AI News: The AI Slowdown: What You Need to Know

There's a lot to talk about in the world of AI this week. Hey, what are you doing? Oh, shoot. Sorry. It has been a hectic week and I'm going to do my best to break it all down for you. So last week, the narrative in the AI world was all about these people coming out of Anthropic and OpenAI saying they're worried about the potential of what AI can do. And there was this Huge risk of it destroying humanity. Well, if that was the main story of last week, the main story of this week has to be all of the AI labs responding to that story. Over the weekend, Dario, the CEO of Anthropic, put out this article called We Must Pace the Frontier. Now, it's a pretty long, in-depth article, but I'm gonna give you the TL;DR, and then I'm gonna sort of give you the story beats of the week that followed from this article. Now, Dario starts by acknowledging what everybody is saying about him and the company. They've been trying to inform the public about all of these risks, even if that gets them accused of hype, doomerism, or regulatory capture. They claim they've tried to prioritize caution over speed and prudence over profit. He then goes on to map out 3 ways he thinks that we can, you know, pace the frontier. The first way being embedded evaluators. Anthropic is saying they're committing to this no matter what. He's saying that these embedded evaluators should be from third parties, not people that work for the company, but have ongoing access to permissions and tools similar to those of internal employees who do comparable risk assessments. So these would be people that don't work for Anthropic but literally have desks in their offices, access badges, company laptops, basically operate like they work at the company but are actually working for a third party. His second suggestion in his 3-part plan is regulation that targets all US frontier AI companies. And that covers even those who are unwilling to cooperate voluntarily. He goes on to say that essentially these regulations have to still keep us in front of, well, specifically China. If we slow down by more than this amount, then CCP-associated projects will pull ahead, creating significant national security risk. So regulations, but only to a degree that doesn't put us behind China. What do those regulations look like? Don't sell powerful AI chips or semiconductor manufacturing equipment to China and crack down on smuggling operations, crack down on unauthorized distillation by companies in authoritarian countries. Distillation is essentially when you use another sort of bigger teacher model, you use all of the inputs, the chain of thought, and then the outputs, and you use that information to train a new model that could respond similarly to the model that taught it. That's an oversimplification, but that's basically what it is. And strengthen security at the AI companies and prevent model weight And then the third piece of his plan is global pacing. This is the point where he says we actually will require cooperation with China, but he does go on to say that this is going to be difficult and we shouldn't be naive here. And he goes on to give some examples of things that we should agree globally should not be done with AI. Things like creating biological weapons, agreements to test each other's models before release, some sort of speed limit on the rate of recursive self-improvement. Improvement, which is the rate that new models teach the next evolution of new models. So Dario put that out over the weekend and almost immediately, like every head of every big lab jumped on board to say they agreed with him. Elon retweeted him and said Dario is right. Sam Altman, who sort of famously rarely agrees with Dario, actually said, I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we'll do the same. Demis from DeepMind: Dario's essay points towards the right path forward. The details need working through, but the direction is correct for meeting this critical moment. Satya from Microsoft jumped on board and responded. He said if the AI we build is not helping humanity and under human control, it's not worth pursuing. He goes on to say, we welcome the research focus and deliberate pacing needed to get alignment right as the design goal. Microsoft also welcomes the idea of embedded evaluators. But I think my absolute favorite response to all of this was the response from Mark Zuckerberg, who pretty decidedly did not agree with Dario. His take was essentially, stop asking for regulation and just do the right thing. He says, every lab has the responsibility and incentive to move at the pace required to train its models safely. and the ability to take its own actions to ensure that happens. People won't want to use agents that are misaligned with them and that don't do what they ask. So labs have a strong natural incentive to make their models more aligned. Any lab that doesn't focus on alignment will fall behind. This is my favorite part here. Meta delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing for people and for us. So he's basically saying, stop going and saying we need to regulate, we all need to get on board, and we all need to have these governing bodies and all of this stuff. Just do the right thing and don't release the models until you feel that they are safe to be released. He then goes on to say, engaging independent evaluators and advisors is industry best practice already. MSL, Meta Superintelligence Labs, already does this today in several areas because it helps produce better work. Other labs can just do this too. So on second read-through, I guess he's not really disagreeing with Dario. It's more More like he's just saying, yeah, we all agree with you and like we're already just doing that. Why do you need to go write up about it and try to get everybody on board? Like, do the right thing. Then of course, the US president had to get involved and give his 2 cents. The only control or guardrails that AI needs is a strong and smart, high IQ president. And the USA has that in spades. The Trump administration has stopped AI people from doing bad or potentially bad things like Dario from Anthropic. Who is now pretending to be a perfect little angel. We will continue to do so. We already have tremendous criminal and regulatory power over these companies. There's a sick conspiracy going on against AI and data centers, and the only one that is happy about it is China. Whoever wins AI wins. We are leading China and all others, and we will continue to do so. Conspiracy theorists, treasonists, traitors, and leakers beware. So yeah, the president is basically coming out and saying this is all just a big hoax, a big conspiracy. This stuff isn't that dangerous. Keep pushing forward. And then during the All-In Summit, that happened over this past week when Jensen Huang was on stage. Trump actually called in to chat with Jensen, and Jensen took the call live from stage. I won't play the full 4 minutes. It will be linked up in the description. But here's the highlights. The AI, it's almost this conspiracy.

And the happiest group is China. The data centers are great, and they make people wealthy, and they make states wealthy. And it's the oil of the next 20, 25 years. It's bigger than the internet. The robots will not be taking over the AI will not be taking over the rest of the world. The whole thing is a hoax. Now, with that, we have to be a little bit careful. You know, we have to do things and we have to do them prudently. But that doesn't mean we're going to stop an industry because, you know, as we work on the next 10 years about how to destroy it. So I'm with you all the way. And I have an expression. It's whoever wins, AI wins. That's how big it is. It's bigger than the internet. And whoever wins, AI wins. And we can't let this kind of stuff happen. And that includes, very much includes data centers. There are communities that were dying that have data centers Right now, and now they're wealthy communities. Yeah. Really wealthy communities. We're gonna make sure that everybody wins in the AI race in America. Every industry, every company, every state, every people. Good. Well, I feel strongly about it and I have the position that can do something about it. We're not gonna let that stuff happen.

So I already shared my thoughts in quite a bit of depth in last week's news video about this whole thing. So I don't want to go too deep into it in today's video. I just wanted to give you the landscape of what happened. And the sort of timeline of events over the last week, because last week we had that Jacob Coxon tweet that went viral, and then he was on like every news platform, and then other people started jumping in and saying they agreed with him, and the whole like doom scenario was the talk of last week. And this week Dario responded, and then every other company sort of jumped on board and responded with, "We agree with Dario," except for. The president who's calling the whole thing a hoax. I'm not gonna get political. If I was to give you a quick TL;DR of my thoughts, it's essentially that I agree these companies should be spending more time and more money on alignment focuses over profit. That's kind of how I'll sum that up. I don't think I'm as concerned over the doom scenarios as a lot of people are, but I don't think anybody should dismiss it either. I think we should be putting significant resources into making sure that future AI models are aligned with the goals of humanity. With that being said, let's move on to the rest of the news and talk about like cool stuff that you could do with AI now. We did get a few other updates out of Anthropic. They announced that Claude Cowork and their chat are now just one thing, so you don't need to switch between 'em. I was hoping I could show this off to you, but when I open the Claude app, I still see a chat and Cowork option here, and apparently I'm running the latest version. But when I scroll down on the article, this is rolling out to Pro and Max plans first. In the Claude app on web, desktop, and mobile over the coming weeks to existing and new users on these plans. I am on a $200 a month max plan and I still don't have it yet. If I'm being honest, like a little personal pet peeve of mine, and I don't think I'm alone in this, I wish companies would announce it when it's ready as opposed to saying, here's something we're doing, but not yet. It's coming in the future. Couldn't this blog post just have been written the day they made the feature available? And instead of saying, Here's something that we're going to do in the future. The title of this blog post is Claude Cowork and Chat are now one Claude. No, they're not now one Claude. They're going to be in the future after it finishes its rollout to Pro and Max plans. Make the blog post about it being changed once it's like actually ready for people. Sorry, I'm done ranting. Or am I? Because this is the other thing Claude rolled out. They're changing the way projects work. They put out a blog post called Projects Redesigned from Folder to Conversation. A new experience for Claude projects now available in beta in Claude Code. Now available. But let's scroll down a little bit here. Starting today, updated projects are available in beta to select Claude Pro and Max subscribers who use Claude sessions in Claude Code and don't have any existing projects on the web or desktop. So if you have any existing projects on the web or desktop, well, sorry, you're out of luck. Over the coming weeks, we'll expand access to more Claude Code users on these plans. Updated projects across all of Claude and Team and Enterprise come after that. If you're a Pro or Max user and don't have access yet, you can join a waitlist. Now, this does seem like an improved version of Projects. I like where they're going here. I just really don't like these AI companies rolling things out like this, where they make a blog post announcement saying this is now available, and then you read the blog post and they're like, haha, but not for you. So here's how Projects will work. When you start a project, you select a goal as well as the repo or context. Claude starts suggesting work it can pick up right away. You can configure the project's cloud environment, connectors, plugins, instructions, and model. You can monitor and guide progress in the main project chat or dive into each individual thread to examine and steer the details. You brief Claude in the project the way you'd brief a chief of staff, and it routes work to new or preexisting threads. Under the hood, each thread is a Claude Code Cloud session working on its own branch and copy of the repo. The coordinator keeps the work organized, but if any threads work on the same code, the overlap is resolved as a merge conflict, just like any other PR. The context builds out over time, And I just think it's a better way to do projects. It makes a lot of sense to me. So here's an example video where you can show that it goes and spins off a bunch of other threads and it works through each thread and you can steer each thread individually if you want, or you have the main chat that you originally started and you can use that as the chief of staff and tell that to go and, you know, steer the individual threads. Again, I demo it for you, but you know, See previous rants. All right, let's move on to a new feature that was rolled out into Gemini Notebook, previously known as Notebook LM. Now inside of Gemini Notebooks, you can talk through concepts inside your notebook. Can you explain the theory that objects in motion stay in motion? That theory is actually part of Newton's first law of motion, which is also called the law of inertia. So instead of just having like a podcast that you could interject if you want, you could just have like audio chats with your notes. You You can now capture anything, including thoughts on the go. Starting next week, you can use this new audio recorder in the mobile app to capture lectures and record your thoughts on the go, which I think is a pretty cool idea. I mean, this is a great tool for students, in my opinion. I wish I had this when I was in school where I could put my phone on the table and record a lecture and then go and sort of chat with that lecture later. That would be awesome. Again, not a feature that's fully rolled out yet. Seems to be a theme with AI companies. But a cool feature when it is ready. They're also adding something called interactive learning overviews to weave together summaries and Studio outputs like infographics, quizzes, and flashcards. So some pretty cool new features. It does say you can soon have a real-time conversation with your notebook. And we got an update out of Apple this week in that a profoundly more capable and personal version of Siri is now available. This new version of Siri uses Apple Intelligence. Supposedly it can see what's on your screen. Screen, has context between apps, and can do things on your behalf, unlike the previous version of Siri. Now this is another one where it says available starting today, and when I scroll down to the bottom of this article, it says Apple Intelligence in iOS 27— Siri AI is rolling out in beta in English in iOS 27. I am running iOS 27.2 beta, but when I go to look at Siri AI in my phone, it says I'm on a waitlist. So I don't even have Siri AI yet, despite this article saying available starting today. But some of the features we can expect once they actually give us the thing they announced that we have: you'll be able to, you know, draft emails with it, ask about what a family member suggested, get summaries of articles you're looking at on the web. It'll have visual intelligence, so it can actually see your screen and use that as context. We've already gone over what it was. going to feature back when they had a keynote about it, but now they're making an announcement that it's here, even though most people I guess still don't have it and it's still rolling out. So I guess I'll play with it once it's fully rolled out or I'm off the waitlist or whatever. I can't be alone in this, right? Like being super annoyed by like every single company saying, hey, this is now available and hey, starting today you can do this. And this is how this now works. And then when you go to use it, it's like, no, I— you announced that it's now like this, but when I go to use it, it says not yet. You're on a waitlist or rolling out to people on your tier, but not you yet. Like, why can't companies make the announcement that says this is now available once it's actually available for everybody? Like, Apple already made the announcement about what this new Siri AI will do. Like, they did a whole keynote about it and said, Coming soon. Why do they need to make another announcement saying starting soon again? I don't know. I'm, I'm getting worked up over this. It's just something I find very annoying because I go into these videos to be like, cool, it's starting now. It's available now. They're saying it's now here. And then I go to demo it and it's like, well, nope, not yet. It's just, it's, it's frustrating. Anyway, there's a couple new LLMs that I think are worth mentioning this week. I won't go too deep into them, but Alibaba released a new QWinn model in QWinn 3.8 Omni Flash. Here's some of the benchmarks. They're not really comparing it to the state-of-the-art models. We can see they're comparing it to Gemini 3.8 Flash and Muse Spark 1.2. So not like the best of the best models. Now, this isn't a model I've had a chance to play with yet. I haven't even been able to run it through Bucybench because as of this recording, it wasn't inside of OpenRouter, which is what I use to run Bucybench. But it's a model that is designed for workflows such as video editing, music video creation, film production, and commentary. Audiovisual summarization and real-time conversations. So it seems like it could work with any type of media. It's also got a 1 million token context window and something I'll probably play around with a little bit more once it's popped into OpenRouter and I'm able to test it some more. What was even more interesting for me this week was a new stealth model, and they're calling it Union Alpha. And right now you can use it inside of OpenRouter and OpenCode for free for a week. Now Union Alpha themselves does have an X account, and as recording this video. 8 minutes ago they posted, you found us, Union Alpha. The stealth model on OpenRouter, OpenCode, and Cloudflare is Pareto 26.9 from the Unbiased Company. Apparently Pareto is not a single set of weights, it's not just one model, it is a system. So it's several models, Open and Frontier, working on the same task with a harness that checks the work and calls in a stronger model when the task needs one. So it is some sort of model router that is using a bunch of different models, but it's not necessarily its own model. There was no new training run on this one. Here's their benchmarks on it. Scores a 74 on DeepSWE, which pretty much puts it at the top there. And we don't have a lot of info, but this was one I was able to test as it was available in OpenRouter for me to use. And when we look at BUCyBench here, we can see this is what it generated. It did it in 4 minutes and 44 seconds. And only used 9,588 tokens to do that. Now, why that's significant is if I come back here and I compare Union Alpha to GPT-6 Astra, when I look at the side-by-side of these here, we can see that its 9,588 tokens is less than a third of the tokens that Astra used to make a very similar-looking image. It also only took 4 minutes and 44 seconds, where Astra took over 10 minutes. Looking at these side-by-side though, they're so similar, I wouldn't be surprised if Union Alpha routed it To Astra because they look really, really close in my opinion. I just don't know how Alpha was so much more token efficient and faster if it indeed did route it off to Astra. I don't totally know how that works. I just found it interesting. I have a few more smaller things I want to share this week, so let's go ahead and break them down in a rapid fire. Pew pew pew pew pew pew. Let's start with some news out of Meta. They introduced a new subscription service that gives you more features and AI. Again, it's something that's rolling out gradually, but basically Meta One gives more self-expression features and AI usage. So basically for a monthly fee, you get access to, you know, Instagram Plus, Facebook Plus, WhatsApp Plus, and more AI usage across image, video, and you know, their various AI platforms. They do say Meta AI is still free for everyday use, but Core and Premium are for people who want even more creative freedom from the AI tools. They have plans ranging from as little as $2.99 a month for WhatsApp Plus all the way up to $4.99 a month for their Max plan. And since we're talking about Meta, last week we talked about Meta's Muse, their new agent, which feels kind of similar to like an OpenCLAW type thing. Well, this week they just gave it the ability to make outbound calls to US Again, it's rolling out and they're prioritizing it to people who asked their Muse to let them know when they wanted it first. I haven't had an opportunity to try this and I don't know what the voice sounds like. I'm kind of curious to hear like what does a call sound like when it calls a company on your behalf. But it will be interesting to just like ask your Muse to book appointments for you and it'll make a call and book appointments or, you know, wait on hold for you or things like that. That would be pretty handy. Google did roll out a new model this week as well called Gemini 3.8 Live. And Gemini 3.8 Live Extended Thinking. Now these are live dialogue models, so according to their benchmarks, they're now basically the state-of-the-art dialogue models, beating out GPT Live 1 Astra in pretty much every benchmark that they share here. Hi, can you hear me?

Yes, I can hear you loud and clear, David. I'm your onboarding agent. I can visually see the dashboard like you can, so feel free to ask me anything.

So that's what it sounds like. And the extended thinking version delivers increased intelligence for more complex workflows. So this is really kind of a feature for developers who are trying to build, you know, customer service agents and things like that. Since we're talking about Google, they also rolled out something called CC, which is an agent built for families. Some examples of what they say it can do. Each morning, CC delivers a shared your day ahead brief to everybody in your family. So instead of family logistical information staying trapped in one person's inbox, Everyone gets a clear view of where they need to be and what needs to get done. You can offload event and task management. It could help fill out permission slips or activity registrations. Basically an agent with like shared information in context for like a family to use, and everybody can see what needs to be seen. And it looks like you can kind of give access control to different people in the family. I was going to go and test that, so went to try to log in and got to a page that says the new CC is coming to you soon. We're gradually opening up access to this upgraded CC experience. If you're a coder that uses Grok, more specifically their Grok Build, it now has memory built in as of this week. Basically, as you work inside of Grok Build, notes are written in the background as you work and then are read back when you return to the project. We've gotten a lot of new music models lately. Well, at the end of last week, ElevenLabs introduced Music version 2.5. Here's a quick sample. Wait, wait, listen to this bit.

That's the bit. Started as a hum, now it's got a heartbeat. Yeah, play it in the car, play it at the wedding. You can have it back, babe, I'm not lending. That's it, babe, it's mine, it's mine. Made it out of nothing.

So seems like another decent music model to try. And while we're on the topic of ElevenLabs, if you're a developer, they just introduced their voice, music, image, and video generation models all inside of the MCP. So you can have your various agents— Claude, ChatGPT, Cursor, GrokBot, Hermes, OpenClaw— whatever you use, you give it access to this MCP, and now it can go and generate any of these things on your behalf. So I guess it's not really just for coders, but it's for coders and people who just want their agent to go generate this stuff for them on their behalf. OpenAI is rolling out a new version of their advertising in AI. In this video, you can kind of see how advertising has worked so far, where it puts a little ad below everything. But now if you on it, it actually opens up a separate chat with the company that the ad was for. So now people can go and click on an ad and then interact in a new chat mode with the company that sent the ad. So that's kind of an interesting development of where the advertising is going. And finally, let's end on some robots. This company, UbiTech, is developing a factory that can generate 10,000 robot units per year, and it can build one humanoid robot every 10 minutes. One interesting note about this is that humanoids are helping build the other humanoids. So humanoid robots perform tasks such as depalletizing, palletizing, and material loading, while collaborative robotic arms work with 360-degree rotating stations to tighten screws during assembly. So now we're getting a factory that's cranking out robots, and the robots are helping operate the factory. But anyway, that's what I got for you today. Again, the big story of the week was sort of the response to the doom of last week, but we did get a few updates about things that we'll eventually get that were made as announcements as if we get them now, but most of them we actually didn't. I won't go down that rabbit hole again. Anyway, I record these on Thursdays and publish them on Fridays, so if there's anything I missed that came out like later Thursday evening or on Friday, It'll likely make it into next week's Friday news video. Every week I drink from the fire hose all week. I keep up with all of the news. I overwhelm myself and make sure that I am completely tapped in so that on Fridays I can make a video like this for you and loop you in with everything I think you need to know that happened in the AI world. If that's something that interests you, make sure you like this video and subscribe to this channel. It'll make sure more videos like this show up in your feed. I've also got a new series that I just started where I'm bringing in other really cool people building awesome stuff with AI. My most recent one was with Bilal Sidhu, where he showed off his God's Eye View project. If you haven't watched that video, I'll link it in the end card that'll show up in a minute here. But it's basically people showing their application of a lot of these AI tools and showing really cool, unique ways and things they've built with it. If that's something you're interested in, Another good reason to click that little subscribe button. But again, that's what I got for you. Hopefully you feel more tapped in this week and now you know what's going on. Next week I'm gonna be at Meta Connect. The following week I'm gonna be at OpenAI's Dev Day. So my release schedule of videos might slow down a little bit while I'm doing some traveling, but I'll be just as tapped in as ever 'cause I'll be at the events where all the announcements are coming out of. So I will share all of that with you as I learn it and break it all down in the next handful of Friday videos. So thank you so much for hanging out with me, nerding out with me today. I really, really appreciate you. Hopefully I'll see you in the next one. Bye-bye.