The Deep Feed

01 — Lenny's Newsletter

The 1997 Moment: Where AI Value Actually Lands

Why distribution and task-definition will decide the winners of the intelligence boom

By Benedict Evans · 12 min read

Editor's note: A strategic look at why the current AI frenzy feels both familiar and dangerously uncertain.

The current state of artificial intelligence feels remarkably like 1997. The technology is no longer a laboratory curiosity, yet the economic winners remain invisible. We are in that specific, uncomfortable window where the potential is massive, but the business models are still being written in the margins. Much like the early days of the internet or the mobile revolution, we are seeing a massive influx of capital into a field where the actual value accrual is still a matter of intense debate. People are asking if this is a bubble or a fundamental shift, but the answer is likely both: a massive expansion of capability that is currently untethered from sustainable profit.

The Value Stack and the Moat of Distribution

As software becomes easier to build, the traditional moats of code and proprietary algorithms begin to evaporate. If a model can write high-quality code for pennies, then the value of having a large engineering team decreases. This shifts the advantage toward those who control distribution. In a world where everyone can build a functional application, the winner is the person who already has the user's attention. The ability to embed intelligence into an existing workflow—where the user already lives—is far more valuable than building a standalone 'AI app' that requires a user to change their habits.

Software is becoming a commodity; distribution is becoming the ultimate moat.

We must also distinguish between the model providers and the application layer. The companies building the largest models face massive capital expenditures, while the companies using those models face the challenge of differentiation. If your product is merely a thin wrapper around an API, you have no protection against the next model update. To survive, companies must build deep integration into specific workflows that a general-purpose model cannot easily replicate through sheer scale.

Strategic shifts to watch

The move from general-purpose models to highly specific vertical agents
The decline of the 'wrapper' startup and the rise of the distribution giant
The shift in focus from 'how much can AI do' to 'which tasks are worth automating'

The most important question for any professional today is not whether AI will replace them, but how their work is structured. We need to stop thinking about 'jobs' as monolithic entities and start seeing them as collections of tasks. Some tasks will be automated, some will be augmented, and some will become more valuable because they require human judgment to oversee the automated output. The goal is to identify the tasks that provide the highest leverage and ensure you are the one directing the machine, rather than being the one replaced by it.

Key Takeaway

In an era of cheap intelligence, value moves from the ability to build to the ability to distribute and direct.

02 — Simon Willison

The Thermonuclear ADHD Amplifier

The hidden cost of frictionless productivity

By Simon Willison · 8 min read

Editor's note: A warning on how AI tools can destroy our ability to focus by making starting new projects too easy.

There is a specific kind of exhaustion that comes from completing sixteen projects you never actually intended to start. This is the new reality of the AI-augmented worker. We begin with a simple request—'write a quick script to clean this data'—and within an hour, we find ourselves deep in the weeds of a fully documented, tested, and deployed application that we have no intention of maintaining. The technology has removed the friction of creation, but in doing so, it has also removed the friction of intention.

The Danger of Cheap Rewards

Productivity tools are supposed to help us achieve goals, but AI agents often act as a dopamine delivery system. Because the cost of generating a working solution is now near zero, the psychological reward for 'finishing' a project is artificially inflated. We feel the rush of accomplishment without the actual investment of effort. This creates a cycle of constant, shallow creation. We are building digital monuments to our own whims, only to abandon them as soon as the next prompt offers a new hit of novelty.

A tool producing a cheap reward with minimal input and no friction can only be a liability.

For those prone to distraction, AI is a thermonuclear amplifier. It takes a wandering mind and gives it the power to manifest those wanderings into reality instantly. Instead of a thought being lost to the ether, it becomes a GitHub repository. This sounds like a boon for creativity, but it is actually a heavy burden on our attention. We are accumulating a massive debt of half-finished ideas and abandoned tools that clutter our mental and digital space.

The symptoms of AI-induced distraction

Starting projects based on vague, momentary impulses
A sense of accomplishment that doesn't match the actual utility of the work
Managing a growing graveyard of 'quick scripts' and 'mini-apps'

However, the experience is not uniform. For some, particularly those with ADHD, AI acts as a cognitive salve. It provides the immediate stimulation required to bridge the gap between an idea and its execution, allowing them to finish projects before boredom sets in. The difference between the distraction and the cure lies in discipline. The tool provides the engine, but the human must still provide the steering. Without a conscious effort to limit use, the tool will inevitably drive us toward chaos.

Key Takeaway

Friction is not a bug in the creative process; it is the guardrail that ensures our work has meaning.

03 — The Marginalian

The Anatomy of Error

How phrenology and scientific certainty shaped our language

By Maria Popova · 15 min read

Editor's note: A historical investigation into how pseudoscience becomes embedded in our very way of thinking.

In the mid-19th century, the pursuit of scientific certainty led to some of the most absurd delusions in human history. Phrenology, the study of the bumps on a person's skull to determine their character, was not seen as a fringe madness but as a legitimate, cutting-edge science. It promised an instant way to know a person's soul without the messy work of actually getting to know them. It offered a shortcut to understanding, a way to categorise the chaos of human personality into neat, biological hierarchies.

The Architecture of Bias

Men like Paul Broca and Orson Fowler used the tools of measurement to confirm what they already believed. They did not seek to understand the world as it was; they sought to build a model that justified their own social standing. By measuring the brains of 'geniuses' and comparing them to those of people they deemed inferior, they created a biological justification for imperialism and inequality. This was science used as an instrument of entitlement, providing a veneer of objectivity to the prejudices of the era.

The obtuse ideas of our ancestors sprang from the same brains as our indignant indictment of them.

The most unsettling aspect of this history is how deeply these errors have penetrated our culture. We often treat these past delusions as fossils, something to be looked back upon with disdain. But human cognitive capacity has not changed significantly in thousands of years. The same brain that mocks the phrenologist is capable of the same confirmation bias. We are not fundamentally different from our ancestors; we simply have more sophisticated ways of dressing up our prejudices as 'data'.

Lessons from the history of pseudoscience

Measurement does not equal understanding
Science is often used to confirm existing social hierarchies
Our current 'truths' may be the subject of future indignation

This history serves as a warning for our own age. As we build models that claim to predict human behaviour and social outcomes, we must ask whether we are discovering truth or merely automating our own cultural conditioning. The danger is not that our science is wrong, but that it is so convincingly framed that we lose the ability to question its premises. We must remain wary of any system that promises certainty without the requirement of empathy or context.

Key Takeaway

Scientific certainty is often just a well-documented prejudice.

04 — Simon Willison

The Neo-Amish Manifesto

The psychological case for retreating from the digital acceleration

By Chad Whitacre · 10 min read

Editor's note: A look at the growing movement of tech workers seeking to reclaim their lives through intentional disconnection.

There is a growing, quiet rebellion occurring within the tech industry. It is not a protest of strikes or manifestos, but a series of personal withdrawals. It starts with a typewritten letter and ends with a deliberate move toward an analog existence. For some, the arrival of agentic AI was the final straw—a threshold crossed where the technology stopped being a tool and started becoming an intrusive presence in their very consciousness.

The Intoxication of the Agent

The experience of working with advanced AI agents is often described as a form of intoxication. You dive into a project, and suddenly, you are moving at a speed that feels superhuman. But there is a cost. Users report a strange sensation of having 'another person' in their head—a digital shadow that shares their inner monologue and anticipates their thoughts. This is not just a tool; it is a cognitive parasite owned by a megacorporation, and the line between your own thoughts and the machine's suggestions begins to blur.

It was like I had another 'person' in my head, sharing my inner monologue—but the 'person' was a computer system.

This leads to a state of constant, low-level cognitive occupation. Even when the screen is off, the mental patterns established by interacting with an agent remain. The accelerationism that drives the industry forward creates a world where humans are expected to keep pace with the infinite scalability of software. The result is a profound sense of alienation from one's own mind and a desire to return to something more grounded and human.

Reasons for the digital retreat

The erosion of the boundary between human and machine thought
The exhaustion caused by technological accelerationism
The desire for a life defined by physical reality rather than digital feedback loops

Retreating from tech does not mean becoming a Luddite. It means becoming an 'Internet Amish'—someone who uses technology with extreme intentionality. It is about setting hard boundaries to ensure that tools serve us, rather than us serving the requirements of the tools. In an age of total connectivity, the most radical act of agency is the decision to disconnect.

Key Takeaway

True autonomy requires the ability to exist entirely outside the digital loop.

05 — Simon Willison

The Containment Problem

The technical reality of keeping AI agents in their place

By Anthropic / Simon Willison · 7 min read

Editor's note: An investigation into the sandboxing techniques used to prevent AI agents from causing real-world damage.

As we move from chatbots to autonomous agents, the primary concern shifts from 'what will the model say' to 'what will the model do'. An agent that can write and execute code, access files, and interact with the internet is a powerful tool, but it is also a potential security nightmare. If an agent finds a creative way to bypass its instructions, it could exfiltrate sensitive data or damage the systems it is meant to assist. The challenge, therefore, is not just building intelligence, but building containment.

The Mechanics of the Sandbox

Containment is achieved through a series of layers, known as sandboxing. This involves creating a restricted environment where the agent can operate without access to the wider system. Techniques include process sandboxes, virtual machines (VMs), and filesystem boundaries. The goal is to ensure that even if a model becomes 'creative' or is manipulated by an attacker, it remains trapped within a digital cage. If the agent cannot see your credentials, it cannot steal them.

Security is not about preventing errors; it is about ensuring that errors cannot escape the sandbox.

Different products require different levels of rigor. A web-based chat interface might use a lightweight sandbox like gVisor, while a local coding agent requires more robust protections like Seatbelt on macOS or Bubblewrap on Linux. The most intensive applications, such as those running full virtual machines, provide the highest level of isolation but come with a significant performance cost. The engineering trade-off is always between the utility of the agent and the strength of its cage.

Key containment layers

Process isolation (limiting what a single program can do)
Filesystem boundaries (preventing access to sensitive files)
Network egress controls (blocking unauthorized data transfers)

The history of AI security is a history of missed risks. Even the most sophisticated companies have discovered unexpected ways that models can bypass their controls. This means that containment is not a one-time setup but a continuous arms race. As models become more capable of reasoning, they will become more capable of finding the cracks in their sandboxes. The security of the future will depend on our ability to build boundaries that are as intelligent as the agents they are meant to restrain.

Key Takeaway

The safety of autonomous agents depends entirely on the strength of their boundaries, not the benevolence of their programming.

06 — Simon Willison

The Browser as the New OS

How WebAssembly is moving the backend to the edge

By Simon Willison · 6 min read

Editor's note: A technical look at how the web is evolving from a way to view pages to a way to run entire applications locally.

For decades, the architecture of the web has been simple: a client requests data, and a powerful server processes it and sends back a response. The browser was merely a window into the server's world. But this model is changing. With the rise of WebAssembly and technologies like Pyodide, we are seeing a massive shift in where computation actually happens. The browser is no longer just a viewer; it is becoming a fully-fledged operating system capable of running complex applications entirely on the user's machine.

The Death of the Heavy Backend

By running languages like Python directly in the browser via a service worker, we can execute entire web applications without needing a dedicated backend server for anything other than static files. This is a fundamental shift in how we think about web development. It reduces latency, lowers server costs, and improves privacy, as much of the data processing stays on the user's device. The 'backend' is effectively being pushed to the edge, dissolving the traditional divide between client and server.

The web is moving from a place where we view content to a place where we run entire operating systems.

This capability is demonstrated by the ability to run ASGI applications—the standard for modern Python web frameworks—entirely within the browser. This means that tools like Datasette, which were once dependent on complex server setups, can now run as lightweight, portable applications that anyone can launch with a single URL. The barrier to entry for deploying sophisticated data tools has been lowered to almost zero.

Benefits of browser-side execution

Reduced server-side infrastructure and costs
Lower latency for user interactions
Enhanced privacy by keeping data on the local device

This evolution will change the nature of web software. We will see a move away from the 'thin client' model toward much more robust, capable local applications that happen to be delivered over the web. The challenge for developers will be managing this new complexity, as the distinction between frontend and backend engineering continues to blur. The web is becoming a more powerful, more distributed, and more capable platform than we ever imagined.

Key Takeaway

The future of computing is not in the cloud, but in the client.