More from this creator
Other episodes by Kitty Cat.
More like this
If you liked this, try these.
Transcript
The full episode, in writing.
What’s actually changing in AI right now? As of October 29, four developments stand out—not because they all put a new chatbot in your hands, but because they could change how software is secured, how AI learns about the world, how countries prepare for biological threats, and who can afford to run these systems. Let’s start with the most urgent.
On October 27, Anthropic described using an unreleased, internal research model to find vulnerabilities in browsers and operating systems, then build working exploit chains. In one demonstration, its team says Claude combined multiple Chrome flaws to escape the browser’s protections and run code on a computer. The researchers also describe findings involving Safari, Android, Windows and Linux. Anthropic says it reported the vulnerabilities to the affected vendors; its account identifies fixes for several of the examples. This is Anthropic’s account of its own security research—not evidence that the model has been released, or that these attacks are circulating in the wild. The significance is that finding a bug and combining bugs into a working attack may no longer require the same amount of specialist human effort. That could give defenders a powerful way to find and patch weaknesses. It could also raise the stakes if similar capabilities become widely available. Anthropic’s practical advice to defenders is to scan their own code.
Next, imagine walking through a scene that AI creates while you’re inside it. On October 27, Google DeepMind announced Genie 4, a world model that generates an interactive environment in real time, including sound and responsive dialogue. According to Google, a person can steer a scene with text—changing the weather, for example—and the simulation responds as events unfold. The interesting possibility goes beyond games or filmmaking: varied simulated environments might eventually help researchers test agents or train robots before putting them in the physical world. But that is a possibility, not a demonstrated general-purpose solution. Google calls Genie 4 an early experimental model. It says objects and visual details can drift during longer sessions, while physical interactions and spatial accuracy still need work. Access began as a research preview for trusted testers; Google described broader access as a future step, not something everyone could use on announcement day.
The third development asks what happens when AI’s growing biological capabilities meet public-health preparedness. On October 24, the UK government announced that it was establishing an AI Biosecurity Lab. Its initial focus is earlier detection of emerging biological threats and faster development of vaccines, treatments and other countermeasures. The government says it is backing the lab with £100 million in public funding. It also describes up to £200 million in commitments from OpenAI and Anthropic, including compute, model access and technical support—not simply another £200 million in government cash. The ambition matters because AI could help with disease response while also lowering barriers to harmful activity. For now, this is a funded initiative with stated goals, not proof that it has already detected an outbreak or delivered a vaccine.
Finally, a quieter infrastructure story could shape what gets built next. On October 20, Google announced that TorchTPU, its native PyTorch stack for Google’s AI chips, was available as open source. PyTorch is a widely used toolkit for developing AI. Google says TorchTPU, co-developed with Meta’s PyTorch team, lets developers move existing PyTorch workloads onto TPUs with minimal code changes and use familiar tools for training and serving models. Why care if you don’t run a data center? Making software easier to move between kinds of chips can give AI builders more hardware choices and reduce the work of adapting a model to a particular platform. It does not, by itself, establish that TPUs are cheaper or faster for every task.
Put these stories together and a pattern emerges. AI progress is not just about a more capable model. It is also about whether that capability can be tested safely, defended against when necessary, applied to difficult real-world problems, and run on accessible infrastructure. The striking results are worth watching. So is the distance between a research demonstration, an announced plan, and a benefit people can actually rely on.