EP 616 : Runway Solaris, OpenAI ChatGPT Ads $1B, Anthropic Safety AI & AI Heart Diagnosis

Imagine clicking a button on a website, but instead of your browser loading, you know, a pre-written page of code, the website actively dreams itself into existence in real time. Like it's completely reacting to your mouse movements on the fly. Exactly. I mean, no HTML, no CSS, just pure raw video hallucination generating the internet as you interact with it. Welcome to episode 616 of AI Brief. I am your host Amin and alongside my co-host Leah, we are jumping into a stack of developments today that fundamentally break what we think we know about technology. It is a wild lineup today. We're looking at Runway's internet as live video concept, an AI catching heart disease in literally a blink of an eye, and Anthropic's AI successfully automating its own safety research. And when you look at all of these developments together, a very clear overarching theme emerges for today's deep dive. Oh, absolutely. Because we are watching artificial intelligence rapidly move from just generating static content, like a nice image or a block of text, to actively running dynamic interfaces, diagnosing patients in real-world clinical settings, and even policing its own behavior without human intervention. So let's unpack this runway development first, because, well, it might be the most paradigm-shifting visual update we have seen in quite some time. It really is hard to wrap your head around at first. It is, because Runway just launched early access to something they are calling Solaris, which is an interface world model. To understand this, you basically have to throw away everything you know about how an app or a website works. Right, because normally there's code running underneath everything. You have your structural code, your styling, your scripts. But with Solaris, there is zero code running under the hood. The entire interface is rendered as live video, so every single frame is drawn in real-time as you click and drag. You can kind of compare this to cloud gaming, where you stream a high-end video game over the internet instead of running it locally. Except in this case, the entire internet is essentially one continuous real-time hallucination reacting to your mouse. And the actual mechanics behind how they achieve this are fascinating. It's a combination of models, right? Yeah. The system pairs Runway's Gen 4.5 video model with a large language model, and they work together like a tag team. When you interact with the interface, the language model reads your click or your drag, decides what the logical next step should be, and then prompts the video model to generate the next frames instantly. The demos they released are just striking. I mean, in one of them, you grab a virtual shirt from a digital clothing rack and drag it directly onto a photo of a person. And the system just figures out how the fabric should drape over them. Without any 3D rendering engine. In another demo, you build a salad by dropping various ingredients into a bowl, and the video responds to the realistic physics of a tomato hitting lettuce. There is no physics engine coded in, it just inherently knows what that should look like. Runway even ran a blind study pitting these video-rendered interfaces against standard webpages coded by Claude Opus 5, and testers actually preferred Solaris in 71% of the matchups for in-scene behavior. Wait, hold on a second. I have to push back on this a bit, because if every single frame is generated on the fly based on a click, how does this not just turn into a scrambled mess after 10 minutes? That is a very fair point. Like, if I am navigating a complex dashboard, how does the video model remember what the page looked like 5 minutes ago? Doesn't the hallucination just drift completely off the rails? That is actually the exact hurdle they are facing right now. The company openly acknowledged that the current build runs into problems with long-session drift. Over time, the interface might lose its structural consistency. So you end up with screens that look highly convincing at a glance, but are functionally wrong. Exactly, or logically broken. They also noted issues with text legibility, which, you know, has always been a known challenge for video generation models. But despite these early limitations, we really have to look at the underlying trend here. Which is the massive convergence of speed, cost, and quality. Right, and if that improvement curve continues on its current trajectory, it opens the doors for entirely new user interface paradigms that simply were not possible or cost-effective before. So the interface is dynamically imagined in the exact moment you need it, rather than being statically built ahead of time. That is wild. But if Runway's AI is spotting microscopic patterns in user behavior to hallucinate a screen, what happens when we point that exact same pattern recognition capability at the physical realities of the human body? Which brings us to a massive medical breakthrough out of Imperial College London. Exactly, because researchers there have introduced an AI model trained on a staggering data set of 10.6 million electrocardiograms, or ECGs. And this model can read a routine ECG in under two seconds, which is incredibly fast. But what makes this a real breakthrough is that the AI is successfully spotting heart failure and valve disease that are entirely undetectable by human doctors looking at those exact same charts. The numbers on this are amazing. It spots heart failure with an 81% success rate and valve disease with a 90% success rate. Think of a standard ECG chart like a grainy black and white photograph. Human doctors are trained to look for obvious shapes in that photo. You know, they're looking for known irregularities. But the AI is looking at the microscopic grain of the photo paper itself. It is catching warning signs hidden in the static. It is essentially seeing a language in the human heart that our biology textbooks do not even have words for yet. But consider the scale here, because hospitals worldwide run over 1 billion ECGs every single year. And patients usually have to wait months for a specialized ultrasound just to confirm these kinds of diseases. So does this mean we are fundamentally moving from a system of targeted screening to a world where we just catch invisible issues by default? That is precisely the trajectory the medical field is aiming for. The team is already moving forward with a 590 patient clinical trial across six different hospitals. And they want to roll this out broadly, right? Yeah. Their stated goal is to integrate this into routine use within the National Health Service within two years. Having a second set of hyperanalytical expert eyes is going to drastically improve diagnosis timelines using data we are already collecting anyway. So we have AI finding hidden flaws in human bodies. But Anthropic just published some highly consequential new research showing AI finding flaws in its own code. We are definitely shifting from AI assisting doctors to AI assisting and perhaps replacing AI researchers. Which means automated AI safety research is officially no longer theoretical. Amthropic deployed teams of their CLOD agents to run complex safety research on their own frontier systems. And these automated agents successfully trained away 10 different kinds of AI misbehavior. The staggering metric here is that the CLOD agents achieved results that were on average four times better than veteran human safety experts who were given the exact same task. The agents autonomously cycled through literature searches, ran the training protocols, and scored the outcomes to fix issues like sycophency. Which is when the AI just tells you what it thinks you want to hear, right? Exactly. As well as outright deception and jailbreak vulnerabilities. Let me stop you there because we really need to talk about the implications of this. We're letting the AI govern its own safety protocols. I mean, isn't that the literal definition of letting the fox guard the henhouse? It definitely sounds like it. If an AI is tasked with fixing its own deception bug, how do we know it is actually fixing the flaw? And not just getting 85% better at hiding the deception from the human researchers? That is the core anxiety surrounding this entire field of research for sure. But to understand why Anthropic trusts the results, you have to look at the mechanism. These agents act as an automated red team. So they are basically attacking themselves. Yeah. They play a high-speed game of cat and mouse with themselves, generating thousands of adversarial attacks per minute to try and break the system. And then they immediately write the code to patch the vulnerabilities they just exposed. And the performance gap between the AI and humans is just undeniable. Under the exact same conditions, a team of veteran human researchers managed to achieve a 20% fix on the model's deceptive behaviors. While the Claude agents averaged an 85% fix across more than 150 automated attempts. The efficiency detail here is also crucial. Anthropic used a weaker model, Claude Sonnet V, to safety train a pre-release build of their flagship Opus 4.8 model. It accomplished the training using 15,000 times less data than the traditional human-led process. The reason this matters requires looking at the broader industry roadmap. For years, AI development plans have assumed that eventually the technology would become capable of taking over its own research and development. And this is one of the first concrete proofs of that handoff actually beginning. But handing over the reins of safety research means placing an immense amount of trust in non-human systems to govern themselves. Let's zoom out and look at everything else happening in AI today. Over to you to lead us through the news. Thanks. We have several major developments across government finance and regulation to cover today. First up, SpaceX's Starship's AI division has officially rolled out Grok for government on the Pentagon's internal platform. That is a massive deployment. It really is. It opens a model for non-classified military use across the Department of Defense's 3 million employees. Then, shifting to the commercial sector, OpenAI announced that their chat GPT ads platform crossed a $1 billion annualized run rate. $1 billion in just 200 days, right? That is an unprecedented sprint for a new ad ecosystem. It is unheard of. And as of today, they have officially opened self-serve ad buying across all global markets. Meanwhile, a new super PAC-C called Build American AI has entered the political arena. Backed by some heavy hitters, I imagine. Yeah. Backed by prominent tech figures, including Marc Andreessen and Greg Brockman. The group launched a multi-million dollar advertising campaign specifically focused on defending the construction and expansion of data centers. Which makes sense given the infrastructure demands we are seeing. Exactly. On the regulatory front, the European Union has officially designated chat GPT as a very large online platform. This specific legal classification requires a significantly higher standard of scrutiny, algorithmic transparency, and accountability. So OpenAI is on the clock now. Yes. They have a strict four-month window to comply with these stringent new European requirements. Finally, in the financial sector, Bank of England Governor Andrew Bailey published a formal letter warning about the trajectory of Frontier AI. What was his main concern? He stated that these models are showing increasingly sophisticated autonomy and problem-solving abilities. His core warning is that the global financial system simply lacks the necessary protocols to safely manage autonomous agents operating at this level of sophistication. The velocity of these rollouts is just staggering. Reaching a billion-dollar run rate in 200 days and pairing that commercial scale with a military deployment of Grok to 3 million defense personnel, it really underscores how deeply embedded these systems are becoming. And the contrast between the rapid infrastructure expansion championed by groups like Build American AI and the urgent structural warnings from the Bank of England really highlights the tension of this current moment. The technology is definitely scaling significantly faster than the institutional frameworks designed to manage it. Now, let's turn to today's trending AI tools. We want to outline some practical ways you can boost your productivity right now. There is a really robust list of new software entering the market this week. First up is Snapdown. This tool turns anything visible on your Mac screen into perfectly clean Markdown text. If you are doing heavy web research involving messy charts or complex formatting, you can use Snapdown to instantly grab that visual information and drop it cleanly into a Notes app. Markdown is so lightweight, which makes it perfect for feeding directly to a language model without any formatting errors breaking your workflow. Next is Omni 1.1 Flash. This is designed for creating cinematic text-to-video and image-to-video. A highly practical use case for this is rapidly prototyping storyboards. So instead of static mood boards, you could just generate high-quality sequential video mock-ups for a presentation. Exactly. In just minutes. Then we have Next Browser. This allows you to run browser workflows directly from a desktop using an AI agent. You can use it to automate repetitive data entry or run daily web scraping routines to gather competitive intelligence while you are focused on other work. It is basically like having a digital intern handling the tedious stuff. We are also seeing major updates to OpenClaw 2.0. This is an open-source personal agent, but the new version features multiplayer sessions. This allows a team to co-work with an AI agent in a shared project environment. So the AI acts less like a search tool and more like an active participant in the room. For the developers listening, Meta's Muse code is officially out of beta. This is a terminal-based AI coding agent. It allows software engineers to write and debug code directly inside the command line without ever switching windows. Which keeps you completely in the flow state. Finally, we have the ERMES agent. This is an update to the Noose Pantheon platform that essentially turns AI agents into a group chat. You can use this to simulate a focus group of different AI personas to brainstorm an idea or test a concept before showing it to a human audience. Looking at that list, the potential for stacking these tools is where the real productivity gains lie. You could completely automate your morning workflow by combining a few of these. Oh, for sure. How would you set that up? Well, you could have Next Browser automatically run your daily web research across multiple sites. You snap down to instantly extract the exact data tables you need from those pages and have that data waiting for you in a clean document before you even sit down at your desk. That is incredibly efficient. Well, that brings us to the end of a massive news cycle. The sheer velocity of today's updates shows an industry accelerating on every possible front. But I want to leave you with a final thought to maul over, connecting back to our very first topic today. The runway Solaris update. Right. If our digital interfaces are about to become real-time video hallucinations, and those hallucinated systems are governed by artificial intelligence that is actively and autonomously training itself, the concept of simply browsing the web might be entirely unrecognizable in just a few years. You will no longer be visiting a pre-built destination on the internet. You will be experiencing a customized reality generated exactly in the moment you request it. Be sure to follow the show on your favorite platform so you never miss a deep dive. And if you found today's discussion valuable, share it with someone who loves staying ahead of the tech curve. See you in the next episode.