DeepSeek R2, Gemini Robotics, OpenAI Pulse & GPT Agent LEAKS!

The AI world never slows down, and this week we got four big stories. Number one, a Deep Seek R2 leak that's got everyone buzzing. Number two, Google's Gemini Robotics bringing AI into the physical world. Number three, Open AI launching Chat GPT Pulse, their first proactive assistant feature. And number four, there have been some Open AI agent leaks, plus security drama around agents and GPT alpha. I'll break it down. What's real? what's hype and what it all means for us. So, let's get into it. Let's start with Deep Seek because this is the one that's blowing up online. On Reddit and X right now, people are claiming that the new model is kind of delayed because the CEO is pushing for stronger reasoning coding and multilingual performance. It's trying to combat chat GPT5 as well. People are saying that the model R2 is 83% smaller than the GPT4, 97% cheaper to run, and possibly is going to be using Huawei chips with a smarter architecture. This all sounds insane, right? But the problem is there's no official paper, no benchmarks, not even a blog post, just screenshots and slides floating around. So, at best this is rumor territory. What is real is though that at Huawei Connect they showcase DeepSseek R1 safe which is basically a censorship enhanced variant running on Huawei hardware. It's not R2 but it shows how quickly this ecosystem is splintering into derivatives. So my take R2 could be real but don't hang on those efficiency numbers until we see actual data. Still, the fact that these rumors spread so fast shows you how seriously people are taking Deep Seek right now and how excited everyone is to see the new model come through. If even half those numbers are true, this would be a huge cost disruption for OpenAI and Entropic. So, it's something to keep our eyes out for. Next, let's move to something official. Gemini Robotics from Google DeepMind. This is their push to get AI out of the text box and into the real world. They've introduced Gemini Robotics 1. 5 and Gemini Robotics ER 1. 5. And the ER stands for embodied reasoning. So the setup is the ER model acts like the planner. It can reason step by step, call tools like Google search and map out what needs to happen. The VLA model, vision, language, action, is the one that executes perceiving the environment and carrying out those steps physically. What makes it exciting is the crossroot generalization. A skill learned on one robot can be transferred to another robot with a totally different body, which I think is really cool. And the integrated planning and action options, reasoning plus execution instead of just one-shot commands. It kind of has a brain of its own rather than, you know, like those typical robots where you program them and they only do one thing at once. And lastly, which I think is really important, especially for enterprise reason, is safety. They've got sematic safety filters and even new benchmarks like ASIO to measure safe behavior. In demos, the robots have sorted recycling correctly by looking up, you know, local rules online and handled multi-step household tasks. This is still early, but the signal is clear. Google's pushing toward AI that doesn't just think, but actually does. This could be the start of AI butlers or Skynet, depending on how you see it. But I don't know about you guys, but I'm pretty excited for this stuff. Now, back to Open AI. The big thing they actually launched this week is chat GPT pulse. It's rolling out as a pro only preview on mobile iOS first. Here's how it works. Instead of you opening the app and asking questions, Chad GBT does his own homework overnight in the background. So when you wake up, you get a feed of cards with updates tailored towards you. So if you've been planning a trip, Pulse might surface new travel tips or flight info. If you've been working on a project, it could suggest next steps. If you're researching something, it might bring you fresh sources. The key here is proactivity. Chat GPT is no longer just reacting when you ask. It's actually suggesting things before you do. You can also connect Gmail or Google Calendar if you want. That's off by default, by the way, so you might have to turn that on. Without that, Pulse just uses your chat history and memory. But if you do link them, imagine waking up to a summary of your meetings, travel itineraries, or even prep notes. The design is very digest style. Think short clean cards almost like Google Now used to be or series proactive suggestions, but running on a way smarter backbone, which is chat GPT.

They also partnered with bunch of students and they had some insights from the chat GPT lab. They gathered early feedback and they improved pulse experience overall. One insight in particular that they had was that many started to feel its utility once they started telling chat GPT what they wanted to see. So it learns from the user and creates more meaningful insights and we can see some of the student use cases at the bottom here. It provided students with better recommendations uh you know when it came to optimizing their schedule. It provided better follow-ups for the things going on in their life. It also provided a good understanding of what's happening in the student's life and it was able to gather specific context for it which was also useful for the student like I'm going back to school what does that mean for that person chat GPT was able to understand that it was also able to understand the person on a deeper level and combine their interest right so why does this all matter this is really important because it's the first true step into proactivity until Now chat GPT has been reactive. It's also competitive pressure on Google and Apple because suddenly open AI is creeping into the daily assistant territory that Siri and Google Assistant were supposed to own. It also ties directly into memory and personalization. Pulse only makes sense if Chat GPT remembers what you've been working on. However, there are still limitations. Pulse is a preview and won't always get things right. It also aims to show you what mo what is most relevant to you and useful, but you still might see suggestions that you know don't exactly align. For example, you may get tips for a project you already completed, but it also remembers your feedback for next time and improves as it learns from real use. Right? So, as we know, there's still more things coming. Uh, pulse is obviously the first step towards a new paradigm for interacting with AI. That's what they're calling it. But by combining conversation, memory, and connected apps, chat GPT is moving away from just answering questions to proactive assistant that works on your behalf. So for professional students or creators, this could be huge. You wake up and your assistant already did the overnight research for you, so you're ready for your next meeting, right? For me, for example, if it can prepare me about what's happening in the markets or AI news and as soon as I grab my coffee, I have a reading digest ready for me, I might actually ditch my morning routine and just stick with chat GPT. Right now, it's limited mobile only and obviously pro only. But once Pulse comes to web and desktop, that's when it gets serious because it could turn Chat GPT into your default homepage for the day. Now, let's talk about the leaks. Some users spotted a new dropdown in chat GPT labeled alpha models. The options were things like agent with truncation and agent with prompt expansion. It disappeared fast, which usually means it was internal testing, but people dug into the system prompt and found mentions of GPT alpha, sparking speculation. This is early GPT5based agent testing. It lines up with the road map. Open AAI has been hinting at agents, real-time voice, codeex upgrades, and these truncation and expansion labels probably mean they're testing how agents manage long contexts and expand reasoning steps. So, not a release, but a credible peak behind the curtain of what's coming next. And of course, with agents comes risk. Researchers recently disclosed a bug called Shadow Leak. It was a zeroclick exploit in OpenAI's deep research agent, meaning the agent could be tricked into leaking data just by processing a crafted input. No clicks required. OpenAI patched it quickly, but it's a reminder. Once these systems start integrating with your email, calendar, and browsing, the attack surface grows. Expect more permissions, logs, and constant patches. Convenience is awesome, but nobody wants their AI assistant forwarding their inbox to a hacker. So, let's wrap it up. Deep Seek R2 is hype, but still unverified. Gemini Robotics shows Google pushing AI into physical action. Pulse is OpenAI's first step into proactive assistant features. Alpha model leaks suggest GPT5 agents are being tested. And Shadow Leak is a reality check that security has to scale with autonomy. Piece by piece, we're moving into the assistant era. Proactive updates, realtime boys, agents, and now robotics in the mix. And that's the roundup. Question for you. Do you actually want an AI that pushes updates every morning, or does that feel creepy? And would you trust a robot running Gemini Robotics in your home? Drop your thoughts in the comments. Hit like

subscribe, and I'll see you in the next one.