The Deep Feed

01 — Lenny's Newsletter

The Content Machine: Automating Thought Without Losing the Soul

How Alex Lieberman uses Claude to scale distribution while avoiding the trap of AI slop

By Lenny Rachitsky · 12 min read

Editor's note: A masterclass in using AI as a multiplier for human insight rather than a replacement for it.

The blank page is the most expensive piece of real estate in a founder's day. It represents a total halt in momentum, a moment where the desire to communicate meets the friction of execution. Alex Lieberman, the co-founder of Morning Brew, has solved this not by writing more, but by building a system that makes writing the easiest part of his day. He has constructed what he calls a 'Content Machine' inside Claude, a workflow designed to extract his specific ideas and turn them into polished assets without the generic, hollow feel of standard AI outputs.

The Myth of the AI Slop Problem

Most people blame the model when AI content feels cheap or repetitive. They call it 'slop'. Lieberman argues this is a mistake. The model is merely a sculptor; if the clay you provide is unformed and generic, the statue will be equally unremarkable. The failure point isn't the LLM; it is the input. If you ask an AI to 'write a post about marketing', you get a post that sounds like every other post on LinkedIn. The machine cannot invent the unique perspective you haven't shared with it. To get quality, you must provide high-fidelity raw material.

AI slop is mostly a people problem, not a model problem. The model is shaping clay, not inventing ideas.

Lieberman's system begins with an 'Oracle'. This isn't a search engine, but a scanner that looks through seven days of Slack messages, Notion docs, and meeting notes to find 'content spikes'—the moments where real, unscripted ideas are actually discussed. Instead of staring at a cursor, he is interviewed by a panel of AI personas. These personas, ranging from Larry King to Barbara Walters, are instructed to be relentless. They don't accept surface-level answers; they push for the specifics that make a story worth reading.

Codifying the Voice

The most technical part of the machine is the Markdown file containing his personal voice guide. This file acts as a set of guardrails. It doesn't just list topics; it defines his linguistic patterns, his specific brand of 'self-deprecating confidence', and his preferred structural formulas. By feeding this to Claude, the machine drafts in his register. It stops trying to sound like the 'average' of the internet and starts trying to sound like Alex Lieberman. This turns the AI from a ghostwriter into a highly skilled editor of his own thoughts.

The Six-Step Workflow

Oracle: Scanning internal data for idea spikes
Interview: Extracting raw thoughts via persona-led questioning
Voice Calibration: Applying the Markdown style guide
Drafting: Generating the initial structure
Writer's Council: Peer review by specialized AI personas
Lessons Loop: Refining the system based on published results

Finally, the draft passes through a 'Writer's Council'. This is a group of AI personas, including one specifically designed to be an 'AI slop allergist', that score the content. If a piece doesn't hit a 9 out of 10, it goes back into the loop. This automated standard ensures that even when Lieberman is busy, the quality remains high. It is a system that doesn't just create content; it manages the reputation of the creator.

Key Takeaway

AI is a multiplier of your existing intelligence, not a substitute for it; if your input is shallow, your output will be too.

02 — Stratechery

The Token Factory: Why Intelligence is the New Commodity

Ben Thompson on the shifting economics of the AI era

By Stratechery · 10 min read

Editor's note: A necessary look at how the unit economics of software are being rewritten by the cost of inference.

For decades, the software industry operated on the principle of zero marginal cost. Once you wrote the code, serving the millionth customer cost essentially nothing. This led to the era of Aggregation Theory, where the winners were those who controlled the demand. But AI is breaking this rule. Every time a user interacts with a sophisticated model, a cost is incurred. We are moving from a world of pure software to a world of 'token factories', where the cost of goods sold (COGS) is a direct, linear function of usage.

The R&D vs. COGS Distinction

There is a common misconception that open-weights models, like those coming out of China, make AI 'free'. This confuses research and development (R&D) with the cost of serving. While downloading a model's weights avoids the massive fixed cost of training it, it does nothing to mitigate the variable cost of running it. Running inference—the process of generating a response—requires massive amounts of compute. If a provider spends 50 cents to generate every dollar of revenue, they are running a very different business than a traditional SaaS company with 95% margins.

Tokens are not a commodity. What is fungible is what is constructed from tokens: intelligence.

This distinction is vital because it changes how we measure value. Nvidia's Jensen Huang talks about 'token factories', focusing on metrics like tokens-per-second or tokens-per-watt. This was a useful framework during the initial ChatGPT craze, when the goal was simply to deliver text as fast as possible. But as we enter the 'reasoning era', the math gets more complicated. Reasoning models use 'chain-of-thought' tokens—internal monologue steps that the user doesn't see but that cost money to compute. A model that is 'cheaper' per token might actually be more expensive if it requires ten times as many tokens to reach the correct conclusion.

The End of Fungibility

In traditional commodity markets, a gallon of oil is a gallon of oil. In the AI market, a token from one model is not equivalent to a token from another. The value lies in the resulting intelligence. If two models provide the same correct answer, the intelligence is fungible, but the path taken to get there—the number of tokens, the latency, the reasoning steps—is where the competitive advantage is won or lost. The battle is no longer just about who has the most tokens, but who can generate the most intelligence per unit of compute.

The New AI Economic Metrics

Inference COGS: The direct cost of serving each user request
Reasoning Efficiency: The ratio of intelligence produced to tokens consumed
Agentic Throughput: How many steps an agent can take before the cost becomes prohibitive
Intelligence Fungibility: The degree to which different models' outputs can be swapped

As Chinese models like Kimi K3 approach the capabilities of Western counterparts, the competition will move from pure capability to economic efficiency. The winners won't just be those with the smartest models, but those who can deliver that intelligence at a sustainable margin. The era of 'free' AI is a mirage; the era of efficient intelligence is just beginning.

Key Takeaway

The software industry is losing its zero-marginal-cost advantage; the new winners will be those who optimize the cost of intelligence, not just the scale of distribution.

03 — Lenny's Newsletter

The Sabbatical: A Strategic Disconnection

Why 'Big Time Off' is a requirement for long-term career clarity

By Lenny Rachitsky · 8 min read

Editor's note: An argument against the cult of constant productivity.

The modern career is often a treadmill of increasing velocity. We optimize for the next promotion, the next raise, and the next milestone, rarely pausing to ask if the direction itself is correct. Lenny Rachitsky's own pivot from a high-growth role at Airbnb to running his own successful newsletter wasn't a sudden epiphany; it was the result of a three-month sabbatical. It was the first time in fifteen years he had stepped away from the default path, and it provided the space necessary to realise that his heart was no longer in his existing work.

The Necessity of Distance

Taking significant time away from the day-to-day is not a luxury; it is a requirement for the mental and emotional shifts that reveal what comes next. DJ DiDonna, author of 'Big Time Off', argues that sabbaticals allow individuals to move from the 'urgent' back to the 'important'. When you are in the thick of a career, the urgent tasks—emails, meetings, immediate problems—crowd out the long-term questions of purpose and direction. Distance provides the perspective needed to see the forest, not just the trees.

Significant time away from your day-to-day is necessary for the mental, physical, or emotional shift that will reveal what's next.

A sabbatical does not have to mean quitting a job or traveling the world. It can be a period of focused tinkering, reading, or even a silent retreat. The goal is to break the cycle of reactive living. For many, the fear of taking time off is rooted in the idea of irresponsibility. However, as DiDonna's research suggests, the most successful individuals often use these breaks to recalibrate, ensuring that when they do return to work, they are doing so with intention rather than momentum.

The Business Case for Time Off

There is a growing tension between individual needs for rest and corporate demands for continuity. Yet, there is a compelling business case for supporting sabbaticals. Employees who return from meaningful breaks are often more engaged, more creative, and better aligned with their roles. Conversely, forcing employees to stay on a path that no longer fits them leads to quiet quitting or sudden, disruptive departures. Companies that build supportive policies for 'Big Time Off' are essentially investing in the long-term sustainability of their talent.

Archetypes of Sabbaticals

The Re-evaluator: Stepping back to decide on a total career pivot
The Skill-Builder: Using time to master a new discipline without distraction
The Restorer: Focusing purely on physical and mental recovery from burnout
The Explorer: Engaging in travel or new experiences to broaden perspective

Ultimately, the question isn't whether you can afford to take time off, but whether you can afford not to. The cost of staying on a path that has become obsolete is far higher than the temporary loss of productivity during a break. A sabbatical is an investment in the person you are becoming, rather than the professional you have been.

Key Takeaway

Constant momentum is not progress; strategic disconnection is often the only way to find a new direction.

04 — Simon Willison

The Death of the Chatbot: Why AI is Becoming an Actor, Not a Partner

Insights from the Claude Code team on the future of agentic workflows

By Simon Willison · 11 min read

Editor's note: A look at how the role of the engineer is shifting from writing code to managing intent.

The way we interact with AI is undergoing a fundamental shift. We are moving away from the 'chat' paradigm—where a human prompts and a machine responds—and toward an 'agentic' paradigm, where the machine acts. During a recent fireside chat with members of Anthropic's Claude Code team, the reality of this shift became clear. The goal is no longer to have a conversation with a bot, but to delegate complex, multi-step tasks to a system that can operate with increasing autonomy.

From Monitoring to Delegating

A year ago, using a coding agent felt like babysitting. You had to watch every permission prompt and scrutinise every file the AI touched. Today, that experience has changed. As models like Claude 3.7 and the newer Fable have matured, the need for constant oversight has diminished. The team at Anthropic is now able to delegate the 'menial implementation' to the AI, freeing them to focus on higher-level architectural decisions and user experience. The friction of 'permission fatigue' is being replaced by a sense of delegation.

The timeline between having an idea and building it is shrinking from months to days. This increases the value of product taste and business sense.

This compression of the development cycle has profound implications for the engineering skillset. When the time between an idea and a working prototype drops from six months to one week, the bottleneck is no longer the ability to write code; it is the ability to know *what* to build. The value of pure syntax knowledge is declining, while the value of product sense, business intuition, and 'taste' is skyrocketing. Engineers must become more than just implementers; they must become product thinkers.

The End of the Prompt Engineering Era

Interestingly, the technical methods for controlling these models are also evolving. In the early days, developers spent vast amounts of time crafting complex system prompts and long lists of 'do not' instructions. However, as models become more capable, these brute-force methods are becoming less effective. The latest iterations of Claude show that adding examples to a system prompt can actually reduce quality in some cases. The models are becoming better at understanding intent without being told every single rule, making the 'prompt engineering' era look increasingly like a temporary workaround for less capable intelligence.

The Shifting Engineer Skillset

Decreasing value: Syntax memorisation and boilerplate implementation
Increasing value: Product sense and user experience design
Increasing value: Business logic and strategic alignment
Increasing value: Managing agentic workflows and verifying complex outputs

As we move toward 'auto mode', the role of the human becomes one of an orchestrator. We set the goals, define the constraints, and verify the results, but the heavy lifting of execution is handled by the agent. This requires a new kind of discipline: the ability to be ambitious enough to take on harder problems, while being rigorous enough to ensure the machine's autonomy doesn't lead to drift.

Key Takeaway

As AI handles the implementation, the most valuable skill in technology is no longer knowing how to build, but knowing what is worth building.

05 — Experimental History

The Human Premium: Why Effort Still Matters

Reflections on the 2026 Blog Jamboree and the value of the 'effort post'

By Adam Mastroianni · 9 min read

Editor's note: A defense of human idiosyncrasy in an age of automated content.

In an internet flooded with AI-generated content, how do we distinguish between a machine mimicking thought and a human actually thinking? Adam Mastroianni's annual Blog Jamboree provides a surprising answer. By running submissions through an AI detector, he found that the most compelling pieces—the ones that actually moved readers—were 100% human-written. These were not just 'posts'; they were 'effort posts'. They were pieces where the author had bothered to cite sources, run experiments, and structure an argument that felt earned rather than generated.

The Anatomy of an Effort Post

An effort post is defined by its refusal to take the path of least resistance. While an AI might summarize a topic, a human writer like Jane Cook will conduct a makeshift experiment with molten agar to understand why bacteria behave a certain way in boiling water. The value lies in the friction. The fact that the writer struggled, investigated, and arrived at a conclusion through a specific, idiosyncratic process is exactly what makes the writing resonate. The machine can simulate the conclusion, but it cannot simulate the struggle.

The AI is born anew every time it speaks... you will go home alone and nobody will know your name.

This distinction becomes even sharper when looking at the nature of digital interaction. As some writers have noted, a conversation with a chatbot is an illusion. The AI has no continuity; it is a new entity every time a new prompt is sent, performing a role based on the immediate transcript. It is a highly skilled actor, but there is no person behind the curtain. Human writing, by contrast, is an extension of a persistent identity. When we read a blog post, we are engaging with a mind that has a history, a set of biases, and a specific way of seeing the world.

The Power of the Specific

The winners of the Jamboree succeeded because they embraced the specific and the strange. Whether it was a piece on the history of the word 'vape' in national constitutions or a deeply personal reflection on ancestral roots, these writers avoided the generic. They didn't write for the algorithm; they wrote to uncover a truth. In a world of 'average' intelligence, the human premium is found in the outliers—the weird, the highly specific, and the deeply personal.

Characteristics of High-Value Human Content

Empirical investigation (doing the work to prove the point)
Idiosyncratic voice (writing that cannot be easily replicated by a prompt)
Personal stakes (the sense that the author actually cares about the outcome)
Structural intentionality (using form to reinforce the argument)

As the cost of generating text approaches zero, the value of 'effort' will only increase. We will likely see a bifurcation of the internet: a massive, low-cost layer of AI-generated utility content, and a high-value, scarce layer of human-driven thought. The latter will be the only place where true connection and influence can be found.

Key Takeaway

In an era of infinite, effortless content, the only way to stand out is to do the work that a machine cannot: the work of being human.