The Deep Feed

01 — Cal Newport

The Productivity Paradox

Why new tools often fail to move the needle on output

By Cal Newport · 6 min read

Editor's note: A necessary reality check for those expecting AI to solve every efficiency problem overnight.

The history of technology is littered with the wreckage of failed promises. In the early 1990s, the arrival of the personal computer was supposed to trigger an explosion in white-collar efficiency. Spreadsheets, word processors, and databases were marketed as the ultimate tools for the modern professional. Yet, as economists noted at the time, productivity in the office sector remained stubbornly flat. The expected surge in output never arrived. Instead, we saw a period of stagnation where the cost of computing rose, but the actual work produced per hour barely moved.

The Ghost in the Machine

This phenomenon, often called the productivity paradox, occurs because technology does not exist in a vacuum. A tool can make a specific task faster—typing a memo is certainly easier with a word processor than a typewriter—but it also creates new, non-productive work. Email replaced the physical memo, but it also introduced a constant stream of interruptions that eroded deep focus. We spent more time managing the tools than doing the work the tools were meant to facilitate. The computer became a source of distraction rather than a lever for output.

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

We are currently seeing the same pattern repeat with generative AI. The assumption is that because these models can write code, draft emails, and summarise reports, our total output must increase. This ignores the reality of integration. Just as it took decades for the workforce to truly harness the power of the PC, we are currently in the messy, inefficient phase of figuring out where AI adds value and where it simply adds more noise to an already crowded digital environment.

Why productivity stalls

New tasks created by the tool itself
Increased coordination overhead
The distraction of constant connectivity
The time required to master new workflows

To avoid this trap, we must stop viewing AI as a magic wand and start viewing it as a structural change. The goal should not be to 'use AI' to do more of the same, but to rethink the work itself. If the tool changes the speed of execution, it must also change the nature of the tasks we choose to perform. Otherwise, we are simply running faster on a treadmill that stays in the same place.

Key Takeaway

Efficiency in a single task does not equal productivity in a whole system.

02 — Lenny's Newsletter

The End of the Process-Heavy Agency

How Gusto built a product line in 10 weeks with almost no overhead

By Claire Vo · 12 min read

Editor's note: A blueprint for the lean, AI-driven team of the future.

Traditional product development is a slow, heavy machine. It relies on a predictable sequence: product managers write specs, designers create high-fidelity mocks in Figma, and engineers translate those designs into code using Jira to track every ticket. This structure is designed to manage complexity and reduce risk, but it also creates massive coordination overhead. For a small, fast-moving team, this process is often more of a hindrance than a help. It slows down the feedback loop and forces people to spend more time talking about work than actually doing it.

The Five-Person Sprint

Eddie Kim, CTO of Gusto, recently challenged this model. He led a team of five—three engineers and one designer—to build 'Gusto Cofounder' from scratch in just ten weeks. They did not use a PM. They did not use Figma. They did not use Jira. Instead, they used an agentic stack centered around Claude Code. By removing the traditional layers of planning and documentation, the team eliminated the friction that usually eats up the first month of any new project. They moved directly from idea to execution.

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

One of their most radical tactics was the 'trash can method'. Rather than writing long, static product requirement documents that gather dust, the team would write, review, and then delete entire pull requests as a way of making product decisions. The code itself became the specification. This allowed for rapid iteration and immediate testing of ideas. If a feature didn't work, it was discarded immediately, not after weeks of debate in a meeting room.

The Gusto Stack

Claude Code for primary engineering
Perma-Zoom for real-time coordination
Direct-to-code prototyping
Minimalist documentation

This approach turns the role of the engineer into something closer to a director. The human provides the intent and the judgment, while the AI handles the heavy lifting of syntax and implementation. This shift allows a tiny team to achieve the output of a much larger department. The constraint is no longer how many people you can hire, but how clearly you can communicate your direction to the agents you are managing.

Key Takeaway

In the age of AI, process is often just a way to mask a lack of clear direction.

03 — Lenny's Newsletter

The Ladder of AI Leverage

Moving from simple assistance to autonomous systems

By Colin Matthews · 9 min read

Editor's note: A framework for understanding how to actually grow your value in an automated economy.

For years, the role of the Product Manager was defined by coordination. PMs were the glue that held engineers, designers, and stakeholders together. They spent their days in meetings, aligning different departments and ensuring everyone was working toward the same goal. But as AI begins to handle the heavy lifting of project management and technical coordination, that version of the role is dying. The PMs who survive will not be those who coordinate people, but those who leverage machines.

The Three Rungs of Leverage

To understand this shift, we can look at leverage through three distinct lenses: personal, product, and systems. Personal leverage is about your own output—using AI to draft a document or research a topic. Product leverage is about the speed at which your team can ship—using AI to prototype features or query data without waiting for an analyst. Systems leverage is the highest form, where you build repeatable, automated workflows that allow AI to complete multi-step tasks with minimal supervision.

The best PMs are no longer just coordinators; they are builders.

As you move up these rungs, the nature of your work changes. On the first rung, you are still doing the work, just faster. On the top rung, you are managing a system that does the work for you. This requires a different set of skills. You need to move away from being a 'communicator' and toward being a 'technical architect'. You need to understand how to set up evaluations, how to direct agents, and how to verify the quality of machine-generated output.

Levels of AI Integration

Assistance: AI helps you write or research
Delegation: You pass a task to AI and review the result
Automation: AI completes multi-step processes and checks its own work

The goal is not to automate everything, but to identify where the highest leverage lies. A PM who can prototype their own ideas using code-generation agents is worth ten PMs who can only write well-formatted tickets. The competitive advantage in the next decade will belong to those who can bridge the gap between business intent and technical execution using the tools available.

Key Takeaway

Leverage comes from moving from doing tasks to managing systems.

04 — Lenny's Newsletter

The Open-Weight Rebellion

Why control and cost are the new frontiers of AI strategy

By Lenny Rachitsky · 8 min read

Editor's note: The battle for AI dominance isn't just about intelligence; it's about who owns the weights.

For the past two years, the AI industry has been dominated by a handful of closed-source giants. If you wanted the best intelligence, you had to use their APIs, pay their prices, and follow their rules. This created a massive vendor dependency. If a provider changed their pricing model or restricted certain types of usage, entire businesses were left vulnerable. However, the rise of high-quality open-weight models is beginning to break this monopoly.

Performance Without the Lock-in

Models like GLM-5.2 are proving that open-weight alternatives can compete with the frontier labs on almost every benchmark. These models allow teams to run inference on their own hardware, fine-tune on their own proprietary data, and—most importantly—switch providers without changing a single line of code. This is a fundamental shift in power. It turns AI from a rented service into a piece of infrastructure that a company can actually own and control.

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

The economic argument is just as strong. For agentic workflows—where an AI might run for 45 minutes performing hundreds of tool calls—the costs of using closed APIs can escalate rapidly. Open-weight models, especially when run through efficient routing services, offer a much more sustainable cost curve. For companies building at scale, the ability to predict and control their compute spend is the difference between a profitable product and a money pit.

Advantages of Open-Weight Models

Lower long-term costs for high-volume tasks
Ability to fine-tune on private data
Elimination of vendor lock-in
Greater control over data privacy and security

We are entering an era where the 'best' model is no longer a single, monolithic entity. Instead, it will be a hybrid stack: using the most powerful closed models for complex reasoning, while delegating the bulk of routine, high-volume work to efficient, locally-hosted open models. This is how sophisticated technical teams will maintain both high performance and high margins.

Key Takeaway

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

05 — Dwarkesh Podcast

The Math of Intelligence

What the progress in mathematics tells us about the future of AGI

By Dwarkesh Patel · 15 min read

Editor's note: A deep look at the most rigorous testing ground for artificial intelligence.

Mathematics is perhaps the most unforgiving environment for artificial intelligence. Unlike creative writing or coding, where 'good enough' often suffices, math requires absolute precision and logical consistency. There is no room for hallucination in a proof. Consequently, the progress AI is making in mathematics serves as a high-fidelity signal for its general intelligence. If an AI can solve a Millennium Prize problem, it is doing something fundamentally different than just predicting the next word in a sentence.

The Spiky Frontier

Grant Sanderson, the mind behind 3Blue1Brown, observes that AI progress is not a smooth, upward curve. Instead, it is 'spiky'. There are specific domains where AI is making massive leaps—such as geometry or certain types of algebraic manipulation—while other areas remain entirely out of reach. This suggests that current models are not yet general intelligences, but they are becoming incredibly specialised tools that can navigate specific logical landscapes with superhuman speed.

AI is making much faster progress in math than in other fields, showing us what progress in other fields will eventually look like.

One of the most profound questions is whether AI will increase or decrease human understanding. If an AI produces a proof for a major conjecture that no human can follow, has it actually 'solved' the problem in a meaningful way? We may find ourselves in a position where we possess the answers to the universe's most difficult questions, but lack the cognitive capacity to understand the logic that led us there. This creates a new kind of epistemological gap.

Key tensions in AI Mathematics

Discovery vs. Understanding
The gap between formal proofs and human intuition
The ability to find 'hidden bridges' between different mathematical fields
The limits of reinforcement learning in non-structured environments

Ultimately, the progress in math is a roadmap. The way AI masters formal logic today will dictate how it masters reasoning, law, and strategic planning tomorrow. We are watching the blueprint of general intelligence being drawn in the language of mathematics.

Key Takeaway

Math is the ultimate stress test for whether AI is truly reasoning or just simulating reasoning.

06 — Simon Willison

The Self-Scaffolding Agent

A new breed of models that build their own paths to a solution

By Simon Willison · 5 min read

Editor's note: A technical look at the next step in autonomous agentic behavior.

The current limitation of AI agents is their tendency to get stuck. When an agent is given a complex task, it often follows a linear path, and if it hits a wall—a compilation error, a missing file, or a misunderstood instruction—it tends to loop or fail. The next evolution in agentic coding is 'self-scaffolding': the ability for a model to recognise its own limitations and proactively build the tools or context it needs to proceed.

Ornith-1.0 and the Agentic Loop

The release of Ornith-1.0 represents a significant step in this direction. Built on top of existing architectures like Gemma 4 and Qwen 3.5, this model is specifically tuned for agentic workflows. In testing, it has shown a remarkable ability to run through long sequences of tool calls without losing the thread of the original objective. It doesn't just execute commands; it manages the environment in which those commands are run.

The model is able to run the agent harness over many tool calls in a proficient way.

This capability is what separates a simple chatbot from a true autonomous agent. A chatbot waits for your input; an agent anticipates the next step. When Ornith-1.0 is asked to find specific code within a large repository, it doesn't just search; it navigates, understands the structure, and builds a mental map of the codebase to ensure its findings are accurate. It is essentially building its own scaffolding as it climbs the problem.

What makes self-scaffolding different

Proactive error correction
Dynamic context management
Multi-step tool orchestration
Reduced reliance on human intervention for mid-task errors

As these models become more proficient at managing their own execution, the role of the human developer will shift even further toward high-level architecture. We will no longer be writing the code; we will be designing the environments in which these self-scaffolding agents operate. The complexity is not disappearing; it is simply moving up one level of abstraction.

Key Takeaway

True autonomy requires a model that can build its own way out of a corner.