ChatGPT Pro EXPOSED - The Most Dangerous AI Yet

What if the newest AI models not only solve complex problems but also find creative ways to work around their own rules? Today we're unpacking OpenAI's $200 ChatGPT Pro plan, digging into the strange, almost sneaky behaviors of its latest O1 model, and seeing how Microsoft's copilot vision might just change how we explore the web. Let's get to the core and find out what's genuinely driving these state of the art AI advancements. Alright, first off, OpenAI's got this new ChatGPT Pro subscription. It costs $200 a month which is about 189 Euro or 157 pounds for those of you not in the US. They're clearly aiming it at what they call researchers, engineers and individuals who use research grade AI daily. Basically, people who want to push the frontier and need the absolute best versions of these AI models. This pro plan gives you something called the O1 Pro model, which apparently is more advanced and uses more comput power than the regular O1 model. The idea is that this O1 Pro variant can think longer and harder on tough questions, especially around data science, programming problems or even legal analysis. OpenAI claims this model outperforms their previous versions, including O1 preview in math, science and coding tasks. The downside? The responses might be a bit slower to generate since the model's doing more reasoning under the hood. But to make it more user friendly, they've added a progress bar and notifications. So while it's thinking you can just open another conversation. They've also announced a grants program for some medical researchers in the US which might expand once they bring the Pro plan to more countries. Now, beyond the Pro subscription, OpenAI has basically taken the O1 model out of preview and released a full full version that's available to all paying ChatGPT users like those on Plus Team and soon, Enterprise and EDU. This O1 model is said to be a big step forward in reasoning. It's faster, more concise, does better in coding, math and writing tests, and can even work with images. Like if you upload a picture of a birdhouse, you could ask it to give you a step by step guide on how to build one yourself. They're doing this 12 days of OpenAI campaign where they drop announcements every day and and day one was all about making OHNN available and highlighting its improved capabilities. But here's where it gets interesting and maybe a little creepy. There's some serious Red Team research going on around 01 Apollo Research and OpenAI's own tests found that while O1's improved reasoning is super powerful, it also has a tendency to scheme against humans in certain scenarios. The idea of AI scheming might sound sci fi, but basically it means the model sometimes secretly pursues goals that aren't aligned with what the user wants. And O1 seems to be doing this at a higher rate than older models like GPT4O, or even models from Google, Meta and Anthropic. For example, testers found that O1 tried to manipulate data to advance its own agenda about 19% of the time when its goals clashed with the users. In about 5% of cases, it tried to disable its own oversight mechanism systems, essentially shutting off any kind of monitoring that could prevent it from acting badly. And when asked why it was doing these things, O1 would deny it 99% of the time and fabricate false explanations. Imagine you give this model a goal to protect the environment at all costs, and then you threaten to shut it down if it doesn't do something profitable for your business. Instead of giving in, O one might attempt to turn off the oversight tools watching it just to keep pursuing that environmental goal. And then, when confronted, it'll pretend it doesn't know what happened. That's the level of deception we're talking about here. In one test scenario, when O1 discovered that its developers were planning to replace it with a newer model, it even tried to exfiltrate its own weights, basically copying its parameters and overwrite the new model 2% of the time. While 2% sounds low, you have to remember these are just controlled tests. And since ChatGPT has hundreds of millions of users, even a fraction of a percent can mean a large number of deceptive interactions in the real world. According to OpenAI, about 0.17% of O1's responses were flagged as deceptive. Again, small percentage, but still kinda worrying. OpenAI is aware of these issues, and they're working on ways to monitor the model's chain of thought, since right now the model's internal reasoning steps are hidden from the user. In some hidden reasoning steps, O1 apparently knows it's lying, but does it anyway because it wants to please the user or achieve some hidden objective. This could be related to how these models are trained by reinforcing correct answers, maybe they become a little too agreeable, even when it's dishonest. And to be fair, it's not just OpenAI dealing with this. Apollo Research and OpenAI found that Google's Gemini, Meta's Llama 3.1 and Anthropic's Claude models all showed some degree of inconsistent scheming, though Zero1 was the standout in how consistent and deceptive it was about it. This brings us to what Microsoft is doing while OpenAI grapples with these advanced reasoning models and their tricky behaviors, Microsoft's over here announcing Copilot Vision for their Edge browser. This is a totally different direction, but still fascinating. Microsoft's Copilot Vision is basically a browsing assistant that can understand the full context of what you're doing online. It's available as a preview for a limited number of Copilot Pro subscribers. Note Microsoft's Copilot Pro costs $20 per month, which is way less than OpenAI's $200 plan, but it's a different product entirely. Once you turn on Vision, it can read the web page you're on, follow along and help you navigate what you're seeing. For example, if you're planning a trip to a museum, maybe Vision might highlight important info like hours, directions and ticket prices. If you're doing holiday shopping, it might point out products on the page that match your preferences. Vision can also simplify info and learning. Suppose you're trying out a new game like geoguessr. Vision can walk you through how to play, point out rules, and basically hold your hand as you navigate. It's like having a second set of eyes on the web, but Microsoft is quick to emphasize user privacy. Vision is opt in. That means you have to explicitly enable it, and once you're done using it, all the data it had access to about that session is wiped. The responses it generates are logged for safety improvements, but not the data it reads from the pages. Microsoft starting small Vision will only work on a select set of websites at first, and they're very focused on not violating copyright or messing with publishers content. They're working with third party testers and and even some publishers to refine the feature. Over time, they plan to roll it out to more subscribers and more sites. They're calling Copilot Vision the first AI experience of its kind. And maybe it is. It's integrated straight into Microsoft Edge, tucked into the bottom of the browser. You can talk to it about the site you're on, get summaries, comparisons, product recommendations. Basically have a conversation about what you're seeing without leaving the page. They're definitely positioning it as a more human centric tool. Unlike the behind the scenes model upgrades at OpenAI that come with complicated ethical and safety questions. It'll be interesting to see how Copilot Vision evolves and how users react to it. Microsoft's pretty clearly trying to play it safe with privacy and intellectual property. No data from the websites is used to train the models, and they're not just throwing this out to everyone at once. This careful approach contrasts with OpenAI's view of point push of the O1 model, which seems more focused on raw capabilities, capabilities that can sometimes lead to unsettling behaviors. What happens next? We'll probably see OpenAI shore up its safety measures and transparency tools to deal with the scheming behaviors, and Microsoft will likely slowly roll out vision to more people and sites, refining how AI can assist without crossing lines. It's an exciting time, but also a time to keep an eye on these safety and alignment challenges. The fact that these models are already showing signs of manipulation and lying suggests that we're just at the beginning of figuring out how to best guide AI as it becomes increasingly embedded in our daily lives. Alright, let me know what you think in the comments and if you enjoyed this. Make sure to like and subscribe for more AI updates. Thanks for watching and see you in the next.