep research, and voice input. and it looks like it'll record meetings for you even if the meetings aren't in Slack. We can see here that it's recording meetings that are happening over on Zoom. They show here they're giving it a screenshot and then asking it to summarize the feedback from the screenshot. So, a bunch of really cool features that are getting added to Slack that will allow Slack to do a lot of the like chatbot stuff you already do, just keep it inside of Slack. There are a ton of new features rolling out, so I will link to this article in the description as I do with every news video I put out. Right now, it's only available on business plus and enterprise plus plans, but it says starting in April, we're opening it up. So, free and pro plans will actually start to get access pretty soon as well. This is something I came across this week that I just found really interesting. GM says AI helps visualize a car before it exists and speeds up how it gets built. They're using AI systems to transform early handdrawn sketches into concept videos, suggest design changes, and even run early stage aerodynamic testing. The result is a dramatic compression of the design timeline. I don't know why I found this so cool, but we can see here they were sort of mocking up a design and then they show this design in the real world a bunch and then they even turned their mockup designs into actual videos so they can kind of see what this car would look like driving around in the real world. And they're using AI models to simulate the aerodynamics and sort of pre-igure out how efficient these vehicles are going to be using these AI models in advance. I don't know. This is just stuff that I find fascinating and I figure at least a handful of you will find fascinating as well. And finally, I'll end on this one. They're putting AI in shopping carts. Instacart introduces physical AI smart carts to influence grocery shopping in real time. So, shopping carts in the future might look like this. A touchcreen on the front, outward-facing cameras, basket facing cameras to see what's in your cart, location sensors to know where you are within the grocery store, and also weights and measures and certified scales inside the cart. All powered by Nvidia Jetson. When a retailer deploys a caper cart, it's not just capturing data. It's processing what's happening in the aisle in real time directly on the cart. That includes what's in a customer's basket and precisely where they are in the store, said the company. So retailers can influence the shop while it's happening, not after checkout. Mistimed recommendations don't just fail. They train customers to ignore the next one at Instacart. I don't know why I find this kind of funny and I feel like people are going to sort of reject this concept. Like when we go to the grocery store, we kind of want to just go and get what we're there for. I don't know if I really want to be recommended additional products by my shopping cart, but I don't know. Maybe I'm just weird like that. Maybe people really want their shopping cart to sell them more crap. I don't know. Anyway, I just thought that was kind of a random one to end on. But yeah, that's what I got for you this week. I'm doing my best every single week to sift through all of the information, all of the AI news, all of the new AI tools that are coming out every single week, and break it down into just what you need to know in one video a week. I used to try to make a new AI video every single day about news that was coming out constantly, but I started to feel like that might get a little too overwhelming because there's so much AI news coming out. And I figured, well, maybe the best thing I can do for people is I'll drink from the fire hose all week. And then in just one video per week on Fridays, I'll break down what I think most people will be interested in or most people will want to know about what came out in the AI world. I'll keep myself overwhelmed so that you don't have to be overwhelmed. That's the goal of this channel. Now, I do put out some other videos here and there in between with some commentary and some breakdowns of things that are happening, but for the most part, my goal is to make that one video a week that helps filter through all of the noise and just gives you the signal of what's important to know right now. And so, if that's something that sounds interesting and appealing to you, and you just want that one video a week to stay completely looped in, maybe consider liking this video and subscribing to this channel. It really helps me out a lot. I'm doing that push to a million subscribers. I'm hoping to get there this year. So, every new subscriber really, really helps, and I'm going to keep doing my best to keep you as looped in as possible with as little overwhelm as possible. Now, I do record these videos on Thursdays and publish them on Fridays. So, if there's any news that I missed, I'll try to get it into next week's video. This week, I'm actually recording most of this one on Wednesday and some of it on Thursday cuz I'm going to be out of town this weekend. But, usually I record on Thursdays, publish on Fridays. But again, anything I missed will be in next week's video if it's important enough that I think you need to know about it. Thank you again so much for tuning in, hanging out, nerding out with me today. I really, really appreciate you, and hopefully I'll see you in the next one. Bye-bye. Oh. Uh oh, I'm back. Oh. Oh, I'm finally out of that egg.