The Deep Feed

01 — Cal Newport

The Productivity Mirage

Why new tools often create more work than they save

By Cal Newport · 8 min read

Editor's note: A warning against the assumption that AI will automatically result in more output.

The history of technology is a history of broken promises regarding efficiency. When the personal computer arrived, the assumption was immediate and universal: software would strip away the drudgery of office life. We expected spreadsheets to replace manual ledgers and word processors to eliminate the slow pace of typewriters. Instead, we entered an era of digital noise. The time saved by a faster way to calculate a budget was immediately consumed by the time spent managing a bloated inbox or navigating a dozen different software interfaces. The computer did not make us more productive; it simply changed the nature of our distractions.

The 1990s Stagnation

Economic data from the late twentieth century provides a sobering reality check. Between 1987 and 1993, despite massive capital investment in computing, business output growth saw almost no meaningful boost from these machines. One economist noted that white-collar productivity had effectively hit a wall. The technology was present, but the way humans worked had not yet adapted to it. We were using high-speed machines to perform low-speed habits. This gap between the capability of a tool and the output of the person using it is where productivity goes to die. It is a period of friction where the tool is a burden rather than a lever.

In the digital world, productivity doesn't always match our expectations.

We are currently entering a similar period with artificial intelligence. The ease with which a large language model can draft a memo or write a script is undeniable. However, we must distinguish between ease and efficiency. Making a task easier is not the same as making a person more productive. If an AI allows an employee to write ten times more emails, but the volume of incoming communication also increases tenfold, the net gain is zero. We risk creating a feedback loop where the speed of production merely accelerates the speed of chaos.

The Drivers of the Paradox

The expansion of task volume to fill saved time
The cognitive load of managing new toolsets
The mismatch between tool speed and human decision-making

To avoid this trap, we cannot simply throw more compute at the problem. The steam engine and the power loom changed the physical world because they replaced muscle. AI attempts to replace cognition, which is a much more volatile substrate. Muscle is predictable; thought is not. If we want to see a real return on AI investment, we must redesign our workflows rather than just automating our existing, broken ones. We need to ask not how much more we can do, but what we should stop doing entirely.

Key Takeaway

Efficiency is not the same as output; automating a bad process only makes it fail faster.

02 — Lenny's Newsletter

The Open-Weight Rebellion

Breaking the vendor lock-in of the frontier model era

By Lenny Rachitsky · 9 min read

Editor's note: An analysis of why open-weight models like GLM-5.2 are becoming essential for enterprise control.

For the past two years, the AI industry has been dominated by a handful of closed-door labs. If you wanted the best intelligence, you had to rent it via an API, subject to their pricing whims, their censorship filters, and their sudden policy shifts. This is a precarious way to build a business. The emergence of high-performing open-weight models, such as the GLM-5.2 from Z.ai, represents a shift in the power balance. These are no longer toys for hobbyists; they are production-grade alternatives that allow companies to reclaim control over their own technical stacks.

The Math of Sovereignty

The argument for open-weight models is often framed as a matter of ethics, but for a CTO, it is a matter of unit economics and risk management. When you use a closed model, you are a tenant. When you use an open-weight model, you are an owner. You can run inference on your own hardware, fine-tune the weights on your proprietary data, and switch providers without rewriting your entire application. This decoupling of intelligence from the provider is the only way to build long-term stability in a market that changes every three months.

The decision is no longer about capability ceilings, but about cost, control, and vendor dependency.

Testing GLM-5.2 in real-world coding environments reveals its surprising strength. In autonomous tasks, such as hunting through logs to identify bug patterns, it performs near the level of much larger, closed models. While it may struggle with the high-level abstractions of complex frameworks like React under intense pressure, its ability to handle long-running agentic tasks at a fraction of the cost is significant. For a team running millions of tokens through an agent, the cost difference is not just a marginal saving; it is the difference between a viable product and a burning pile of cash.

Advantages of Open-Weight Models

Zero vendor lock-in
Data privacy through local inference
Predictable cost curves for long-context tasks

We are moving toward a hybrid reality. The frontier models will still exist as the ultimate benchmarks, but the day-to-day heavy lifting of engineering and automation will likely migrate to these leaner, more controllable models. The goal is to put these tools in rotation, using them for specific, high-volume tasks where the cost of a closed API would be prohibitive. The winners will be those who build systems that can swap brains as easily as they swap servers.

Key Takeaway

Control over your model is as important as the model's intelligence.

03 — Lenny's Newsletter

The Death of the Process

How Gusto built a product line in ten weeks without a PM

By Claire Vo · 10 min read

Editor's note: A case study in radical speed through the removal of traditional software coordination.

In most software companies, the path from idea to launch is paved with documentation. You have Product Managers writing specs, designers creating Figma mocks, and engineers managing Jira tickets. This structure is designed to prevent errors, but it has a secondary effect: it creates massive coordination overhead. For a small, high-velocity team, this process is often more of a hindrance than a help. Eddie Kim, the CTO of Gusto, recently proved this by building a new AI product line with a team of only five people, bypassing almost every traditional rule of software development.

The Trash Can Method

The team's approach was built on the idea that speed is better than perfection. Instead of spending weeks on a design document that might be wrong, they used what Kim calls the 'trash can' method. They would write a full pull request—code that actually does something—and then review it. If it worked, it stayed. If it didn't, they deleted it and moved on. This turns the development process into a series of rapid experiments rather than a slow march through a series of approvals. It replaces the 'plan then build' model with a 'build to learn' model.

When AI does the building, coordination overhead doesn't scale the engineering; it just slows it down.

This radical reduction in process was made possible by a specific tool stack. By using agentic coding tools like Claude Code, the team collapsed roles. The designer was able to ship code at a high level, and the engineers were able to handle product decisions on the fly. They replaced standups and Slack threads with a 'perma-Zoom' setup—a constant, open connection that allowed for immediate, unscripted communication. The goal was to eliminate the gap between a thought and its execution.

The Gusto Stack

Claude Code for autonomous engineering
Perma-Zoom for real-time coordination
Zero-doc development for rapid iteration

The result was a tier-one product launch in just ten weeks, starting from zero code. This is a new blueprint for the modern startup. In the past, the first six months of a company were spent building infrastructure and hiring managers. Now, a tiny team can use AI to act as their own PM, designer, and DevOps engineer. The barrier to entry is no longer how many people you can hire, but how effectively you can direct the machines.

Key Takeaway

In the age of AI, coordination is a tax that small teams can no longer afford to pay.

04 — Lenny's Newsletter

The Taste Economy

Why judgment is the only skill that survives automation

By Lenny Rachitsky · 11 min read

Editor's note: An exploration of how AI is collapsing professional roles and elevating the importance of aesthetic and functional judgment.

The traditional hierarchy of the software company is being dismantled. For decades, we have separated the 'thinkers' from the 'doers'. Product managers define the what, designers define the how, and engineers execute the build. But as AI agents become capable of handling the execution, this division becomes an obstacle. If an engineer can use a tool to generate a UI or a PM can use a tool to write a functional prototype, the value of the specific technical skill begins to decline. We are entering an era where the ability to build is common, but the ability to decide what is worth building is rare.

The Rise of Professional Taste

Andrew Ambrosino, a lead at OpenAI, argues that 'taste' is emerging as the most important professional capability. In this context, taste is not merely about aesthetics; it is about the ability to judge quality, utility, and coherence. When a machine can generate a thousand variations of a feature, the human's job is to act as the ultimate filter. You need to know which version is elegant, which is intuitive, and which is merely a clever trick. Without this judgment, you are simply drowning in high-speed mediocrity.

When everyone can build anything, the only differentiator is taste.

This shift requires a new kind of worker—the 'zone defense' professional. Instead of being a specialist in a single silo, these individuals operate across the entire product lifecycle. They understand the constraints of the code, the needs of the user, and the vision of the business. They do not just manage a process; they direct an outcome. This is a much harder way to work, as it requires a broad understanding of multiple disciplines, but it is the only way to remain relevant as the technical barriers to entry continue to fall.

Skills for the AI-First Era

High-fidelity judgment
Cross-disciplinary fluency
Directional leadership over execution

The danger is that we might try to preserve our old roles for the sake of comfort. We might try to keep the PMs and the designers in their separate boxes, even when the tools make those boxes obsolete. This will only lead to friction and slow development. The companies that win will be those that embrace the collapse of roles and focus entirely on the quality of the final product. The goal is no longer to manage people; it is to manage the standard of excellence.

Key Takeaway

As execution becomes cheap, judgment becomes the most expensive asset.

05 — Simon Willison

The Agentic Frontier

Self-scaffolding models and the move toward autonomous coding

By Simon Willison · 7 min read

Editor's note: A look at the technical shift from models that chat to models that act.

The current era of AI is defined by the chat interface: a human asks a question, and a model provides an answer. It is a reactive relationship. The next era, however, will be defined by agency. We are seeing the first signs of this in models like Ornith-1.0, which are designed not just to predict the next token, but to use tools and build their own scaffolding. These models do not just wait for instructions; they attempt to solve problems by interacting with the environment, running tests, and correcting their own errors.

Self-Scaffolding Architecture

The concept of self-scaffolding is a significant technical leap. In traditional agentic workflows, a human must provide the 'harness'—the set of tools and the loop that allows the AI to act. Self-scaffolding models are trained to understand how to set up these loops themselves. They can identify when they lack a specific capability and attempt to write the code to acquire it. This moves the AI from being a passenger in the development process to being a co-pilot that can actually take the controls when the situation demands it.

The shift is from models that chat to models that act.

Practical tests of these models show they are already handling complex, multi-step tasks with ease. For example, asking a model to find specific logic within a large codebase and then modify it requires a level of reasoning that goes far beyond simple pattern matching. It requires an understanding of the intent behind the code. As these models become more proficient at navigating these agentic harnesses, the role of the human coder will shift from writing syntax to defining intent and verifying results.

Key Features of Agentic Models

Tool-use proficiency
Autonomous error correction
Multi-step reasoning capabilities

This transition will not be without its risks. An autonomous agent that can write and execute code has the potential to cause significant damage if it lacks proper guardrails. The challenge for the next year of development will be finding the balance between giving these models enough agency to be useful and enough constraint to be safe. We are building machines that can think for themselves; we must ensure they still think in ways that serve our interests.

Key Takeaway

The future of AI is not in better conversation, but in better action.

06 — The Marginalian

The Cosmic Lottery

On the absurdity of the self and the necessity of love

By Maria Popova · 8 min read

Editor's note: A philosophical reflection on the randomness of existence and the fragility of identity.

The self is a fiction. We like to believe we are the authors of our own lives, the steady captains of our own souls. But if we look closely at the variables that make us who we are, the illusion begins to crumble. Our DNA, our birthplace, the specific culture into which we were born, the exact sequence of chance events that led to our parents meeting—none of these were chosen. We are the result of a cosmic lottery, a series of improbable accidents that have coalesced into a single, thinking entity. To claim ownership of this identity is, in a sense, an absurdity.

The Illusion of Authorship

Iris Murdoch observed that the self is a place of illusion. We spend much of our lives building up a defensive wall of opinions, identities, and certainties. We use these to convince ourselves that we are solid, permanent things. But this is a way of hiding from the world. True clarity comes from the attempt to pierce this veil—to see the world as it is, rather than through the distorted lens of our own egos. This is not a simple task; it requires a discipline of the mind and a willingness to let go of the stories we tell ourselves.

There is no reason for you to be here, to be you.

Blaise Pascal felt this weight with terrifying precision. He looked at the infinite immensity of the universe and felt the crushing smallness of his own existence. He was amazed, and terrified, that he existed at all. This is the fundamental tension of being human: we are finite, accidental beings living in an infinite, indifferent universe. We occupy a tiny sliver of time and space, and yet we possess a consciousness that can grasp the scale of the whole. This mismatch is where the terror lies.

The Components of the Self

Biological chance (DNA)
Environmental contingency (Birthplace)
Historical accident (Culture/Time)

If reason leads us to the conclusion that we are accidental, then what is left? Perhaps it is love. Not the romanticized version found in novels, but love as an affirmation of existence itself. Love is the way we respond to the improbable nature of life. It is the substance that outweighs the antimatter of chance. In a universe that offers no inherent meaning, the act of loving—of connecting with the world and with others—is the only way to survive the helplessness of being born. It is our only way to make the chaos bearable.

Key Takeaway

Identity is a story we tell to manage the chaos of our own accidental existence.