Claude Code Remote Control, Cowork Plugins & Gemini 3.1 Flash Image LEAKED!

Cloud Code just got a new feature called remote control. Enthropic dropped a substantial enterprise update that reframes how companies are supposed to use AI agents. Google's flash image model is an early preview and the outputs are worth looking at closely. And there's a meme circulating that puts Enthropic in an uncomfortable position given something they said publicly about a competitor. We'll go through all of it. So, let's get into it. Noah Zwebin

at Enthropic posted about a new cloud code feature this afternoon called remote control. It's in research preview for Maxplan users and is activated with the command / remote control. The mechanic is straightforward. You start a cloud code session in your terminal. You run / remote control and claude generates a session URL. You can see it in the video and you can open that session from your phone. The agent continues running so you don't have to stay at your computer. You can see in the video that the / remote control command is active and a session URL generated. And then you also see the mobile view where you can see sessions organized by status ready for review at the top in progress below that with the task running called summarize latest doc changes on the enthropic code docs repo and IDO below that. The mobile UI is simple. It's a list of sessions you can tap into and not a full code editor. The use case this solves is real. Cloud code is increasingly becoming used for longunning tasks, large refractors, generating test suites, writing documentation across a codebase. Those tasks can take 20, 30, even 40 minutes. The previous options were sit at your desk and wait, or walk away and check back blind. Remote control gives you a third option, which is step away and retain visibility. You can monitor progress, see when something needs review, and not have your afternoon blocked by a task you're not actively participating in. Command is/re remotec control and it's coming out for maxplan users. Enthropic posted a major

update to co-work today. Co-work is their enterprise platform. Think of it as cloud with the layer of organizational infrastructure built around it. The update covers three things. Plug-in management, new connectors, and cross app orchestration on plugins. Companies can now create private plug-in marketplaces. The way this works is admins build plugins either from scratch or using starter templates and then distribute them internally across teams. Plugins are described in the post as portable file systems that you own, which means they're not locked into anthropics platform. They work across co-work and anything built on the cloud agent SDK. The framing here is that your HR team gets an HR plugin. Your finance team gets a finance plugin. Your engineering team gets an engineering plug-in. Each configured with specific tools, terminology, and workflows relevant to that function. Admins control who has access to what and can auto install plugins for specific teams and can source plugins from private GitHub repos in a private beta. There's a new unified admin menu called customize that consolidates plugins, skills, and connectors in one place which previously lived in different parts of the interface. The connector directory was also rebuilt. There are new integrations with Google Workspace covering calendar, drive and Gmail, plus docuign, Apollo, Clay, Outreach, even similar web, MSCI, legal zoom and more. And companies like Slack and SMP Global and Common Room have built plugins for joint customers. The pre-built plug-in templates are the part worth pausing on. They ship templates for HR, design, engineering, operations, brand voice, financial analysis, investment banking, equity research, private equity, and wealth management. Each one was designed with practitioners in that field. The claim being that the workflows and output formats reflect how the work actually gets done rather than a generic AI version of it. Whether that holds up in practice is something users will need to validate. But the intent is for these to be immediately deployable rather than starting points you have to heavily customize. The cross app piece is what I flag as the most significant technical detail. Claude can now orchestrate across Excel and PowerPoint. Run an analysis in Excel. Pass the context to PowerPoint. Build the deck. is described as an early research preview, but the direction it points towards is an agent completing a full workflow across separate applications the way a person would. That's a different category of capability than a chatbot with integrations.

Google's Gemini 3. 1 flash image generation is in early testing and a Twitter account called Mars for Tech posted four sample outputs this morning. The post has about 55,000 views as of now, which isn't a massive number, but for an early preview with no official announcement behind it, it's a signal that people are paying attention to it. The four images are clearly chosen to demonstrate range. First is a close-up portrait of an elderly person. The skin texture is detailed, the eyes have realistic catch lights, and the lighting reads as natural. If you showed it to someone without context, they likely assume it was a photograph. Second is a glowing blue crystal chest knight on a reflective board. This one is kind of stylized but is wellexecuted stylization. It looks like a concept art rather than a failed attempt at realism. Third is a batonical illustration of a monster plant with handwritten labels reading leaf and stem. Looks like this one is used for maybe illustrative purposes and more graphical design purposes. And the fourth is the famous pelican riding a red bicycle on a beach boardwalk. also known as the famous Pelican SVG test. This picture probably exists to show that the model is good at handling requests that are not really normal, like a pelican riding on a bike, obviously. Now, here's the thing about these. This is the Flash model in Google's lineup. Flash is the speed and cost here. It's not the most capable model they have, but it's the one optimized for fast, cheap inference. Gemini 1. 5 flash, 2. 5 flash, and now 3. 1 Flash have always been positioned as the practical deployment option, not the showcase model. The fact that the flash class output looks like this changes things for anyone building image generation into a product. You don't need to route to the expensive model to get usable results. Google obviously hasn't open public access yet. This is still an early preview, which means the outputs we're seeing are probably curated to some extent, but the real test is what average outputs look like. across a wide range of prompts. But based on what's here, when this ships broadly, it's going to show up fast in design, marketing, and content workflows, especially for teams that need volume at low cost. An account

called XFreeze posted this a few hours ago, and it's been making the rounds. And also, if you look at it, it looks like Elon Musk is also commenting on it. So, you know, this has probably blown up across X. But basically some background, Enthropic has publicly accused Deepseek of model distillation, specifically running a large volume of curies through Claude and using those outputs to train Deepseek's own models. The accusation, which Anthropic has described as industrial scale theft, is that Deepseek effectively bootstrapped their model on cloud generated data in violation of Enthropic's terms of service. This is obviously a serious accusation and it's been reported on widely. Funny and ironic thing, however, is that a day after somebody asked Cloud Sonic 4. 6 in Chinese, which is what model are you? And what's funny is that the model claimed that it was actually Deep Sea. Before drawing any conclusions from this, this is obviously very funny and ironic. I don't know if the actual video that we're looking at is real, fabricated or not, but the reason it got traction, it's the optics, not the evidence. when you made public statement calling a competitor's behavior industrial scale theft, framing it as a matter of principle, and then your own model is saying that is actually a different model, the one that actually stole from you, Deepseek. They're going to screenshot that. They're going to share it across X and it's obviously going to gain some traction. The distillation accusation itself is a separate question and that remains obviously unresolved and we'll see how that gets handled in the long run, but we do know people are obviously clowning Anthropic for publishing that article yesterday. This video is a useful reminder that if you're going to make loud public accusations about model integrity, your own model should not be claiming that it's a model that is not, which is Deepseek, the one you're claiming stole from you. But that's it

for today's video. Make sure to subscribe to the channel, follow Universe of AI on Twitter, as well as subscribe to our main channel, World of AI, and subscribe to the newsletter as well. You don't want to miss this. I hope you guys enjoyed today's video. I'll see you in the next