The Deep Feed

01 — Simon Willison

The $165,000 Rewrite

How agentic engineering turned a decade-long impossibility into an eleven-day sprint

By Simon Willison · 12 min read

Editor's note: A case study in how LLMs are making the 'impossible' rewrite a standard engineering task.

For decades, the rule in software engineering was simple: never rewrite a large, successful system from scratch. It is a fool's errand that consumes years of developer time and almost always results in a product that is less stable than the original. Joel Spolsky famously codified this in the early 2000s. But the arrival of high-reasoning coding agents has broken this rule. Jarred Sumner recently completed a total rewrite of the Bun runtime from Zig to Rust, not over years, but in a matter of days. This wasn't a human-led effort assisted by AI; it was a coordinated parallel agentic workflow that treated the original codebase as a set of constraints to be satisfied.

The Memory Problem

The decision to move to Rust was driven by a specific, recurring headache: memory safety. Bun, written in Zig, struggled with a mix of garbage collection and manual memory management. This combination led to a persistent list of bugs—use-after-free errors and double-frees that kept the team awake at night. Rust solves this through its ownership model and RAII-like automatic cleanup. By moving to Rust, the goal was to turn runtime crashes into compile-time errors. The challenge was the sheer scale of the task: adding over a million lines of code.

Our bugfix list felt bad and I was tired of going to sleep worrying about crashes in Bun.

The process relied on a language-independent TypeScript test suite. This suite acted as the ultimate arbiter of truth. The agentic harness would attempt to port segments of the code, run the tests, and when a failure occurred, the agent would iterate on the code until the assertions passed. This created a feedback loop where the human's role shifted from writing lines of code to managing the process that generates the code. It was an adversarial review system where the agent had to prove its work against a million assertions.

The Agentic Workflow

Language-independent conformance testing
Adversarial code review loops
Automated error correction via model prompting
Human-in-the-loop monitoring of high-level workflows

The result was a 10% speed increase on Linux. While most users noticed nothing, the engineering reality changed forever. The ability to perform massive, high-fidelity migrations in days rather than years changes how we think about technical debt. We are moving from a world where we must live with our early mistakes to a world where we can simply rewrite our way out of them.

Key Takeaway

Coding agents turn the 'never rewrite' rule into a strategic option.

02 — Not Boring

The Nuclear Renaissance

How data centers are funding the next generation of energy

By Packy McCormick · 10 min read

Editor's note: The massive demand for compute is inadvertently solving the energy crisis.

On July 4th, 2026, Aalo Atomics reached criticality, marking the fourth advanced nuclear company to hit its mission deadline in a single year. Just three years ago, the idea of multiple advanced reactors reaching criticality within such a tight window seemed like a fantasy. This rapid acceleration is not a coincidence. It is the direct result of a massive, concentrated demand engine: the data center. As the world builds out the infrastructure for AI, the hunger for reliable, carbon-free, 24/7 power has moved from a theoretical concern to a primary driver of industrial policy.

The Customer Shift

Historically, the primary customers for nuclear power have been massive utilities. These organizations are often slow, risk-averse, and burdened by legacy structures. However, the new wave of advanced, smaller reactors is finding a different kind of buyer. Silicon Valley's hyperscalers—the companies building the massive clusters required for frontier models—are willing to pay a premium for speed and reliability. They need power now, and they have the capital to fund whoever can deliver a working solution most quickly. This shifts the competition in the nuclear sector from pure science to manufacturing and deployment speed.

The field shifts from science to manufacturing. Whoever can deploy most quickly, with a product that customers want, at the best price, will win.

This demand is also driving a secondary market in gas turbines. Companies like American Turbine are moving away from large, complex machines that require specialized expertise, opting instead for small, highly manufacturable turbines. The goal is speed. If a data center needs more megawatts, they don't wait for a massive plant; they simply buy more small turbines. In the American context, this leverages a massive domestic advantage: abundant natural gas. The ability to turn gas into power to fuel ideas is becoming a core component of national competitiveness.

Drivers of the Nuclear Boom

Hyperscaler demand for 24/7 clean energy
Shift from utility-scale to modular, manufacturable designs
The convergence of AI compute needs and energy policy
Competitive pressure to move from R&D to commercial deployment

We are witnessing the creation of a domestic supply chain built on real-world demand rather than just subsidies. When the buyers are willing to pay for speed, the entire pace of innovation accelerates. The 'Age of Miracles' isn't just a concept; it's being built, reactor by reactor, to keep the lights on in the data centers that run our world.

Key Takeaway

Data center demand is the turbocharger for the advanced nuclear industry.

03 — Lenny's Newsletter

The Sol Benchmark

Why taste is becoming the ultimate metric in AI performance

By Claire Vo · 15 min read

Editor's note: Standard benchmarks are failing to capture what actually makes an AI useful for high-level work.

The era of measuring AI solely by logic puzzles and math problems is ending. As models become more capable at general reasoning, the real differentiator is no longer just 'correctness,' but 'utility'—how well the model understands intent, style, and the unstated requirements of a professional task. Claire Vo's recent benchmark of the GPT-5.6 suite against Claude Fable 5 highlights this shift. Using a 'Claire Weighted Index,' which places 70% of the weight on human taste and 30% on technical benchmarks, the results show a clear winner: GPT-5.6 Sol.

The Pedantry Trap

One of the most striking findings was the failure of Claude Fable 5 in collaborative workflows. While Fable demonstrated high precision, it suffered from what can be described as a 'pedantry problem.' It would often get stuck on minor technicalities or refuse to move forward without perfect input, making it difficult to use in the messy, iterative process of product management. In contrast, GPT-5.6 Sol showed a capacity to 'break through'—to make reasonable assumptions and maintain momentum even when the initial prompt was imperfect. For a product manager, a model that can iterate is far more valuable than one that is perfectly, but uselessly, precise.

Precision is a baseline; the ability to navigate ambiguity is the real competitive advantage.

The benchmark tested several high-stakes categories, including PRD writing, prototyping, and agentic voice. Sol's victory wasn't just about being 'smarter'; it was about being more useful in specific, real-world workflows. For example, in browser automation, the ability to execute complex, multi-step tasks like managing LinkedIn replies without constant human intervention became a decisive factor. The winner is the model that acts more like a teammate and less like a calculator.

Key Model Differentiators

Agentic Voice: The ability to maintain persona and intent
Iterative Reasoning: Moving past technical roadblocks
Contextual Assumption: Making smart guesses in ambiguous tasks
Workflow Integration: How well the model fits into existing tools

As we move toward agentic systems, the metric of success will shift from 'how much does this model know' to 'how much can this model do.' The models that win will be those that understand not just the text of a prompt, but the goal of the human behind it.

Key Takeaway

In the age of agents, utility and taste matter more than raw logic.

04 — Lenny's Newsletter

The Authenticity Tailwind

How Instagram is preparing for a synthetic world

By Lenny Rachitsky · 14 min read

Editor's note: Adam Mosseri discusses the changing role of the designer and the creator in an AI-saturated ecosystem.

Adam Mosseri, the head of Instagram, is facing a fundamental question: what happens to a social platform when the content is no longer guaranteed to be human? The rise of AI-generated content is often viewed as a threat to authenticity, but Mosseri argues it could actually act as a tailwind. As synthetic content becomes ubiquitous, the value of genuine human identity and connection will likely increase. Instagram's challenge is to build the infrastructure that distinguishes the two, ensuring that creators can maintain their identity even in a world of infinite, easy-to-generate imagery.

The Death of the Specialist

This shift is also transforming the internal structure of product teams. The traditional model of siloed specialists—PMs, designers, data scientists—is dissolving. In its place, we are seeing the rise of the 'product staff': lean pods of generalist operators who can handle the entire lifecycle of a feature. This is a direct response to the increased velocity enabled by AI. When the technical barriers to prototyping and data analysis are lowered, the premium moves from functional expertise to strategic orchestration.

The rise of AI-generated content is a tailwind for Instagram, provided we solve for creator identity.

Mosseri's perspective is shaped by a history of both massive successes and notable failures. He cites Facebook Home and the early iterations of Reels as lessons in the danger of misreading user intent. The current era requires a different kind of foresight: anticipating how users will interact with synthetic media. It is not just about detecting AI content, but about creating a space where human-led creativity feels distinct and rewarded.

The New Product Landscape

Generalist operators replacing functional specialists
Identity-centric design in a synthetic content world
Algorithmic transparency as a tool for user trust
Rapid prototyping as a standard, not a luxury

For Instagram, the goal is to ensure that as the cost of content production drops to zero, the value of human connection does not follow suit. The platform must become the place where authenticity is not just possible, but verifiable.

Key Takeaway

AI makes human identity the most valuable commodity on social media.

05 — The Marginalian

The Fractal Nature of Loneliness

Finding meaning in the spaces between us

By Maria Popova · 18 min read

Editor's note: A philosophical look at the human condition through the lens of isolation and connection.

Loneliness is often treated as a temporary state to be solved, but it is more accurately described as a fundamental condition of being. We are born into isolation and we die in it. Every human experience is contained within a single mind and a single body, creating an inherent gap between ourselves and the rest of the world. This is not a flaw in the human design, but its defining characteristic. All attempts at connection—art, poetry, friendship—are attempts to bridge this gap, to reach out from one island of consciousness to another.

The Three Core Lonelinesses

Drawing on the work of Jungian analyst Robert A. Johnson, we can see that loneliness is not a monolith. It branches into different emotional hues. There is the past-oriented loneliness of mourning what is gone, and the future-oriented loneliness of longing for what has not yet arrived. These are temporal lonelinesses, rooted in our relationship with time. But there is a third kind: an existential loneliness that exists outside of time. It is the disorientation felt when one's smallness meets the immensity of the eternal.

Loneliness is fractal—the closer you look at the granularity of life, the more you see it branching into myriad lonelinesses.

This existential loneliness is not something to be cured; it is something to be inhabited. It is the source of wonder and the driver of creativity. When we face the vastness of the universe or the depth of love, we experience a specific kind of isolation that is both terrifying and beautiful. It is the feeling of being a finite creature standing at the edge of the infinite. Poets, more than philosophers, seem to understand this truth, using language to map the territories that logic cannot reach.

Kinds of Loneliness

Past-oriented: The ache of what once was
Future-oriented: The longing for what could be
Existential: The encounter with the eternal
Social: The pain of being misunderstood or invisible

To live fully is to accept these various forms of isolation. We do not overcome loneliness; we learn to use it as a compass, a way to navigate the immense ocean of time and chance that defines our existence.

Key Takeaway

Loneliness is the baseline of human existence, not a problem to be solved.

06 — Stratechery

The Data War

Verifiability as the new frontier in the AI race

By Stratechery · 12 min read

Editor's note: As models scale, the battle shifts from compute power to the quality of verifiable data.

The current arms race in artificial intelligence is moving past the stage of mere parameter counts and compute clusters. We are entering a phase where the primary differentiator is access to verifiable, high-quality data. As frontier labs exhaust the available supply of public internet text, the struggle to secure unique, reliable datasets—ranging from proprietary enterprise data to specialized scientific repositories—has become the central strategic tension in the industry.

The Verifiability Gap

The problem with the current web is that it is increasingly polluted by synthetic content. If models are trained on the output of other models, they risk a feedback loop of degradation. This makes 'verifiable data'—data that can be traced back to a human source or a physical reality—incredibly valuable. Companies like Meta, xAI (with Grok), and the major frontier labs are all pivoting their strategies to secure these 'clean' data streams. The ability to prove that a model is learning from reality, rather than from a hallucinating peer, will define the next generation of intelligence.

The battle for verifiable data is increasingly defining the AI race.

This competition is reshaping the economics of information. We are seeing a move toward closed ecosystems where data is treated as a strategic asset rather than a public good. This has profound implications for the open web and for how companies like Google or OpenAI negotiate with publishers. The data war is not just about what a model knows, but about the integrity of its foundations.

Strategic Data Priorities

Securing non-synthetic, human-generated datasets
Developing methods for verifying data provenance
Building proprietary data moats through enterprise partnerships
Mitigating the risks of recursive training loops

In this new landscape, the winners will not necessarily be those with the most compute, but those with the most reliable map of the real world.

Key Takeaway

The next AI frontier is defined by the quality and provenance of data, not just scale.