Resources:
Join the discussion on LinkedIn: Got something to say? Let us know on LinkedIn and network with other AI leaders
Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup
Connect with Jordan Wilson: LinkedIn Profile
Start Here Series in our Inner Circle Community: Join for free access
Enterprise AI Strategy: Critical Updates from OpenAI, Anthropic, Apple, and Microsoft
AI advancements continue at a rapid pace, with leading organizations making strategic moves that demand direct attention from business leaders. Several developments this week underscore shifts in AI model development, software integration, and enterprise readiness. These updates signal new opportunities and potential risks for businesses seeking to stay competitive and secure.
AI Model Development: OpenAI’s “Spud” and Anthropic’s Emerging Offerings
OpenAI has completed pretraining on a new model codenamed "Spud," which is expected to be revealed within weeks. Internal messaging emphasizes its potential to accelerate both productivity and the broader economy. OpenAI is realigning focus toward enterprise tools, shelving high-profile consumer projects like Sora, and doubling its workforce to prepare for Spud's launch.
Anthropic, meanwhile, experienced a leak exposing development efforts on new models, including "Claude Mythos" and a larger-tier offering codenamed "Capabara." Internal documents describe these as a "step change" in capability, especially in cybersecurity. Notably, access to these models may be tightly controlled, excluding even high-paying customers from initial rollouts. Documents cite enhanced cyber capabilities and a need for organizations to harden systems before broad deployment.
Business Impact: Continuous model improvement and the narrowing gap between enterprise-grade and consumer AI introduce higher productivity tools, but also require proactive evaluation of access, deployment, and information security.
Advanced AI Agents: Anthropic’s Claude Automates Mac Workflows
Anthropic rolled out a research preview of Claude’s ability to directly operate a Mac computer for users. This feature allows Claude to open files, manipulate data, control browsers and developer tools, and even execute multi-app actions such as batch-resizing photos or transferring assets to other devices. The feature is limited to paid users on macOS and includes user consent steps and safeguards against common attack vectors.
Business Impact: Highly specific automation capabilities can streamline knowledge work and IT operations, reducing friction between AI systems and day-to-day applications. However, enterprises will need to balance efficiency gains with considerations around data sensitivity and operational reliability, as the feature remains in early, sometimes buggy, stages.
Productivity Ecosystem: OpenAI Codex Plugins and Microsoft Copilot Developments
OpenAI has quietly reintroduced plugin support for its Codex coding assistant, enabling skill and integration expansion via pre-packaged scripts. This development reduces the risk of hallucination, lowers inference costs, and allows for connection to external services (e.g., editing Google Drive files or reviewing GitHub changes). Codex, paired with GPT-4 high tiers, has demonstrated superior performance for in-depth engineering and general knowledge tasks.
Microsoft, on the other hand, has accelerated Copilot CoWork access through its Frontier Program, allowing for more robust, long-running collaborative workflows across the Microsoft 365 ecosystem. Notably, Microsoft is implementing a “critique” feature that uses an initial GPT model output, reviewed by Anthropic’s Claude for accuracy and citation completeness, along with a model council to compare outputs from different LLMs side-by-side.
Business Impact: Plugin architectures and cross-model review create tangible opportunities for custom process automation, error reduction, and increased transparency in enterprise workflows. These capabilities unlock automation possibilities not just for software engineering, but for regulatory compliance, document management, and collaborative decision-making.
Platform Integration: Apple’s Opening to Third-Party Chatbots
Apple is preparing to allow third-party AI chatbots – including Google’s Gemini and Anthropic’s Claude – to integrate with Siri on iOS 27, iPadOS 27, and macOS 27. This move ends the prior exclusivity arrangement with OpenAI and empowers users to route Siri queries to their chatbot of choice. Apple is also planning a major Siri overhaul, including a proprietary chatbot based on Gemini models.
Business Impact: Businesses invested in the Apple ecosystem will soon have easier access to advanced conversational AI tools directly from mobile and desktop platforms, offering greater flexibility for workflow automation, client communication, and information access.
Legal and Security Considerations: Anthropic’s Supply Chain Dispute
A federal judge in California has temporarily stopped the Pentagon from enforcing directives that would have blocked government use of Anthropic’s AI systems, after the designation of Anthropic as a supply chain risk – a first for a US firm. The legal action centers on whether concerns were based on security or political grounds, with Anthropic arguing that the “risk” designation damaged business and free speech rights. The core of the dispute is policy on AI used in autonomous weapons and related oversight.
Business Impact: As legal scrutiny intensifies, enterprises must closely monitor the regulatory environment surrounding AI usage in sensitive and defense-related sectors. Decisions made in this space can directly affect vendor relationships, procurement policies, and compliance protocols.
Technical Efficiency: Google’s TurboQuant Reduces AI Memory Demands
Google researchers introduced TurboQuant, a technique for cutting the memory required by transformer model caches down to one-sixth their original size, without loss of accuracy or additional retraining. This approach is compatible with both new and existing models and enables AI applications requiring significantly less hardware investment.
Business Impact: TurboQuant and similar optimizations make high-performance AI more affordable and accessible, paving the way for local deployment of complex models on a broader range of hardware. Organizations can benefit from reduced cloud dependency, shorter inference times, and decreased total cost of ownership.
Fast-Moving Landscape: Weekly Releases and Announcements
Several additional updates merit close tracking:
OpenAI’s ChatGPT ad program surpassed $100M in annualized revenue, demonstrating rapid monetization of the platform.
Meta rolled out new AI shopping features for Facebook and Instagram users.
Luma AI and Google both launched new or improved multimodal and translation models.
Microsoft expanded Copilot’s skills to include both one-time tasks and repeatable workflows, showcasing the trend toward persistent, context-aware agent assistants.
Regulatory and competitive clarity continue to evolve, forcing faster evaluation and adoption cycles for enterprise leaders.
Conclusion
The latest advances push AI further into the core of enterprise operations, not just as background tools but as primary actors in workflow automation, knowledge management, and security-sensitive applications. Leaders must account for rapid product iteration, evolving regulatory scenarios, and nuanced platform interoperability in their 2024 AI strategies. Careful attention to feature-specific details and business layer integration will separate opportunistic adopters from those who struggle to extract value from the ongoing AI surge.
Topics Covered in This Episode:
- OpenAI SPUD Model Leak Details
- Anthropic Claude Computer Use Goes Viral
- Codex Plugin Support and Integration
- Apple Siri AI Chatbot Integration iOS 27
- Anthropic vs. U.S. Government Legal Battle
- Google TurboQuant Model Compression Method
- Anthropic Claude Mythos and Capabara Leak
- Microsoft Copilot CoWork Frontier Launch
- Multimodal Model Counsel Feature in Copilot
- Major AI Industry Weekly Feature Roundup
Episode Transcript
Jordan Wilson [00:00:16]:
I think right now, the best AI model in the world, if you have the budget and you have the patience, is GPT five four Pro from OpenAI, which seems like it was released months ago. Crazy enough, it was announced this month. Yet, we already have new AI news that OpenAI is working on a newer model, codenamed Spud, and it could be right around the corner. And OpenAI must think it's pretty good as they say it'll reshape the economy, and they're doubling headcount to prepare for it. And, well, they're not alone. As their biggest AI competitor, Anthropic, well, recent leaks show that they're also cooking up a new model that could arrive soon. And, apparently, it's so powerful the company is concerned about its capabilities. Yeah.
Jordan Wilson [00:01:07]:
So this week was a heavyweight week for both the big AI companies, a ton of AI news, leaks, and updates, and we've got it all. Oh, and if you stick around toward the, well, the end of the livestream version of our show, you'll actually be the first in the world to hear of a new AI update from one of the big four that will break to you live that, well, we think it's actually really good. Alright. I'm excited to get to you the most important AI news of the week. I hope you are too. Let's dive into it. If you're new here, welcome. What's going on? My name is Jordan Wilson.
Jordan Wilson [00:01:43]:
This is Everyday AI. It's a daily livestream podcast and free daily newsletter helping everyday business leaders like you and me keep up with the non stop avalanche of AI updates because they are literally every single day. I tell you what's important, what's not. You take that information to be the smartest person in AI to grow your company and your career. So it starts here with the unedited, unscripted live stream podcast, but take it to the next level. Make sure you go to our website at youreverydayai.com. Each day, we recap the, livestream in the newsletter so you can quickly catch up, and we give you all of the other important AI feature updates and everything else that you need to know to stay ahead. So without further ado, let's get straight into it and give you the AI news that matters for the week of March 30.
Jordan Wilson [00:02:32]:
Well, the first one, it's a short name for a, apparently, really big model. So OpenAI, their new model nicknamed SPUD, well, could be dropping any week or any month now. So according to reports, OpenAI has completed development of a new AI model, codenamed Spud, capability the company says could significantly boost productivity and help refocus its own internal strategy toward focusing more on enterprise tools and integrated products. This update comes as open a OpenAI has shelved a bunch of, well, kind of popular products such as Sora, and they've delayed some of their other efforts as well to concentrate on their core business. And, well, at least it looks like for the future that maybe whatever this new codenamed spud model is actually called. Right? Whether that's a GPT five five, a GPT six, we're not sure. But according to the information, OpenAI recently finished pretraining work on spud and plans to reveal it within weeks with a CEO, the CEO, OpenAI, CEO Sam Altman reportedly saying the model can really accelerate the economy, a claim that signals a push for broad commercial impact. Spud is expected to support OpenAI's plan to build a productivity super app by combining chat GPT, codecs, and the company's browser called Atlas, potentially making multi tool workflows faster and more tightly integrated.
Jordan Wilson [00:04:10]:
So while specific technical details about Spud's architecture or capabilities were not disclosed in the reports, the emphasis from leadership suggests improvements in productivity, multimodal understanding, or more advanced reasoning and tool use. The timing of Spud's completion aligns with OpenAI's decision to redeploy resources away from consumer experiments like Sora and their adult mode, which they've delayed, and toward enterprise and robotics, implying the spud will be a cornerstone of that strategic shift. So, pretty big news here from OpenAI. Yeah. I was, like, looking at the calendar and I'm like, wait. What has it been, like, two months since GPT five four? And I'm like, no. We're still in March. And GPT five four was released in March.
Jordan Wilson [00:05:02]:
So the pace of development, obviously, if you've been following AI for more than a couple of months is straight up blistering year. Right? And I think maybe one of the reasons we've seen this is continued, pressure and great models and improved, productivity gains specifically, I think, from Anthropic where I think they've been winning 2026 so far, at least when it comes to shipping products to consumer and then obviously from Google as well. Alright. Speaking of things from Anthropic, that's our next piece of AI news, and this one was big. Well, it was at least extremely viral. We'll see how big it ends up being, but I think this is something, it could kind of, shape the next agentical layer. So Anthropic has rolled out a research preview feature that let its that lets its chatbot, Claude, actively use a Mac to perform task for users, making a step toward more autonomous agent style AI that can operate apps, type, click, and use connectors for services like Google Calendar and Slack. So the new feature is in a limited research preview, and it's right now only available to paid Claude users.
Jordan Wilson [00:06:20]:
So if you're on the pro or the max plan and it is restricted to computers running macOS, making it immediately relevant only to yeah. If you're a paid user on Apple. So here's what it is and, well, why it's pretty cool. It can perform real world actions on your computer. Right? So a a lot of these capabilities were, previously available kind of in Claude CoWork, in Claude code, but here's what's new. Well, Claude can now perform real world actions on your computer. It can open files, use your browser and developer tools in the browser. It can type and move the cursor and even choose connected apps that it has access to inside Claude or manually control to complete tasks such as transferring files to, your phone or batch resizing photos.
Jordan Wilson [00:07:12]:
Anthropic says, Claude will always ask for permission before acting and allows users to stop the agent at any time aiming to balance, convenience with user control. Also, the company implemented automated safeguards to scan for prompt injections and other common attacks against agents, and it disables some apps by default. But it warns the feature is new, it may contain errors, and it should be used with highly it should not be used with highly sensitive data. So I don't know. Live stream audience, podcast people, let me know in a comment. Have you used this yet? I've used it. I think it's extremely impressive. We may go, give this the the total hands on treatment on Wednesday.
Jordan Wilson [00:08:00]:
So, yeah, if you are new to the show, on Mondays, we go over the AI news. On Wednesdays, we go pretty deep, hands on demo with one tool. And then on Fridays, we kinda do a recap style of new features. And Tuesday, Thursday, you know, we rotate different shows in there. But we may go super in-depth with this one on Wednesday. My quick, take on this is it's buggy. Obviously, it's the first of its kind, but it's extremely impressive. And I do think it represents the, not the final layer.
Jordan Wilson [00:08:29]:
Right? But it it represents the next layer of agentic work. You know, I think until, some of the protocols improve, you know, the the eight the eight to eight agent to agent protocol, MCPs. Right? All of these other things. I think there's still even with all of these, you know, protocols that essentially help AI agents talk to other AI agents and to help large language models talk to other large language models. Even until these things improve, I still think the next layer is, well, it's agents using your actual computer. And this is a very, you know, open cost style update here. Right? But this is different. This isn't just being able to save files, upload files, access your local directory.
Jordan Wilson [00:09:11]:
Yes. That has already been available in Cloud Code work and in Cloud Code. This is actually controlling your entire computer. So I've liked it so far. Like I says, when like I said, when it works, great. Doesn't always work. But being able to open different documents, on your computer being, having Claude, the desktop app control open, use other apps on your, computer. That's kind of the new capability here.
Jordan Wilson [00:09:38]:
Right? You know, some of my testing, I'm I'm having, you know, Claude open perplexities, Comet browser, open OpenAI's Atlas browser. Right? Use different browsers, use different desktop, programs that normally might not be able to talk to or access other AI. So pretty big one. If you think we should give this the the Wednesday, deep, deep dive treatment, just say computer use, in a comment there. So I I I I know if we want this. So it was crazy viral, though. So when open a or sorry. When Anthropic released this, on Twitter, I think the release video got, like, 75,000,000 views almost instantly.
Jordan Wilson [00:10:17]:
So pretty big. Alright. Next piece of AI news. This might seem like a smaller one, but I actually decided to highlight this as one of our main AI news stories for the week. Although, there's probably even bigger AI news stories from OpenAI, but I think this one is actually really important. That's because, well, plugins are back. Right? If you've been a long time listener of this show, you know I was, very, very bullish on plugins early on inside chat GPT. Super bummed that OpenAI actually disabled them, you know, discontinued support.
Jordan Wilson [00:10:50]:
Eventually, you know, it led to GPTs and apps and other things, but now OpenAI has quietly updated its codex program to let developers create custom plugins and let general users use them. So OpenAI has introduced plugin support for codex, allowing developers to add custom skills and external integrations to their coding assistant. So the most newsworthy detail here is that plugins can include prepackaged scripts, which are skills, plus configuration files so codex can run tested code snippets instead of just generating code from scratch each time, which that obviously reduces the hallucination risk and cuts inference cost and response time as well. So plugins can connect codecs to external services through MCP servers and developers can upload MCP configuration files to control sandbox middleware and environment behavior. So OpenAI also with this launched a plugin directory with more than a dozen prebuilt integrations, including plug ins that let codecs edit Google Drive files in review GitHub repo changes. So this follows a similar capability launched by Anthropic for Cloud Code a couple months ago. They're kind of plug in ecosystem, and Anthropics already supports sub agents, and Codecs just rolled those out as well. So OpenAI says plug ins are designed to help software teams keep development tools and configurations synchronized across multiple engineers, OpenAI accounts, reducing code inconsistencies on collaborative projects.
Jordan Wilson [00:12:29]:
Here's why I think this is super important. Number one, everyone, including OpenAI, is straight up sleeping on codecs. Right? I do think maybe this is one of the reasons that OpenAI is, shifting away from having codex as its own dedicated app and moving forward with the super app where it kind of rolls codex chat g b t and the Atlas browser all into one. That's because codex is a freaking beast. Yes. It is better. I don't care what anyone says. I use these tools more than 99.9% of the population.
Jordan Wilson [00:13:01]:
Right. Codex right now with GPT five four high, extra high is better than Claude Code with Opus four six, better than Claude CoWork with Opus four six. It is just better. You know, if you want something that looks nice, and done quickly, you can use Claude Code, Claude Co work. If you want something done the right way, and you have the time, I think codex is by far better. Yeah. It creates ugly front ends. We get it.
Jordan Wilson [00:13:29]:
But it just gets the code right. But here's the thing. Even if you are not using it to code, it is amazing at just doing everyday, knowledge work tasks. Right? So in the same way that I think that, anthropic kind of, not maybe pivoted, but kind of remarketed Claude Code essentially as Claude Cowork. Right? Because they realize that it's great for even non coders. I think that's what we're gonna see with OpenAI in this new super app because maybe they're slowly starting to realize that codex is actually really good for people that are not running software, not software development teams. Yes. It's good for those people, obviously, but it's good for everyday knowledge work.
Jordan Wilson [00:14:10]:
And I think plugins combined with the, kind of the news of the upcoming super app is really significant of that, and it's showing that I think the average knowledge worker is gonna be using the codecs platform a lot. Whatever that looks like in the new super app, I mean, that remains to be seen. But what this means to you, if you have not used codecs yet, start using it now. Right? These plugins, I think, do make it easier, and you can probably find some great use cases. Right? Even one of the plugins I just mentioned there, Google Drive. Right? Being able to edit Google Drive files, that's pretty big. Right? That's pretty big. It's a, you you know, a big shortcoming a lot of the, you know, front end AI connectors not being able to actually edit files.
Jordan Wilson [00:14:55]:
Alright. Our next piece of AI news. Well, Apple. Right? We're getting all the big companies here. So Apple may actually get AI that works. We'll see. They've been saying that for years. Like the the boy that cried wolf.
Jordan Wilson [00:15:07]:
We'll see if we believe them this year, but, maybe it will happen. Because according to Bloomberg, Apple will soon let third party AI chatbots such as Google's John Eye and Anthropic's Claude integrate with Siri starting in iOS 27. That means iPhone users can route unanswered Siri's queries to their preferred chatbot app. So, also, Apple announced that at their upcoming WWDC keynote, they're gonna have a lot more of an AI focus and third party, integrations that will work with iOS. So, essentially, it seems like Apple is saying like, yeah. We've spent billions of dollars in, three or four years, and we can't get it right. So, essentially, we're gonna ope open up the generally extremely restricted Apple ecosystem and allow users to integrate with other AI chatbots. Right? So if you've been like me and you're using your iPhone and you're like, this thing is, you know, a a very expensive dumb phone and there's no AI that works in it, it's kind of my thought.
Jordan Wilson [00:16:12]:
Right? Siri doesn't work. It doesn't do anything. You know, yeah, they have these, oh, these right tools. They're they're useless. Right? So maybe now we can all finally get AI that actually works on our devices. So according to reports, users will choose which services Siri can access through the new extensions settings in iOS 27, iPad OS 27, and Mac OS 27 found in the Apple Intelligence and Siri panel of settings. So that the change ends the practical exclusivity of Apple's prior OpenAI tie up. Right? So they've had this feature, not very well integrated, I would say, right, where essentially Siri can just kick things over to chat gbt, but OpenAI's chat gbt will still remain supported, but will no longer be the only external bot that Siri can call.
Jordan Wilson [00:17:05]:
Apple also is still planning a major Siri overhaul, which we've been hearing about for years, and will ship its own Siri chatbot built on the Google Gemini models, which we've talked about pretty extensively on that show. On this show, while extensions give users the option to direct requests to other chatbots instead of just Siri. So Bloomberg reports that Apple will announce these new features and the new, updated Siri that can talk to other chatbots at their WWDC keynotes, this summer in June with the feature arriving in iOS 27. Alright. We'll see if that actually happens. Right? Apple. Right? It's it's it's it's it's kinda funny because two years ago, right, Apple was like, oh, Apple intelligence. You know, we are AI, and it's gonna be the best AI ever.
Jordan Wilson [00:18:01]:
And then they got sued because they didn't release anything that actually worked. There was nothing actually intelligent that Apple released. And then they kind of quote unquote took a year off, you you know, from their big WWDC. That's the worldwide developer conference. Right? That's their one big time of year where they come out with all their announcements. So, you know, two years ago, they're like, oh, yeah. We are Apple Intelligence. We are the smartest AI in the world.
Jordan Wilson [00:18:24]:
They got sued because they couldn't deliver. So then they quote unquote took a year off from AI. They didn't really, you know, announce anything of substance at last year's WWDC. And now apparently, this year, they're back to being the Apple Intelligence. So, hopefully, they've learned their lesson and they won't lean into trying to redefine the actual AI category. Probably not a good idea, especially if all they're ultimately doing here is allowing users to use better and smarter AI. So it should actually be pretty telling, how they approach this from a branding angle. But I do think, however, the markets, right, if you care about that, I think the markets will actually like this move from Apple because they're like, yeah, Apple.
Jordan Wilson [00:19:06]:
We understand you can't build AI. So you should probably start integrating with as many third party, AI chatbots as possible. Alright. Our next piece of AI news, well, this one is the ongoing. AI moves too fast to follow, but you're expected to keep up. Otherwise, your career or company might lag behind while AI native competitors leap ahead. But you don't have ten hours a day to understand it all. That's what I do for you.
Jordan Wilson [00:19:39]:
But after 700 plus episodes of Everyday AI, the most common questions I get is, where do I start? That's why we created the Start Here series, an ongoing podcast series of more than a dozen episodes you can listen to in order. It covers the AI basics for beginners and sharpens the skills of AI champions pushing their companies forward. In the ongoing series, we explain complex trends in simple language that you can turn into action. There's three ways to jump in. Number one, go scroll back to the first one in episode six ninety one. Number two, tap the link in your show notes at any time for the start here series, or you can just go to starthereseries.com, which also gives you free access to our inner circle community where you can connect with other business leaders doing the same. The Start Here series will slow down the pace of AI so you can get ahead. Rama.
Jordan Wilson [00:20:36]:
And it's not over, but we have a new chapter in the, novella. So a federal judge has temporarily stopped the Pentagon and other federal agencies from enforcing directives that would have immediately halted the government's use of Anthropic's AI tools, keeping the company's widely used clawed system available while a broader legal fight plays out. So a US district judge in California issued an order, late last week preventing the enforcement, directives from president Trump and defense secretary Pete Hegseth that sought to bar anthropics tools from government use for now. Alright. So, essentially, the federal government labeled anthropic a supply chain risk, for a couple of reasons, which we'll get to here in a minute. But essentially, a judge said, nope. Doesn't make sense. Right? So the order now lets Anthropics AI continue to be used inside the government for now, and by outside contractors working with the military while the lawsuit proceeds.
Jordan Wilson [00:21:48]:
So the judge wrote that the government actions looked aimed at, quote, unquote, crippling anthropic and a chilling public debate. And the judge characterized statements by officials as appearing to be classic first amendment retaliation. So the dispute began after public criticism from presidents Trump and HECSIF who labeled anthropic a supply chain risk. And the first public use of that designation against a US company ever in a label typically reserved for, companies tied to adversary nations. So yeah. It was very strange that the federal government decided to label, Anthropic a supply chain risk because that's never literally happened against a US company. Then Anthropic, in response, sued the Department of Defense and other agencies earlier this month, saying the government's designation in public attacks harmed its business and violated its free speech rights. So the judge noted that officials public comments attacking Anthropic were more on political grounds.
Jordan Wilson [00:22:55]:
For example, they called the company Woke and its employee in in Anthropic's employees, quote, unquote, left wing nut jobs rather than pointing to specific security defects. Right? So this drama is not yet over. Right? Anthropic essentially, you know, they, had two little clauses that they wanted to have in their agreement with the government for reasons they said that would ultimately give them more protection over how the military would not use its AI. As an example, they said they didn't want the government to use its clawed systems for, fully autonomous weapons in war without human oversight. Right? So we'll see how this continues to go on, but it is, again, one of those stories, that is going to continue to drag on. But the latest, the latest one here, pretty big update. A judge essentially saying, no, government. You were wrong.
Jordan Wilson [00:23:54]:
You can't do this, and this was political. And there's really not a lot of merit. So, yeah, it will continue, to be legislated. Alright. Our next piece of AI news, a little technical one here, but it actually had pretty big, ramifications both instantly and in the long run. So Google researchers have introduced a new methodology called turboquants, a two stage vector compression method that reduces transformer key value cache memory by about six times while preserving downstream accuracy on tested workloads. So the most newsworthy part here is that turbo quant lets models quantize key value caches down to three bits without retraining, which could sharply improve, lower memory requirements for large language models. So the system requires no model retraining, which that's huge, making it potentially compatible with existing open models and easier to adopt in current production stacks.
Jordan Wilson [00:25:03]:
So here's, in a simple way. Right? This is technical. I had to read this, this one a couple of times because, you you know, as much as I talk about AI, I am not super technical on the pre training side. So it's kind of like a a a super shredder, right, for an AI's memory. So when you chat with an AI, it has to store a lot of data to remember what you just said. And usually, that takes up a ton of, well, expensive memory. And turboquants from Google, so essentially a new, quant quantized, technique. Right? It just kind of squishes the data.
Jordan Wilson [00:25:38]:
So it shrinks the memory to about one sixth of that size without losing any of the information. And by doing that, it obviously speeds things up because the data that it has to remember, right, about your conversations, is smaller. The AI can, quote unquote, read it up to eight times faster. And the impressive things here as well, this new technology, or technique can apparently be applied to any new model. So it doesn't have to be only new models that haven't been developed yet. So there's no extra work. So you don't have to retrain the AI. So this can be applied to open models.
Jordan Wilson [00:26:15]:
So as an example, it can work instantly on models maybe like Google's open source Gemma. So also interestingly enough, so this happened late last week and it instantly triggered a pretty sharp but probably temporary sell off, on the stock market of, you know, memory chip stocks. So, pretty big deal here. We'll see what Google does with this technology, how it may be used across the spectrum, but this could ultimately, right, bring way more powerful models to way smaller devices. Right? So as an example, as of recently, there's been this big, you know, OpenClaw and OpenClaw esque, kind of surge. Right? And a lot of people are buying very expensive, you know, $10,000 Mac studios or, you know, NVIDIA DGXs, to run the most powerful local models that they can so they don't have to pay for cloud inference. Right? They don't have to rack up API bills with Infropic, Google, or OpenAI. But the problem is, well, you have to have a huge, very expensive computer to run these models.
Jordan Wilson [00:27:20]:
Well, maybe you might be able to, get a six times more powerful model on the same size or just have a much smaller computer, or a, you know, not as powerful GPU to be able to bring all of these things. So, pretty pretty exciting news, for the future of AI development where we just might have way better local models available on phones and computers in the future, if turbo quant ends up being, what it could be, which is a game changer for how large language models work. Alright. Here's well, our pretty big biggest story of the week. Although, we are gonna be able to break, break some news here in about five minutes from one of the big companies. So, a major leak at Anthropic has exposed nearly 3,000 internal files. And, well, probably 99.9% of them weren't very important except there was a new unpublished draft about a new powerful model from Anthropic called Claude, Mythos. I think that's how it's pronounced.
Jordan Wilson [00:28:27]:
Right? In details of an even larger tier named Capoeira. So, Anthropic did confirm the reporting to Fortune that an accidental data leak exposed nearly 3,000 assets uploaded to its content management system on its website, but marked as private, making them publicly accessible in a data lake. So the leaked collection included unused marketing assets, PDFs, images, all that stuff. Also, employee and corporate event information, but the big thing was a draft blog post about a new AI model labeled Claude Mythos. So according to the leaked drop, draft, Mythos is by far the most powerful AI model that Anthropic has ever developed, and the company calls its performance a step change. The model is in trials right now with selected early access customers. So the documents revealed a planned new top tier as well, right now, codenamed, Capabara, I think that's how it's pronounced, which would sit above Anthropic's current highest tier, which is Opus, making Capabara the company's late, largest and most capable offering. So Anthropic's leaked materials warns that Claude Mythos and Capabara could significantly increase cybersecurity risk, saying the model the models are currently far ahead of any other AI model in cyber capabilities.
Jordan Wilson [00:29:58]:
The leaked draft states Anthropic intends to study and share findings about near term cyber risks so defenders can prepare, and it is giving early access to organizations, to give them time to harden their code bases against potential AI driven exploits. But I think the potentially bigger news here might ultimately be this new model's access because, reports point to a very limited early access rollout with select customers only. And those leaked materials described access expanding gradually through the cloud API rather than broad availability in a standard paid plan. So, yeah, this might be the first, well, major model we've seen from any company that might not be available to regularly paying business users. Right? So, if you have a, you know, Claude, you know, paid plan where you're paying monthly and you expect, oh, well, I'm gonna have always the, you know, most powerful models from INTROPIC. Well, maybe not or maybe not right away, which I think is actually a a pretty big pivot, overall with how these companies work. Right? We've seen reports that OpenAI and Anthropic are likely both going public this year, and this could be a new, kind of tactics to bring in more revenue maybe, right before in IPO. We'll see.
Jordan Wilson [00:31:29]:
But, again, reports are saying that this might even if you're a 200, you know, like myself, you know, paying $200 a month, for Claude Max, you might not get access, to the new Claude mythos, whenever it's released. Alright. Speaking of release, well, we're safe now. The embargo has lifted, and we can talk about new updates from Microsoft. So, yes, if you are listening here on the live stream, you are the first in the world to hear about this. So Microsoft is now just introducing that Copilot Cowork is available through its Frontier program. And there's some new, which I think are really, really good updates that they're rolling out to Researcher. Alright.
Jordan Wilson [00:32:14]:
But first, Copilot, CoWork. Alright. So we know that Microsoft Copilot, CoWork isn't new. They announced this a couple of weeks ago, but this is now pretty big news because they're making it available via the Frontier program. So that's when a lot of companies are now going to get access to it. Also, they're introducing a Microsoft three sixty five Copilot feature designed for long running multi step work, instead of just the one one app one off chat prompts. So that's with, you know, Cowork. You know, this is a kind of partnership, with Anthropic.
Jordan Wilson [00:32:47]:
It's not really a white labeled version of Claude Cowork, but it kind of is. Right? They're leveraging, that technology, and, obviously, Microsoft is a big investor in Anthropic. So, initially, this was rolling out to a very small beta group. So now it's going to be rolling out into a pretty big group, in the frontier program. So, the company said Copilot Cowork lets users describe the outcomes they want then creates a plan, works across files and tools, and then shows visible progress. Right? So you kind of, the screenshot here I have for our livestream audience, you can kind of see it work through its entire plan. Very similar to if you've used, Infropix Claude co work. Right? But Microsoft style.
Jordan Wilson [00:33:31]:
So Microsoft said Copilot co work includes skills from both Claude and Microsoft. Right? So it's not just just, in in exact, you know, duplicate of, Copilot co work because it does include, obviously, a lot of specific in Microsoft exclusive capabilities, including calendar management and daily briefing. And the company said it can handle both one time tasks and repeatable workflows such as monthly budget review. Alright. But here's actually some of the announcements that I am maybe more personally excited about. So Microsoft has announced a new and improved researcher agent. Here's the big part though. Built on multimodal intelligence.
Jordan Wilson [00:34:15]:
So it's aimed at a more complex knowledge work by synthesizing information across sources and producing cited reasoning analysis that users can act on. So one of the most important additions is the new critique feature, where OpenAI's GPT model will draft a response and then Anthropics Claude model reviews it for accuracy, completeness, and, citation integrity before delivery, showing that Microsoft is now productizing a multimodal review workflow rather than relying on a single model's first pass. Y'all, I can't believe it is, essentially now 2026, and we're now just getting this for the first time from one of the big players. Granted, it's really only Microsoft or Google that could do this. Right? Essentially, bringing a new model in from a different company. Right? So that's completely different training. It works in a completely new way. And essentially, having those two models work well with each other, but technically against each other.
Jordan Wilson [00:35:23]:
Right? So having, OpenAI's GPT create something, and then Infrafx Claude essentially tear it apart. Right? For anyone that's a power user of AI, if you're using it for, high value work, this is what you've been doing manually. Right? For me, this is probably what I spend 70% of my time, doing. Right? I think there are some great third party solutions, such as perplexity's model counsel, but it's honestly very expensive. Right? It eats up a bunch of credits. There's some other, you know, third party less popular tools that kind of do this, but it's actually a pretty big deal that Microsoft is the first of the big four. Right? So, that's Microsoft, OpenAI, Anthropic, and Google. They're the first one to provide this.
Jordan Wilson [00:36:07]:
And like I said, they can't really anyone else can't really do this. I think technically Google could. Right? Because Google is also a big investor in Anthropic, but I think they're probably competing a little bit too closely on the model size, on the model side as well. Because right now, Microsoft is even though they are developing, kind of their next generation of models, right, in their, Microsoft AI, you know, Mustafa Suleiman now working on that side. Right? They don't really have, today at least, you know, frontier level models. So this is pretty big both expanding, Copilot, co work, to the, frontier program, which is a lot of enterprise organizations here in The US. But then also with these new updates in, the model counsel, which is huge. Right? So that's a new feature.
Jordan Wilson [00:36:58]:
So you have the critique feature in researcher, but then you also have the model counsel feature, which lets users compare responses from different models side by side so they can see where answers agree, where they diverge, and what each model contributes. So like I said, that's something that's been kind of available inside perplexity, but pretty cool to see this offer now from one of the big four. Alright. That's it for our big stories of the week, but let's roll into our what's new and what's next. So this is the kind of bullet point roundup of all the other big stories in other weeks. Some of these might have been some of the biggest stories of the week, but there's a lot going on this week with all the new leaks, the big bottles, released from Microsoft, a ton going on. So let's go over, this what's new and what's next. So, Infropic released a new economic index, which shows a widening AI fluency gap.
Jordan Wilson [00:37:52]:
Google released their updated Lyria three pro. You can create, audio tracks up to three minutes. OpenAI reportedly closed their funding round that's reaching a $120,000,000,000. Alright. Meta introduced Sam 3.1. So that's their their segment anything model. So that allows you to segment any object in image or videos from simple prompts like click or boxes. SoftBank reportedly secured a $40,000,000,000 loan tied to their OpenAI investment.
Jordan Wilson [00:38:23]:
A report said that ChatGPT's new ad pilot passed a $100,000,000 in annualized revenue. So we saw some reports that even though maybe it wasn't as successful as they had hoped, well, they've already brought in a $100,000,000 in annualized revenue. My hot take is they're gonna, like, 10 x that by the end of the year. It's gonna be a cash machine. The White House announced members of the president's council of advisors on science and technology, including Mark Zuckerberg, Jensen Huang, and others. Google released Gemini 3.1 flash live and search live. We went over that on our Friday show, going over our new Friday features, our weekly show. OpenAI upgraded their Chat GPT shopping and product discovery, really emphasizing shopping less and product discovery more.
Jordan Wilson [00:39:09]:
Google released an updated version of Google Translate. It's actually really good. I've been using it. We went over that Friday as well. Arm introduced that it will start making its own chips. OpenAI revamped Chad GPT shopping with a Gentic commerce protocol and the Walmart app. Luma AI launched a pretty impressive and kind of out of nowhere, new multimodal image model called UniOne. Fairly impressive so far.
Jordan Wilson [00:39:37]:
Drugmaker Eli Lilly agreed on a $2,700,000,000 deal with Insilico for AI developed drugs. Sam Altman reportedly has now shifted his own priorities and will no longer directly oversee security or safety at OpenAI. Chat GPT rolled out its library feature for central file management. Anthropic is reportedly working on a computer use for mobile after their very viral computer use for Mac OS. In Gemini business, Google is testing skills. Meta conducted another round of layoffs, this time around 700 employees, amid their AI shift. OpenAI shelved reportedly their adult mode indefinitely alongside cutting and, killing off Sora as a dedicated app for now. Meta rolled out new AI shopping experiences on Facebook and Instagram.
Jordan Wilson [00:40:31]:
OpenAI officially relaunched their nonprofit arm and named leaders, and they planned to spend at least a billion dollars in doing so. Manus now offers a computer control from their, iPhone app. I haven't tried that one yet. I do have a Manus subscription, so I'll have to give that one a try. There's a new benchmark in town, Arc AGI three. Since Arc AGI one and two have essentially been saturated, but we're still saying we haven't achieved AGI. Well, now there's ARC AGI three, a new benchmark that has launched with $2,000,000 plus in prizes. Apple will reportedly like I said, we already covered that one.
Jordan Wilson [00:41:06]:
They are opening, Siri to rival AI assistance. Google Gemini is launching a chat history and memory import tools to make it easier to start with Gemini, to transfer over. ChatGPT is getting unified Google Drive integration for business and enterprise. Suno released version 5.5. A lot of cool features in there allowing you to also kind of use your own voice, which is cool. And then last but not least, agile robotics in Google DeepMind announced a strategic research partnership. Y'all, that was a ton happening this week in AI. If you missed anything, well, we just covered it.
Jordan Wilson [00:41:45]:
So you don't have to spend hours each and every week saying, oh, what's happening in AI? What's worth paying attention to? What's not? What should I be using in my company? That's what we do on our Monday show. So I hope this was helpful. If so, please subscribe to the show. Please leave us a rating. I'd really appreciate that. Takes, like, thirty seconds whether you're listening on Apple Podcasts or Spotify. And then if you haven't already, make sure you go to your everydayai.com. Sign up for the free daily newsletter where we recap not just each day's podcast, but everything else you need to know to be the smartest person in AI at your company.
Jordan Wilson [00:42:17]:
So thank you for tuning in. We hope to see you back tomorrow and every day for more everyday AI. Thanks, y'all.