Edition #8. Zuckerberg is building an AI agent that supports his role as CEO. OpenAI pulls the plug on Sora.
Zuckerberg builds a personal AI CEO agent
Mark Zuckerberg runs a company with 78,000 employees. And he wants an AI agent to help him with that. According to the Wall Street Journal, Meta's CEO is working on a personal AI agent that supports him with strategic decisions, processing internal information and managing his agenda. Not as an experiment. As a daily working tool.
That sounds like a fun company-wide pilot project, until you hear that AI usage is by now a factor in performance reviews at Meta. You're not just encouraged to use it. You're judged on it.
The direction is clear: Meta sees AI agents not as a tool alongside your work, but as an integral part of how you work. When 78,000 agents talk to each other's agents, you're essentially building a distributed decision-making system on top of your organization.
The question is no longer whether AI agents will enter the workplace. The question is how quickly your organization falls behind if you don't.
OpenAI shuts down Sora
Sora, OpenAI's video generation model, is being shut down completely. The app, the API and Sora.com are all closing.
It's a striking choice. Sora was one of the most-discussed AI products of the past two years. The waitlist was huge. The demos were impressive. And now OpenAI is pulling the plug.
The reason is strategic: OpenAI is shifting compute toward world simulation for robotics. Instead of generating videos that look realistic, they want to build models that understand the physical world. How objects move. How gravity works. How a robot arm picks up a cup without breaking it.
For users and companies who built on Sora: time to find alternatives. Runway, Kling and Veo are ready to absorb Sora's customers.
Google compresses AI memory 6x, without quality loss
Google Research published TurboQuant this week, a compression algorithm that shrinks the working memory of AI models by a factor of six, without quality loss.
The technical idea: large language models store intermediate results in temporary memory during a conversation. That memory grows with how long a conversation or document is, and quickly becomes a bottleneck on complex tasks. TurboQuant compresses that memory drastically, without the model needing to be retrained.
What it does mean: powerful AI models will soon just run on your phone. Within 24 hours after publication, developers were already busy implementing the algorithm in popular tools for local AI.
