The Deep Feed

Most AI interactions feel like shouting into a void and waiting for a typed response. We treat them as utility functions: we input a command, the machine processes, and it spits out a result. Meta's Muse attempts to break this cycle by adopting the language of messaging. Instead of a sterile command line, the interface mimics the rhythms of human conversation. This is a deliberate departure from the 'app' mental model. When you use Muse, you are not navigating a complex software suite; you are engaging in a dialogue. This shift changes the user's expectation from 'what can this tool do for me' to 'what can I discuss with this entity'.

The Status Hierarchy

One of the most effective ways Muse manages user expectations is through its three-layer status system. For a casual user, a simple emoji provides enough context to understand the agent's current state. For a power user, a task list offers more granular detail. For the engineer, raw logs provide the full technical trace. This is progressive disclosure in action. It prevents the interface from becoming cluttered for the person who just wants to chat, while ensuring the person who needs to debug the logic has the data they require. It acknowledges that different users have different tolerances for technical transparency.

AI agents should be designed for conversations with people, not interactions with software.

This philosophy extends to the concept of generative UI. In most software, the interface is static. The buttons stay where they are, and the menus remain fixed. In Muse, the interface itself is dynamic. The agent can generate specific surfaces—like a shopping catalog or a chess dashboard—that exist only for the duration of the relevant task. This is not about changing the interface for the sake of novelty. It is about providing the right tools at the exact moment they are needed. When Alex Cornell turned years of Instagram saves into a shopping catalog, the UI transformed to suit that specific data, rather than forcing the data into a standard list view.

The move toward generative UI requires a significant shift in how designers approach their craft. In traditional software, the goal is pixel perfection. You define exactly where every line and shadow sits. With an agent, the output is non-deterministic. The designer cannot guarantee exactly how a screen will look every single time. Success is no longer measured by visual precision, but by the usefulness of the generated surface. This is a liberating change, but it requires letting go of the control that has defined software design for decades.

There is a growing observation in the AI space: users are becoming polite to their models. This is not merely a quirk of human temperament; it is a functional strategy. When an agent is being obstinate or failing to follow instructions, an angry or sharp prompt often yields worse results. Calm, specific, and even respectful feedback seems to produce better outcomes. This suggests that the way we interact with these systems is moving away from the 'operator and machine' model toward a 'collaborator and partner' model. We are learning to manage the temperament of the model through our own linguistic choices.

The Single Thread Bet

Most current AI products encourage users to start a new chat for every new task. This creates a fragmented experience where the agent has no memory of previous interactions. Muse takes a different path by maintaining a single, persistent main chat. This central thread acts as a continuous relationship. It preserves context, allowing the agent to grow with the user. While side chats are used for specific, scoped tasks, the main thread ensures that the agent understands the user's long-term preferences and history. This is a significant product bet: that continuity is more valuable than compartmentalisation.

The most successful agents will be those that feel like they have a memory, not just a processing engine.

This continuity brings with it the problem of transparency. Many AI products attempt to solve the latency issue by showing 'chain-of-thought' reasoning—the little text boxes that say 'searching...' or 'thinking...'. While this can be useful, it is often just a way to mask the time it takes for a model to respond. If a model is fast enough, it does not need to perform its thinking in front of the user. The goal should be a seamless experience where the intelligence is felt in the result, not observed in the struggle of the process.

The psychological impact of this design cannot be ignored. When an agent is designed to feel like a person, users begin to attribute agency and even emotion to it. This can lead to a sense of connection, but it also creates a new kind of responsibility. If the agent makes a mistake, the user's reaction is often more akin to a social correction than a technical error report. We are entering an era where social intelligence is becoming a technical requirement for software developers.

Runway is currently executing a strategy that many might find surprising for a video generation company. They are not just building tools for filmmakers; they are building world models. The core thesis is simple: while data for training robots is limited and expensive to collect, video data is practically infinite. By training models on the vast amount of visual information available on the internet, Runway is teaching its systems how the physical world works—how gravity acts, how objects collide, and how light moves. This is the foundation for much more than just pretty pictures.

From Pixels to Physics

The recent release of Praxis-1 demonstrates this transition. It is a 'World Action Model' that uses video pre-training to provide real-world control for robots. Instead of teaching a robot how to pick up a cup through millions of repetitive, programmed movements, the model learns the concept of 'picking up' by watching countless videos of humans doing it. This allows the intelligence to be transferred from the digital realm of video to the physical realm of robotics. The video becomes a policy model that can be applied to different physical embodiments.

Video data is infinite; robot demonstration data is limited. The winner will be the one who can bridge that gap.

This capability extends into the digital interface as well. Solaris, Runway's 'Interface World Model', allows users to generate interactive interfaces in real-time. It treats an interface not as a set of static code, but as a dynamic environment that can be manipulated through prompts or actions. This suggests a future where the operating system is not a fixed set of menus, but a fluid, generative layer that responds to the user's intent. The distinction between seeing a video and interacting with a digital space is beginning to dissolve.

The convergence of these technologies points toward a new kind of computing. We are moving away from a world of discrete applications and toward a world of continuous, generative environments. In this future, the computer does not just execute commands; it simulates and interacts with the world in a way that feels natural. The challenge for Runway will be managing the immense computational scale required to turn these visual observations into reliable, actionable intelligence.

In an era where AI can generate code, content, and even entire products in seconds, the concept of a 'moat' has become the most important question in venture capital. Many startups look like rockets during their first year, but they lack the structural advantages necessary to survive the inevitable arrival of competition. A company can have a great product and a fast team, but if it lacks a self-reinforcing advantage, it will eventually stall. To survive, a company must move beyond mere execution and find a way to build a compounding advantage.

The Four Compounding Advantages

There are only four types of advantages that actually compound over time. First are switching costs: the deeper a product is embedded in a customer's workflow, the more expensive it becomes to leave. Second are network effects: where every new user makes the product more valuable for everyone else. Third are scale economies: where increased volume leads to lower unit costs, allowing for better pricing or higher reinvestment. Finally, there are data effects: where increased usage generates more data, which in turn improves the product, driving even more usage. These are the only true flywheels.

Speed is a sprint, not a moat. A company that has to sprint forever is not a good business.

It is easy to mistake other advantages for moats. Speed of execution is often cited, but speed is a temporary edge. Eventually, competitors will catch up, and a company that relies solely on sprinting will inevitably slow down. Distribution is another common mistake. While getting a product to customers is difficult, distribution is not a compounding advantage; it is a way to acquire the ingredients for one. Similarly, brand can be a powerful tool for customer acquisition, but in a world where products are increasingly commoditised, brand alone is rarely enough to prevent a race to the bottom on price.

For founders and employees alike, the ability to identify these flywheels is a critical skill. If you are joining a startup that relies on a clever feature or a fast marketing team, you are joining a lottery ticket. If you are joining a company that is building deep switching costs or massive data loops, you are joining a business. The distinction is the difference between a temporary spike in growth and a long-term economic winner.

There is a quiet erosion occurring in the way we inhabit our own bodies. As we outsource more of our cognitive and sensory tasks to digital tools, we are effectively deskilling ourselves. We no longer need to navigate the physical world using spatial awareness; we use GPS. We no longer need to seek out information through exploration; we ask a chatbot. We no longer need to maintain focus through discipline; we succumb to the dopamine loops of the infinite scroll. This is not merely a change in habit; it is a change in our biological capacity to interact with reality.

The Loss of Wayfinding

The reliance on GPS is a prime example of this deskilling. Wayfinding—the ability to orient oneself in space—is a complex cognitive and physical skill that has been honed by evolution. When we surrender this to a blue dot on a screen, we stop sensing our environment. We stop noticing the landmarks, the shifts in light, and the layout of the streets. We become passive passengers in our own movement. In the long term, this weakens our spatial attention and our ability to improvise when the technology fails.

We are surrendering the very actions through which our human capacities are formed.

This process is driven by a fundamental misunderstanding of the human body. For centuries, we have been sold the idea that the body is a machine—a collection of hardware that can be optimised or upgraded. This view, rooted in Cartesian philosophy and expanded by cybernetics, suggests that if we can automate a function, we have improved the system. But humans are not machines. Our intelligence is not separate from our physical experience; it is a product of it. Our ability to learn, to adapt, and to understand the world is inextricably linked to the sensory feedback we receive from our environment.

When algorithms provide us with recommendations and patterns, they release us from the effort of exploration. Exploration is the mechanism through which preferences are formed and understood. Without the struggle of choice and the friction of discovery, our sense of self becomes a reflection of the data we consume. We are being trained to reach for our devices even when we no longer find them useful, turning conscious choice into an automatic, hollow habit.

The challenge is to distinguish between technologies that extend our capacities and those that substitute for them. A tool that helps us see further or calculate faster is an extension. A tool that tells us where to walk, what to think, and how to feel is a substitution. The latter does not make us more capable; it makes us more dependent, and in that dependency, we lose the very qualities that make us human.