The Deep Feed

01 — Lenny's Newsletter

The Muse Dilemma: Agents vs. Aesthetics

Meta's new frontier models excel at workflows but fail the design test

By Claire Vo · 8 min read

Editor's note: As AI shifts from simple chat to autonomous agents, the gap between functional logic and creative soul is widening.

Meta has released Muse Wonder, a model designed to act as a personal and professional agent. On paper, the capability is impressive. In practice, the results suggest a widening chasm between what an AI can do to manage a task and what it can do to create something worth looking at. When put through the 'How I AI' benchmark—a rigorous test covering coding, prototyping, and writing—Muse shows a distinct personality: it is a tireless clerk, but a mediocre artist. It handles the heavy lifting of agentic workflows with speed, yet its visual outputs, particularly in SVG and short-form video, lack the intentionality required for high-end design.

The Rise of the Agentic Clerk

The real value of Muse lies in its ability to manage long-running tasks. Meta has leaned into the 'active focus mode,' a feature that allows the model to maintain coherence over extended operations. This is where the model shines. If you need an agent to navigate a series of email responses or manage a coding sequence, Muse is a formidable tool. It moves quickly through one-shot tasks and shows promise in the 'agentic' category—tasks that require a sequence of reasoning rather than a single burst of generation. However, this efficiency comes at a cost to the creative edge.

Muse is a master of the workflow, but a failure of the aesthetic.

In the realm of design, the model struggles. While it can generate code and manage logic, its ability to produce visual assets like SVGs or edited video is underwhelming compared to competitors like Claude or specialized video models. It produces work that is technically correct but visually hollow. For an agency owner, this means Muse might be a great tool for managing the process, but it is not a replacement for the designer who understands why a specific curve or colour choice matters.

Where Muse Wins

Long-running agentic tasks
Email management and voice assistant logic
Rapid one-shot coding tasks

The takeaway for businesses is clear: do not mistake efficiency for quality. Meta is building a tool for the 'doing' of work, not the 'dreaming' of it. As these models become more integrated into our professional lives, the premium on human taste and aesthetic judgement will only increase, even as the cost of clerical labour drops to zero.

Key Takeaway

Automation handles the process, but humans must still own the taste.

02 — Not Boring

The World Model Thesis

Why Runway is betting on video to build the future of robotics

By Packy McCormick · 10 min read

Editor's note: Runway is moving beyond video generation into the much larger territory of how machines understand physical reality.

Runway is currently experiencing a period of intense, almost frantic, product release. In a single quarter, they have moved from video generation to real-time interface models, robotics research, and even an operating system preview. To the casual observer, this looks like a company throwing everything at the wall. To those looking closer, it is a singular, coherent bet on 'World Models'. The core idea is that by training on the near-infinite supply of video data, an AI can learn the underlying physics and logic of our world better than it ever could through text alone.

From Pixels to Physics

The acquisition of Kinetix and the release of Praxis-1 signal a shift from digital art to physical control. Praxis-1 is a 'World Action Model' that uses video pre-training to provide a foundation for robotics. Traditional robotics relies on limited, expensive demonstration data. Runway's approach is different: they use the vastness of YouTube and other video sources to teach a model how objects move, how gravity works, and how humans interact with their environment. This creates a policy model that can be adapted to different robotic bodies.

Video data is infinite; robotics data is limited. The winner will be whoever masters the former to solve the latter.

This extends into Solaris, their Interface World Model. Solaris doesn't just generate a static image of a website; it generates an interactive interface frame by frame, in real time. It is a glimpse into a future where the operating system is not a fixed set of buttons and menus, but a fluid, generative environment that responds to your intent. The line between the user interface and the world itself is beginning to blur.

Runway's Strategic Pillars

Video Generation: The foundation of visual understanding
Interface Models: Real-time, generative operating systems
Robotics: Applying visual logic to physical movement

If Runway is right, the next decade of computing won't be about learning how to use software, but about software learning how to perceive the world. We are moving from a period of 'command-line' thinking to a period of 'observation-based' intelligence. The company is not just making tools for filmmakers; they are building the sensory cortex for the next generation of autonomous machines.

Key Takeaway

The path to true robotics lies in teaching machines to watch the world, not just read about it.

03 — Lenny's Newsletter

The Moat Fallacy

How to identify real compounding advantages in the age of AI

By Dan Hockenmaier · 9 min read

Editor's note: In an era where anyone can build a product with a prompt, understanding what makes a business defensible is the most important skill for a founder.

The term 'moat' is being used to describe almost everything in the current startup climate, which makes it nearly meaningless. Speed of execution is often cited, but speed is not a moat; it is a requirement for entry. A company that has to sprint just to stay in place is a fragile business. To find a real moat, one must look for compounding advantages—self-reinforcing flywheels that make a company harder to compete with every single day. If a business doesn't have one of these, it is merely a temporary occupant of a market segment.

The Four Pillars of Defensibility

There are only four true compounding advantages. First, switching costs: when a product becomes so deeply embedded in a workflow that leaving it causes significant pain. Second, network effects: where every new user increases the value for all existing users. Third, scale economies: where increased volume leads to lower unit costs, creating a price or quality advantage that competitors cannot match. Fourth, data effects: where more usage generates more data, which in turn improves the product, attracting more users.

Speed is a shot on goal, not a way to win the game.

Many founders mistake distribution or brand for moats. While vital, they often fail to compound. Distribution helps you reach customers, but it doesn't stop a competitor from reaching them too. Brand can drive acquisition, but in a world of hyper-efficient AI competitors, a brand can be eroded if the product experience itself is commoditised. A real moat must be structural, not just perceptual.

The Moat Checklist

Does the product get better as more people use it? (Network Effects)
Does it become harder to leave the more it is used? (Switching Costs)
Does more volume lead to better margins? (Scale Economies)
Does usage create a proprietary data loop? (Data Effects)

For the agency owner or investor, this framework is a filter for longevity. In the AI era, the barrier to building a product has vanished. This means the competition will be relentless. If you cannot identify which of these four engines is driving a startup, you are likely looking at a company that will eventually be crushed by the very tools that helped it launch.

Key Takeaway

A moat is not a temporary edge; it is a structural advantage that compounds over time.

04 — Aeon

The Deskilling of the Body

How our reliance on digital tools is eroding human biological capacity

By Melanie Challenger · 12 min read

Editor's note: We are trading our sensory and spatial fluencies for the convenience of algorithms.

There is a quiet erosion happening in the way we inhabit our own bodies. As we outsource more of our cognitive and sensory tasks to digital tools, we are not just saving time; we are losing capacities that took millions of years of evolution to hone. We rely on GPS to navigate, and in doing so, we weaken our spatial attention. We rely on algorithms to suggest what to read or watch, and in doing so, we lose the capacity for serendipitous exploration. We are becoming passive recipients of information rather than active participants in our environment.

The Machine Lie

This deskilling is driven by a long-standing philosophical error: the idea that the human body is merely hardware for the mind. From Descartes to the cyberneticists of the 1950s, we have been taught to view the body as a machine that can be upgraded or bypassed. If the body is just a vessel for information, then automation is seen as a pure gain. But the body is not a machine; it is a living system that requires engagement to maintain its function. When we substitute a technological fix for a physical action, we are not just being efficient; we are undergoing a form of atrophy.

We are trading the effort of exploration for the ease of recommendation.

The distinction between helpful technology and deskilling technology lies in whether the tool strengthens or replaces our capacities. A tool that helps us see further is an extension of our senses. A tool that tells us where to look, and what to think about it, is a replacement. Algorithms that regurgitate patterns do not just save us effort; they disrupt the very curiosity and play from which human preference and identity are formed.

Signs of Biological Deskilling

Weakened spatial wayfinding due to GPS reliance
Reduced attention spans from algorithmic feedback loops
Loss of sensory engagement through screen-mediated life

The cost of this convenience is a body that is less able to assess its own condition, improvise, or regulate itself. As we move further into the AI era, the temptation to automate away every friction point will grow. But friction is often where learning happens. If we remove all the resistance from our lives, we may find that we have also removed the very things that make us capable, sentient animals.

Key Takeaway

Convenience is often a euphemism for the slow erosion of human capability.

05 — Experimental History

The Finite Desires Fallacy

Why the 'lump of labor' is a myth and why we will never run out of work

By Adam Mastroianni · 8 min read

Editor's note: AI won't run out of things for us to do; it will simply reveal how infinite our appetites truly are.

The fear that AI will lead to mass unemployment is built on a logical error known as the 'lump of labor' fallacy. This is the mistaken belief that there is a fixed amount of work to be done in an economy, and that as machines take over tasks, the available pool of jobs must shrink. If this were true, we would have reached 100% unemployment long ago. Instead, history shows that as technology automates old tasks, humans do not sit idle; we invent entirely new categories of work that were previously unimaginable.

The Mouse and the Cookie

To understand why we fail to see this, we can look to the children's book *If You Give a Mouse a Cookie*. The book describes a cycle of endless, conditional demands: a cookie leads to milk, which leads to a straw, which leads to a napkin, and so on. The boy is trapped because he cannot predict the mouse's next desire until the current one is satisfied. Human desire works exactly the same way. We cannot imagine the jobs of the future because those jobs are conditional on the technologies of today. We cannot imagine being a 'social media manager' or a 'cloud architect' until the underlying infrastructure exists to make those roles necessary.

The lump of labor fallacy is actually the finite desires fallacy.

We often mock new, seemingly frivolous professions as evidence of economic decadence. But these jobs are not signs of a dying economy; they are signs of a growing one. Every job, from the most essential to the most seemingly absurd, exists because humans have the capacity to upgrade their 'nice-to-haves' into 'must-haves'. As AI satisfies our current needs, our appetites will simply shift toward new, more complex desires.

Why Work Evolves

Automation of repetitive tasks creates capacity for higher-order tasks
New technologies create new needs (e.g., the internet created SEO)
Human desire is conditional and infinite

The real challenge of the AI era is not a lack of work, but a lack of preparation. We are not facing a shortage of tasks, but a rapid shift in the *nature* of those tasks. The economy will not run out of things to do; it will simply become a more complex reflection of our ever-expanding, ever-changing list of demands.

Key Takeaway

We don't run out of work; we just run out of old ways to be busy.

06 — Aeon

The Cost of Memory

Tracing the legacy of resistance in the face of genocide

By Aeon · 6 min read

Editor's note: A look at the quiet, perilous work of rebuilding a community from the literal ground up.

In the wake of the ISIS genocide against the Yazidi people in Sinjar, the battle for survival has moved from the front lines to the soil itself. The land is littered with the remnants of conflict—landmines and unexploded ordnance that turn everyday life into a lethal gamble. For Hana Khider and her all-female de-mining team, the work is a meticulous, high-stakes struggle to reclaim the physical space necessary for a community to breathe again. It is a task that requires absolute precision and a total lack of hesitation.

The Geometry of Danger

The danger is often disguised. Explosives are hidden in ordinary objects, sometimes even children's toys, turning the symbols of innocence into instruments of death. The work of the Mines Advisory Group is not just about technical expertise; it is about the psychological endurance required to face these objects daily. Khider’s team operates in the tension between the professional necessity of their task and the personal reality of being a wife and mother in a community still reeling from trauma.

Rebuilding a community begins with making the ground beneath it safe to stand on.

The film *Into the Fire* captures this duality. It shows that de-mining is not just a technical process of disarming explosives, but a social process of restoring agency. When a field is cleared, a farmer can plant; when a path is cleared, a child can walk. The work is slow, incremental, and often invisible, yet it is the fundamental prerequisite for any meaningful recovery or peace.

The Reality of Post-Conflict Recovery

Physical safety is the foundation of economic stability
The psychological toll of hidden threats remains long after combat ends
Community rebuilding requires reclaiming the physical landscape

This is the quietest form of resistance: the refusal to let the tools of war dictate the future of the living. By clearing the land, Khider and her team are not just removing explosives; they are removing the lingering grip of the genocide, one square metre at a time.

Key Takeaway

True recovery is not found in grand gestures, but in the painstaking reclamation of safe ground.