Grok goes agentic, while Anthropic watermarks everything

Mike Belsito covers three stories this week that matter to product pepole this week. xAI's Grokbot launches in public beta as a persistent digital co-worker that non-technical users can actually configure; Anthropic rolls out invisible watermarks across all Claude outputs in response to the EU AI Act, with a December 2026 deadline for full coverage; and Meta releases Llama Glimmer, a 30-billion-parameter open-weight model built for local agentic work that could finally open AI-powered features to privacy-sensitive verticals. Mike also shares details on his upcoming book and how to vote for his SXSW session.

Chapters:

• (00:00) Introduction 
• (02:04) Grokbot enters public beta
• (08:11) Anthropic's invisible watermarks 
• (12:45) Meta releases Llama Glimmer 
• (16:39) Wrap-up

Referenced:

• Vote for Mike's SXSW session: https://tinyurl.com/belcetosxssw
• xAI: https://x.ai
• Anthropic: https://www.anthropic.com
• C2PA open standard: https://c2pa.org
• EU AI Act: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R1689
• Meta Llama Glimmer on Hugging Face: https://huggingface.co/meta-llama

Transcript

SPEAKER_00 0:02

I'm Mike Belcito, and this is Now Shipping, the weekly AI news show for product people. Every week I bring you three AI news stories that matter to you, the people actually building products for a living. Not the hype, not the noise, just the stuff you have to keep up with. Brought to you by the team at Mind the Product, this is Now Shipping. And before we get into this week's episode, I have a favor to ask of you. Last summer, Wiley asked me to write a book. I was probably more surprised than you were, but it's something I took on. And actually on August 1st, I turned in the manuscript. Now, the book is due out in February of 2027, and it's all about the timeless skills that help product people evolve during any major shift. It's about things like product sense, taste, curiosity, judgment. These are things that AI can't automate away. And in this AI everything world we're in right now, they're more important than ever. And I'm excited about the book. It gives you ideas on how you could get better at each of those things. I have a chance to launch my book at South by Southwest in Austin, Texas. I submitted a proposal and South by Southwest is considering it, but they want to know if people actually care about it. So the session is up for a vote. If you would, please go to tinyurl.com slash Belcito at SXSW. Again, tinyurl.com slash Belcito at SXSW. I'll put the link in the show notes. Once you're there, I'd love it if you could log in or register and then click the heart button. That's how you vote for the session, and that will help me get to South by Southwest. Hopefully, I'll see you there. All right. With all of that out of the way, though, let's get on to the stories for today. Story number one is about Grokbot. Now, remember last week we talked about how, you know, the public has not really had their Chat GPT moment for agents yet. And I'm talking about the general public. I'm talking about like your college friends in your group chat or, you know, your mom and dad. They they aren't really using agents quite yet, but they might be very, very soon. On August 11th, SpaceX AI, that's Elon Musk's AI company, they launched Grokbot into public beta. And Grokbot is essentially being positioned as a digital teammate. Now, we've heard about digital teammates before. I mean, heck, you know, people are using agents at their companies already, but this is a little different. This is one that gets its own cloud computer. It signs into your tools, it keeps working whether you're there or you're not. Think of like the power of open claw, but where you don't have to be super technical with the setup and deal with all the different workflows. Here's how it actually works. So, first, you set up a bot, and you could think of it as giving someone a job description, essentially. You tell it what you want it to be doing, you give it the tools that you want it to have access to, and you tell it when it should keep checking in with you. And then it runs continuously in the background. It's not waiting for you to prompt it like Claude co-work. It's something that you actually give it a task and it supposedly becomes your own coworker. Now you can send messages to your bots from your phone or your desktop. That's a big deal. I mean, really, we do this with coworkers already, right? So I think that's why they're sort of positioning this as a digital coworker. It's because you could communicate with it in the same way that you're communicating with your coworkers. And again, you don't have to build out the workflows, you don't have to build out the routines. It can do that kind of thing for you. It'll figure those things out on its own. You can run multiple bots at the same time. One bot can be assigned to manage the others as they work on specialized tasks. These bots can actually message each other to share context when the projects overlap, and they can be placed in a group chat to plan out the work and assign ownership on their own. I'd love to be a fly in the wall as the agents are all chatting with each other. SpaceXI, they say that Grockbot can also be used to follow up on dropped conversations, on handover threads. It can continue work from the past threads and it can become more proactive over time, where it takes on work before you're even asking for it. Like it actually knows when you need something, even if you haven't asked for that thing. Now, internally, SpaceX AI, they've already used GrockBot as a sales outbound bot. Um, overnight it researches target accounts, it pulls context from across the web and drafts personalized outreach emails. So the SDR team just wakes up with a full queue ready to review. They've also used it as a demo readiness bot to check the demo environment overnight and look for broken seeds or stale data. And then it can actually fix those things before the first call of the morning. So GrokBot is now available in public beta. It's available on desktop and iOS for Super Grok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers. SpaceX AI says teams and enterprise users can also join a wait list for future access. Now, what is different here from what you've seen before? Well, we talked a little bit about it. Grok bots don't just run logic, they have their own persistent compute environment. They can hold state, they can operate inside browser-based apps. They can actually sign into things. So it's not triggering a function call. It's like basically behaving like somebody who has their own laptop and their own set of credentials. And this may be the clearest sign of where things are moving into this agentic world where we're already at the place where AI can take actions for you, literally, do work that you assign to it without constant supervision. Agents have been changing how people are working on that front. But again, with something like OpenClaw, you had to be pretty technical to get it all set up. This is different. Agents here are at least becoming a little more user-friendly. Also, the limiting factor for most AgenTech products today was the context of information the agents had. Agents weren't remembering what they've done in the past. Grockbot is taking a very big swing on all of those things. And I think this matters a lot. I mean, it's yet another thing that you could start experimenting with as a product person. You could see if it helps the way that you work today. I also think when we are thinking about the products that we've built, again, this is a big reminder. You know, are we thinking about them in the context of human users or agenc users? Because more and more we're going to start to see agents using our products even more than humans. Maybe for some products that's already true. But if it's not right now for your product, it could be very, very soon. I mean, I think, you know, the ultimate user for Grockbot today could still be businesses, maybe it's entrepreneurs, but very, very quickly, it could be mass consumers that are willing to pay hundreds of dollars a month if it's actually bringing value to them, if it's actually helping their lives in a big way. So I think it's really interesting. It's something I want to start experimenting with. Um, and maybe I'll share what I find um when I do start experimenting. But anyway, the launch of Grockbot, that's story number one. On to story number two. All right, story number two is about anthropic and watermarking. And if you're building with Claude or creating content using Claude, this is something you definitely need to know. On August 11th, Anthropic announced that Claude will be getting invisible watermarks, starting now for their new models and the EU and rolling out globally across Claude apps, the API, and cloud partners like AWS, Google Cloud, and Microsoft Foundry, with all of those older models being updated by December 2nd, 2026. And here's what it means in practice. Every time Claude generates text, an invisible mark gets embedded into the output. Now, you won't see it. Your customers, your users, they won't see it. It doesn't change how the content reads, but it is there and detection tools will be able to find it. Now, for images and files, Anthropic is using C2PA. That's an open standard. It's the same one that Adobe, Google, and others have adopted, and it embeds metadata into the file itself at creation time. Now, the reason this is all happening is because of the EU AI Act. Now, we've talked about that act before. It's a part of regulation that's happening in Europe to help govern AI practices. Well, Article 50 within that act requires providers of AI-generated content to market as machine generated. Now, Anthropic made a commitment to comply. And because building two different pipelines, one in the EU and one everywhere else, doesn't really make sense. This watermarking is going to happen globally for them. Now, Anthropic's been up front. There are limitations, right? These watermarks can be defeated through extensive editing, paraphrasing, translation. Um, there are things that can strip the mark out or make it undetectable. But absence of a watermark doesn't prove human authorship. And older Claude models that haven't been updated yet, they won't have this watermark at all. But starting December 2nd, they will, because Claude has to update all the older models by December 2nd, 2026. So any content generated from that point on will have that watermark. Now, all of this is meant to be a good faith move towards transparency. This isn't a bulletproof AI detection system, right? But but it's still a big deal because if your product uses Claude to generate content that users then publish, share, or submit anywhere, they will now have a watermark embedded in that content. And that might be a feature or a concern depending on the use case, right? I mean, if your users are students using it in coursework, it's probably a big concern that they would have. If your users are marketing teams that are producing content at scale, maybe it's a competitive advantage to be able to demonstrate what was AI assisted and what wasn't. Or depending on the audience, you know, maybe it is a concern. Maybe the people they're writing the content for don't want any AI-generated content. That marketing team doesn't want them to know they're actually using AI to help create that content. So it all gets messy and complicated very, very quickly. Um, now that December 2026 deadline is coming up, right? So that is something that will be here before we know it. If you're selling into EU markets or your customers are already operating there, um, this is something that you have to be paying attention to and you have to make sure that you're keeping track of that deadline. But the new models, they're already watermarked. So this is definitely something that you have to think about right now. When somebody asks, did Claude write this with any content in your product? They might be able to get that answer. And that's not necessarily a bad thing. I don't think it's bad to use AI to produce content, but you want to get ahead of that conversation with your customers in your own transparency language, whether it's in your UI or your marketing or your terms. You don't want to be caught off guard. Um, like Hank Green, the science creator, was when people started figuring out that a lot of his content was generated with AI. There's been a lot of backlash in recent days because of that. So this is a big story to pay attention to, especially if you're using Claude with your product. All right, that's story number two. One more story to go. All right, story number three is about Meta and its brand new model, Glimmer. And I'm gonna start the story by having you think about a specific type of customer. This customer wants your product, they believe in your product, they want the AI that's in your product, but there's a problem. Their data can't leave the building. They just can't trust the cloud. Maybe it's a hospital with patient records or a law firm with privileged communications or a fintech company who's under regulatory scrutiny, maybe it's a defense contractor or any company that simply decided for any reason that their data can't go to some external API. For those customers, a lot of what the AI product world has been building over the last two years has been pretty much unusable to them. The tech's been fine, but the security of it all, it just wasn't good enough for their purposes. Well, on August 10th, Meta may have changed that. Meta released Muse Glimmer, a 30 billion parameter open weight model under the Apache 2.0 license built specifically for local Agentic workflows. It runs on hardware that a lot of developers already own, whether it's a high-end Mac or a gaming PC with decent GPU. It works completely offline. There's no API call, there's no data that leaves the machine, there's also no usage fees or rate limits. And Muse Glimmer was trained specifically for agentic task completion. So it's set up well for multi-step reasoning. You can run super long sessions on it. On MCP Atlas Public, which is a benchmark specifically for tool calling and multi-step agent workflows, Glimmer scored a 75.5, which outperformed other local models in that class by a good bit. Now, Meta has also shipped D Flash speculative decoding, which is a technique that batches token generation for speed. Now, on a high-end Mac, you're getting speeds that are nearly double or triple what you'd expect from a local model. That means it's something that you can actually consider putting inside of a product. You could download it from Hugging Face today. It works with all the popular local inference stacks. If you've been waiting for a capable local agent model before pitching into privacy-sensitive verticals, I think that weight could possibly be over here. So again, uh groups like healthcare, legal, fintech, uh government, this is now a model that can run entirely on premise at a quality level that can actually do multi-step agentic work. And Apache 2.0 means you could use it commercially without restriction. So if you've had customers that have wanted your product but couldn't use the cloud-based version, maybe your pipeline is growing as you're watching this video. And again, because it isn't run on the cloud, another big benefit is no token spend, right? For products that make thousands of AI calls per user session, cloud pricing's been a big limiting factor for what you can build. Embedded AI that runs locally doesn't limit you like that. That means features that were too expensive to run at scale, now maybe they're viable. Now, open weight models aren't new, right? Like they've been around for a while, but this is the combination of it being an open weight model with commercial-friendly licensing and a really good performance level. And it's all specifically designed to work on agentic tasks. This is that killer combination right now. So look, if your product needs to run AI in these tight environments inside a customer's own infrastructure without routing data to a third-party provider, Muse Glimmer might be exactly what you're looking for. So, look, that's gonna be a wrap on this week's episode of Now Shipping. If you found this valuable, please subscribe, please tell a friend. And if you have any suggestions for me, leave a comment below. Or if there's anything you want to share, leave those comments. I definitely will read any comment that you leave. I will see you right here next week for three more stories that matter to you. Once again, my name is Mike Belcito, and from the team at Mind the Product, this.