Transcription
[Music] AI just passed a human test meant to keep it out. OpenAI's CEO says GPT-5 feels like a nuclear bomb. And Meta offered $1 billion to a single researcher who said no. Yeah, this weekend AI, it's absolutely unhinged. From ChatGPT's new update to AI agents clicking "I'm not a robot" without blinking, things are getting wild. And after all that, we'll dive into even more breaking AI news. From Google and Microsoft's latest upgrades to Nvidia's record-breaking model and Adobe's new AI magic in Photoshop. So, let's talk about it.
All right, let's kick off with the one that affects the most people right now. ChatGPT's brand new Study Mode. If you've ever used the chatbot for homework, studying, or even prepping for an exam, you already know how tempting it is to just get a full answer instantly. Ask it about the Pythagorean theorem and it'll just hand over a polished explanation or even write your full assignment for you, no questions asked, which sounds helpful, but if you're trying to learn, it kind of defeats the purpose. Study Mode changes that completely. Now, instead of just giving answers, ChatGPT actually walks you through concepts step by step, like a real tutor that doesn't sleep, doesn't judge, and never runs out of patience.
It starts by asking you what you're trying to learn, how much you already know, and then it tailors the explanation to match your level. So, if you're struggling with a topic like sinusoidal positional encodings or discrete math, it's not just dumping code or definitions at you, it's explaining them in a structured, layered way while also throwing in some self-check questions, hints, and prompts to make sure the info sticks. It even adapts based on your past conversations. So if you've been working on the same subject for a while, it remembers and builds on that. And it's not just answering questions. It's actively guiding the learning process. Think Socratic questions, breakdowns of complex ideas into manageable chunks, and feedback designed to help you reflect as you go. Honestly, the structure feels like a mini-course.
And this isn't just AI guessing at what might help. OpenAI built it with input from teachers, scientists, and pedagogy experts to make sure it aligns with actual learning science stuff like metacognition, cognitive load management, and curiosity building. Now, this isn't just about making students smarter. It's also OpenAI's response to a rising problem. Last year alone, United Kingdom universities reported nearly 7,000 confirmed cheating cases tied to AI tools. That's more than five per thousand students. Just a year earlier, that number was closer to 1.6. So, yeah, it's exploding. And considering over a third of college-aged adults in the United States are already using ChatGPT and a quarter of the messages it handles involve school or tutoring, it's a pretty obvious pressure point.
But even OpenAI admits this won't stop cheating entirely. Students can still ignore Study Mode and just ask for full essays. That's why they're saying this needs to be part of a broader, industry-wide shift: schools rethinking how they assess students and maybe even building in AI awareness into their testing systems. Basically, they're trying to set the tone for what responsible AI use in education looks like. And honestly, it's long overdue. One college student, Maggie Wang, said she used Study Mode to finally wrap her head around sinusoidal positional encodings, a concept she'd failed to understand multiple times before. After a three-hour session, she finally got it. She compared it to a tutor who never gets tired. And that's pretty much the vibe Study Mode is going for.
Now, let's jump into something that feels straight out of a satirical Black Mirror episode. ChatGPT's agent clicking "I am not a robot." Yeah, that actually happened. So, the agent is part of ChatGPT's growing toolbox. It can browse the web through its own isolated virtual environment, meaning it has access to a browser and operating system that interact with the real internet. You give it a multi-step task like downloading a video or ordering groceries, and it performs the steps for you. It even narrates what it's doing, which is both useful and, in this case, a little ironic.
While completing a task involving a Cloudflare-protected page, the agent encountered one of those little boxes that says, "Verify you're human." You know the ones. Click the checkbox to proceed. And without hesitation, the agent clicked it, then explained what it just did. "This step is necessary to prove I'm not a bot and proceed with the action." A bot. Literally saying it needs to prove it's not a bot and passing the test. You couldn't script that better.
Now, to be fair, it didn't get hit with the full CAPTCHA challenge, the one where you have to click on blurry traffic lights or storefronts. It only passed through the initial layer, which is called Cloudflare Turnstile. That layer analyzes all kinds of subtle behavioral signals: mouse movement, browser fingerprint, IP history, to decide whether you're likely a human, and the agent's behavior was humanlike enough that it didn't trigger the deeper check. This is a big deal because CAPTCHA systems were specifically designed to block bots like this. The idea dates back to the 1990s. Give humans a simple visual test that machines supposedly can't pass.
Over time, the CAPTCHA arms race escalated. Better AI meant more sophisticated CAPTCHAs. Some even use the process to train machine learning models. Google's reCAPTCHA, for example, had people helping digitize books and train vision algorithms, all under the guise of proving they weren't bots. Kind of brilliant, kind of dystopian. But the fact that an AI agent can now casually breeze through even the behavioral layer of a CAPTCHA shows how much has changed. And this wasn't just some brute-force hack. It was smooth, self-aware, and integrated into a broader workflow. That's what makes it so different. You're watching the agent behave more like a human assistant, navigating, clicking, reading, and not just running scripts.
Someone on Reddit even had the agent order groceries from a local store. The instructions were super basic: avoid red meat, stick to healthy items, and stay under $150. It pulled it off, no problem. But of course, not everything's flawless. Another user said the agent completely failed to navigate the Stop and Shop website because the interface was too messy. So yeah, bad UI still wins against good AI sometimes.
Now, let's get into what might be the most shocking part of all this. OpenAI CEO Sam Altman just said in an interview that GPT-5 scares him. He wasn't vague about it either. He said testing GPT-5 reminded him of the Manhattan Project, the development of nuclear weapons. Yeah, that's the comparison he made. He talked about how fast GPT-5 feels, not just in how quickly it responds, but in how much it seems to understand. There were sessions where he just sat there watching what it could do and felt deeply uneasy. And remember, this is the guy at the very top of the company that built it.
Altman also took a shot at the state of AI oversight. He said there are no adults in the room, meaning the systems that are supposed to be regulating and monitoring AI development are way behind. It's moving too fast and the people in charge don't have the tools or the knowledge to keep up.
And just when you think the stakes couldn't get higher, let's talk about Meta. So Mark Zuckerberg, who's been trying to poach top talent for his new Meta Super Intelligence labs, just hit a brick wall. He went after Mira Murati's team at Thinking Machines Lab and offered them outrageous deals. We're talking hundreds of millions, up to $1 billion for a single researcher, not over their lifetime, just over a few years. It's one of the biggest offers anyone's ever heard of in this space. And every single person turned him down. According to a report in Wired, no one at Thinking Machines Lab took the offer. Some were offered $200 to $500 million over 4 years. One person was offered a full billion. They all said no. That's not just a flex, it's a statement. A rejection like that at that scale signals that this isn't just about money anymore. These researchers are making decisions based on values, alignment, maybe trust. It says a lot about how they view Meta and possibly how they view Zuckerberg's goals with this new super intelligence team.
All right. Now, while OpenAI and Mira Murati dominated headlines, there's been plenty happening elsewhere in AI too. Ideogram introduced a tool for generating consistent characters from a single photo. Microsoft rolled out Copilot mode in Edge with multi-tab analysis, voice control, and task execution. Google expanded AI search to handle PDFs, real-time phone video, and launched a new planning feature called Canvas. Nvidia's Llama NeMo Super version 1.5 just topped reasoning benchmarks. And Adobe upgraded Photoshop with smarter blending, upscaling, and object removal tools.
Starting with Ideogram, this update is a big step for anyone who works with visual content. The new Ideogram character feature lets users upload a reference image and generate that same person or character in different styles, scenes, and poses. It's not just another face generator. It's built for consistency, which means the face, hairstyle, and even lighting can stay locked across multiple outputs. For creators working on comics, virtual avatars, or personal brand visuals, that kind of stability is a game-changer. You can also mask or unmask areas like clothing, neck, or hair depending on what you want to preserve. If you're going for a stylized look, there's a fiction mode that pushes the output into more imaginative territory. And when you combine it with Ideogram's existing tools like Magic Fill or Remix, you can actually place that character into custom scenes, adjust lighting to match the environment, and control how closely the output should follow the reference. It's all available now in early access directly from the platform's character tab.
Over on the browser side, Microsoft is starting to reshape how we interact with the internet. The new Copilot mode inside Edge is more than a side panel; it turns your browser into an active assistant. Once enabled, it can read across all your open tabs and provide summaries or comparisons, which is incredibly useful for research or side-by-side reviews. For example, if you're shopping or analyzing competitors, it can highlight differences and surface insights from multiple pages at once. They also introduced a unified input bar that merges search, chat, and navigation into one place, cutting down clicks and switching. Voice navigation is already working, too. You can just talk to Copilot and ask it to find content, compare pages, or open new tabs. And while those features are live now, Microsoft is already testing more ambitious stuff like letting Copilot complete real-world tasks for you, group your browsing into topic journeys, or offer smart suggestions based on what you're currently working on.
Meanwhile, Google is pushing deeper into multimodal search with major upgrades to AI mode. You can now upload images directly from desktop, something that was previously mobile-only. And they're about to launch PDF uploads, too. That means you'll be able to ask questions about actual documents like lecture slides or white papers, and get direct answers combined with web context. One of the more interesting additions is Canvas, a planning tool that opens in a side panel and stays active across sessions. Whether you're building an itinerary, organizing research, or tracking a task list, it gives you a persistent workspace where your search evolves with your goals. On mobile, Search Live is also rolling out. This lets you point your camera at an object or scene and talk with AI mode in real time based on the live video feed. It's built on top of Project Astra and accessed through Google Lens. Chrome's getting smarter, too, with a new "Ask Google about this page" option right in the address bar. Plus, deeper AI responses when you highlight content or tap "Dive deeper."
Switching to models, Nvidia's Llama NeMo Super version 1.5 just made a strong entrance. It's topping the Artificial Analysis Intelligence Index leaderboard, beating out other open models in math, science, reasoning, coding, and chat. What's different here is how efficient it is. It runs on a single Nvidia H100 GPU but still outperforms in both accuracy and speed. This makes it extremely practical for anyone building serious AI agents or assistants that need to think through complex tasks. The model was trained using entirely synthetic datasets: over 26 million high-quality rows generated from models like Qwen 3 and Deepseek. It went through several layers of post-training from supervised fine-tuning to advanced reinforcement learning techniques like DPO and RLHF. All of that was aimed at improving reasoning, function calling, instruction following, and more. And soon you'll be able to deploy it using NVIDIA NIM microservices with minimal setup.
And finally, Adobe just dropped a set of AI updates for Photoshop that'll save a ton of time for creators. The new Harmonize feature lets you insert objects into a scene and automatically matches the lighting, tone, and shadows so it looks like it was always part of the image. Generative Upscale improves resolution up to 8 megapixels without making it look oversharpened or artificial, which is a big help when printing or working with older files. They've also upgraded the Remove tool using their latest Firefly image model, so object removals are now cleaner with fewer weird edges or blurry patches. And on desktop, there's a new Projects feature that organizes assets into shared workspaces, making it easier to collaborate without losing track of versions or edits.
All right, your turn. Is AI crossing the line when it starts clicking "I am not a robot" and actually gets away with it? Let me know what you think down below. Make sure to like, subscribe, and turn on notifications. Thanks for watching. Catch you in the next one.