The Deep Feed

01 — Lenny's Newsletter

The Death of the Text Spec

Why HTML is replacing Markdown in the age of AI agents

By Thariq Shihipar · 12 min read

Editor's note: As AI agents take over more of the heavy lifting, the way we instruct them is shifting from simple text to rich, interactive interfaces.

The era of the thousand-line Markdown file is ending. For years, Markdown has been the standard for technical documentation and planning because it is easy for both humans and machines to parse. But as AI agents like Claude Code begin to handle more complex, multi-stage engineering tasks, Markdown is proving to be too flat. It lacks the density and interactivity required to manage the sheer volume of information an agent generates. When a plan spans thousands of lines of text, a human engineer stops reading and starts skimming. This is where the connection breaks. The engineer loses sight of the logic, and the agent drifts from the intended goal.

The HTML Advantage

Anthropic engineers are finding a better way: HTML. By using HTML as the primary format for planning and communication, they are turning static specs into living artifacts. HTML allows for scrollable sections, visual mockups, and interactive elements that make information much easier to digest. Instead of a wall of text, an engineer receives a dashboard. This isn't just about aesthetics; it is about information density. HTML allows an agent to present a plan that includes code snippets, visual representations of a UI, and data tables in a way that a human can actually engage with. It makes the planning phase a visual experience rather than a reading chore.

99% of your AI-generated tokens should go to planning and interfaces, not production code.

This shift represents a massive reallocation of compute. In the old model, we used tokens to write the final code. In the new model, the vast majority of tokens are spent on the 'meta-work': the dashboards, the custom interfaces, and the interactive plans that ensure the human stays in the loop. We are building 'micro-software' on top of our main software. If a specific part of a plan is too complex to read, the engineer can ask the agent to build a throwaway UI just to edit that one section. This tool exists for ten minutes, solves the problem, and is then discarded. It is a highly efficient way to manage complexity without being constrained by rigid templates.

The HTML Workflow

Brainstorming via visual mockups rather than text lists
Using interactive HTML tables to edit complex data rules
Maintaining living design systems that travel within the repo
Generating status updates that are actually readable by managers

The goal is to make the work legible enough that you actually want to participate in it. When the interface is rich, the human remains a director rather than a mere spectator. This is the future of the developer experience: moving away from the terminal and toward a high-fidelity command centre where the agent does the typing, but the human manages the vision through interactive, visual feedback loops.

Key Takeaway

Rich, interactive interfaces are more effective than plain text for managing the complexity of AI-driven workflows.

02 — Lenny's Newsletter

The Silicon Body

Why the next frontier of AI is physical, not just digital

By Caitlin Kalinowski · 10 min read

Editor's note: The transition from LLMs to robotics is the most significant hardware challenge of the decade.

The current obsession with large language models has largely ignored the physical reality of intelligence. We have built incredibly sophisticated brains, but they are currently trapped in boxes. The next phase of the AI boom is not about better chat interfaces, but about embodied intelligence—the marriage of advanced models with sophisticated hardware. This is the transition from software that thinks to robots that act. While the industry has spent years perfecting the digital side, the hardware side is only just beginning to catch up.

Lessons from the VR Era

The failure of virtual reality to achieve mass-market dominance provides a vital lesson for the robotics industry. VR struggled with the friction of the interface—the weight of the headset, the discomfort of the lenses, and the isolation from the real world. Robotics faces similar hurdles. A humanoid robot that is too heavy, too slow, or too expensive to maintain will remain a novelty. However, the technologies developed for VR—computer vision, spatial mapping, and low-latency sensor integration—are the exact foundations required for modern robotics. The tools built for the virtual world are now being repurposed to navigate the physical one.

The real AI boom hasn't even started; we are still just teaching the machines how to see.

One of the most immediate bottlenecks is the supply chain, specifically memory. As models become more capable, the demand for high-speed, high-capacity memory is set to skyrocket. We are approaching a price shock. Startups in the robotics space are already being advised to pre-buy memory components to avoid being priced out of the market as the demand from larger players becomes unavoidable. This is a classic hardware race where the winners will be determined by their ability to secure the physical components of intelligence before the rest of the world wakes up.

Hardware Bottlenecks

Memory price volatility and supply shortages
The difficulty of mass-producing humanoid actuators
The energy requirements for real-time spatial reasoning
The transition from digital models to world models

We are moving toward a world where AI is not something you look at on a screen, but something that moves through your environment. This requires a fundamental shift in how we think about engineering. It is no longer enough to optimize for latency and accuracy in a cloud environment; we must now optimize for torque, battery life, and physical durability. The intelligence is becoming embodied, and the hardware is the new battlefield.

Key Takeaway

The next stage of AI development depends on solving the physical constraints of robotics and the supply chain of specialized hardware.

03 — Cal Newport

The Efficiency Trap

Understanding the Jevons Paradox in the age of automation

By Cal Newport · 8 min read

Editor's note: Efficiency does not always lead to less work; often, it simply creates more of it.

In 1865, economist William Stanley Jevons observed something counterintuitive about the industrial revolution. As steam engines became more efficient and used less coal, the total consumption of coal did not drop. Instead, it exploded. Because coal was now cheaper and more effective, people found more ways to use it. This is the Jevons Paradox: increasing the efficiency of a resource often leads to a massive increase in its consumption. As we apply this logic to AI, we see a pattern that challenges the fear of job loss.

The Labor Market Shift

The common fear is that if one programmer can do the work of five using AI, companies will simply hire fewer programmers. But the Jevons Paradox suggests the opposite. If the cost of software development drops significantly, the demand for software will skyrocket. Every small business, every local government, and every individual will want custom tools. The total volume of software being built will grow so much that the demand for engineers—even those managing AI—will likely increase. We aren't shrinking the pie; we are making the pie so cheap that everyone wants a slice.

Efficiency is not a zero-sum game; it is a demand multiplier.

However, there is a darker side to this phenomenon. When we make communication more efficient, we don't spend less time communicating; we just communicate more. The arrival of email and Slack created a world where the average knowledge worker is interrupted every two minutes. We have become more efficient at sending messages, which has led to a total collapse of deep work. AI presents the same risk. If an AI can generate a report in seconds, we will expect ten times as many reports. We risk drowning in a sea of hyper-efficient, low-value communication.

The Risks of Hyper-Efficiency

The explosion of low-value digital noise
The erosion of deep, focused work
The constant pressure for instant responses
The increased cognitive load of managing more tools

We must be careful not to mistake efficiency for productivity. A person who responds to 100 emails an hour is efficient, but they are not necessarily doing anything of value. As AI accelerates the pace of work, our challenge will be to ensure that we are using our newfound efficiency to do better work, rather than just doing more trivial work at a faster rate.

Key Takeaway

Technological efficiency often triggers an explosion in demand, which can lead to increased productivity or a descent into constant, shallow activity.

04 — Lenny's Newsletter

The Compute Allocator

How the role of the engineer is being redefined

By Thariq Shihipar · 7 min read

Editor's note: The value of an engineer is shifting from the ability to write syntax to the ability to manage resources.

For decades, the measure of a great engineer was their mastery of syntax and their ability to write clean, efficient code. But as AI agents become capable of generating vast amounts of production-ready code, that skill is becoming commodified. The new essential skill is not writing code, but deciding what is worth building. We are seeing the emergence of a new professional archetype: the compute allocator. In this role, the engineer acts more like a fund manager, deciding how to spend a budget of tokens and processing power to achieve a specific outcome.

Managing the $500 Task

When an AI agent can run for several hours on a single complex task, you are no longer just managing time; you are managing a significant financial cost. Deciding whether to let an agent iterate on a problem for eight hours is a decision about how to spend hundreds of dollars in compute. This requires a different kind of discipline. The engineer must define the boundaries of the task, set the constraints, and know exactly when to intervene. The goal is to maximize the return on compute, ensuring that the agent's output is not just correct, but strategically valuable.

The engineer of the future is a strategist, not a typist.

This shift places a premium on architectural thinking and problem definition. If you give an agent a poorly defined task, it will spend a large amount of money producing the wrong thing. The ability to decompose a complex business problem into a series of executable, constrained instructions is the new core competency. You are no longer building the machine; you are directing the machine that builds the machine.

The Allocator's Skillset

Precise problem decomposition
Strategic constraint setting
Cost-benefit analysis of compute spend
High-level architectural oversight

As the cost of generating code approaches zero, the value of the ideas behind the code approaches infinity. The engineer's job is to ensure that the abundance of compute is directed toward solving real problems, rather than just generating more noise. The future belongs to those who can navigate the tension between agent autonomy and human intent.

Key Takeaway

As AI handles the syntax, the engineer's primary value shifts to strategic decision-making and the efficient allocation of compute resources.

05 — Stratechery

The Infrastructure War

The social and physical friction of the AI boom

By Ben Thompson · 6 min read

Editor's note: The massive expansion of data centres is hitting a wall of local opposition.

The expansion of AI requires a physical footprint that the world is not entirely prepared to accommodate. Data centres are massive, energy-hungry, and often perceived as parasitic by the communities they inhabit. They consume vast amounts of electricity, strain local grids, and occupy large tracts of land, often without providing the kind of visible, community-facing benefits that traditional industries might. This has led to a growing wave of discontent, as local governments and citizens push back against the rapid deployment of AI infrastructure.

The Cost of Proximity

The opposition to data centres is not merely a matter of zoning laws; it is a fundamental clash of interests. On one side, you have the global tech giants who need massive amounts of compute to stay competitive. On the other, you have local communities concerned about their rising energy bills, the stability of their power grids, and the environmental impact of large-scale cooling systems. This tension is creating a new kind of regulatory friction that could slow the pace of AI development if not addressed through more sophisticated engagement strategies.

The solution to infrastructure opposition is rarely technical; it is almost always financial.

Solving this problem requires moving beyond simple land-use agreements. Companies must find ways to make the presence of a data centre a net positive for the local area. This might mean investing in local renewable energy projects that stabilize the grid for everyone, or providing direct economic benefits that go beyond mere tax revenue. The era of 'build it and they will come' is over. The era of 'negotiate and integrate' has begun.

Strategies for Infrastructure Integration

Direct investment in local energy resilience
Transparent communication regarding resource usage
Creating tangible community-facing value
Proactive engagement with local regulatory bodies

The physical reality of AI is that it cannot exist in a vacuum. It requires land, power, and water. As the demand for these resources grows, the companies that succeed will be those that treat local communities as stakeholders rather than obstacles. The infrastructure war will be won by those who can integrate their needs into the existing social and physical fabric of the world.

Key Takeaway

The rapid expansion of AI infrastructure faces significant social and regulatory pushback that can only be resolved through direct local investment and integration.

06 — Simon Willison

The Open Source Retreat

Security, privacy, and the danger of closing the gates

By Simon Willison · 5 min read

Editor's note: The NHS's decision to restrict open source access highlights a growing tension in public sector security.

In the pursuit of security, it is easy to fall into the trap of thinking that secrecy is the same as safety. This is the logic currently driving a retreat from open source within parts of the UK public sector, most notably the NHS. In response to reported vulnerabilities, the decision was made to close access to certain open source repositories. While the intention is to protect sensitive code, the move ignores a fundamental principle of modern cybersecurity: openness allows for scrutiny, and scrutiny leads to better security.

The Illusion of Secrecy

Closing a repository does not make the vulnerabilities within it disappear; it only makes them harder for the good actors to find. When code is open, a global community of researchers and engineers can identify and fix flaws before they are exploited. When code is private, vulnerabilities can remain hidden for years, known only to attackers. The Government Digital Service (GDS) has argued that openness should remain the default, noting that making everything private adds significant delivery costs and reduces the ability for other agencies to reuse and improve existing work.

Openness is a security posture, not a liability.

The culture of the civil service often struggles with this concept. There is a tendency to view transparency as a risk rather than an asset. This is reflected in the internal friction seen when disagreements over policy spill into the public eye. The decision to retreat from open source is a regressive step that increases the long-term cost of maintaining public infrastructure and reduces the overall resilience of the system.

The Costs of Closing Code

Reduced ability for cross-agency reuse
Higher long-term maintenance and security costs
Loss of community-driven bug fixes
Increased risk of hidden, unpatched vulnerabilities

For the public sector to build truly resilient digital services, it must embrace the benefits of the open source ecosystem. Security is not found in the darkness of a private repository, but in the light of public scrutiny. The move toward secrecy is a short-term fix that creates long-term fragility.

Key Takeaway

Retreating from open source in the name of security often increases vulnerability by reducing the community scrutiny necessary to find and fix flaws.