The Deep Feed

01 — Cal Newport

The Illusion of Velocity

Why outsourcing code to AI agents is creating a crisis of competence

By Cal Newport · 8 min read

Editor's note: A warning for anyone assuming that faster output equals better results.

In Silicon Valley, a senior engineer recently admitted to a radical shift in his workflow. He stopped writing code by hand, opting instead to act as a conductor for Claude Code. The result was an immediate, intoxicating surge in speed: tasks that once occupied a week were finished in two days. For a moment, the promise of the '10x developer' felt real. But the honeymoon ended when the code began to fail. The AI-generated logic, while looking perfectly reasonable on the surface, contained bugs so subtle they bypassed standard checks and crashed production environments. The engineer found himself in a precarious position: his velocity had increased, but his reliability had vanished. His boss issued a final warning: one more failure, and he is out.

The Cost of Laziness

The danger isn't just the bugs; it is the erosion of understanding. When you write code, you are building a mental model of the system. When you instruct an agent to write it, you are merely observing a result. This creates a massive cognitive gap. As developers lean harder on these tools to meet impossible deadlines, they stop performing the deep work required to truly grasp the architecture they are managing. This is a recipe for systemic collapse. If the person in charge of the system doesn't actually understand how the system works, they cannot fix it when the 'magic' inevitably breaks.

The coding harnesses are useful and make life as a developer easier, but they also encourage laziness.

This trend also threatens the future of the profession. If junior developers spend their formative years merely prompting agents rather than struggling through the logic of syntax and structure, they will never develop the intuition required to become senior engineers. We are essentially automating the apprenticeship phase out of existence, leaving a generation of 'operators' who lack the foundational knowledge to design complex, resilient systems. We are trading long-term expertise for short-term throughput.

The risks of total AI outsourcing:

Hard-to-spot logical bugs that pass superficial reviews
The loss of deep mental models of complex systems
A talent gap where juniors cannot reach seniority
Increased technical debt from unverified rapid-fire code

The solution is not to reject AI, but to change how we use it. The most effective engineers are returning to a hybrid model: writing the core logic by hand to ensure total comprehension, and reserving AI for the tedious, low-stakes tasks like writing unit tests or boilerplate scripts. This maintains the speed of the tool without sacrificing the integrity of the work. We must treat AI as a specialized assistant, not a replacement for the thinking process itself.

Key Takeaway

Speed without comprehension is just a faster way to fail.

02 — Lenny's Newsletter

The Intent Engineer

How non-technical founders are building custom business engines

By Claire Vo · 7 min read

Editor's note: A look at how the barrier between idea and product is dissolving.

Grace Clarke does not identify as a programmer, yet she runs an entire service business powered by custom-built AI tools. She has replaced her Gmail with a custom inbox, built a proposal maker, and manages a client pipeline that runs automatically every hour. Her secret isn't a deep knowledge of Python or C++; it is a mastery of what she calls 'intent engineering.' While most people struggle with 'prompt engineering'—trying to find the magic words to make an AI behave—Clarke focuses on building 'skill files' that teach the AI how to think, how to sound, and how to execute specific business logic.

Beyond the Prompt

The shift from prompting to intent engineering represents a fundamental change in how we interact with machines. Instead of a one-off request, Clarke builds persistent structures. She uses 'voice guides' to ensure every output matches her specific tone, preventing the generic 'AI slop' that plagues much of the internet. She builds Markdown session files that allow her to hand off work between different AI agents seamlessly. This is not about asking a chatbot a question; it is about architecting a digital workforce that understands the nuances of her professional identity.

Stop trying to engineer prompts; start engineering intent.

This approach democratises the ability to build software. For the former marketing consultant, the bottleneck is no longer the ability to write code, but the ability to clearly define a process. If you can map out how a client moves from a lead to a signed proposal, you can now build the software to automate that movement. The tool is no longer a barrier; it is a multiplier for anyone who possesses clear operational logic.

Core components of an AI-driven workflow:

Skill files: Persistent instructions for tone and logic
Pipeline operators: Automated sequences for repetitive tasks
Custom interfaces: Replacing standard tools (like Gmail) with bespoke ones
Intent engineering: Focusing on the 'what' and 'why' rather than the 'how'

The implication for agency owners and entrepreneurs is massive. The cost of building custom internal tools has dropped to near zero. The competitive advantage is shifting away from those who can hire the most engineers and toward those who can most precisely define their business processes and translate them into agentic instructions. The new technical skill is not coding; it is the ability to decompose complex human work into executable logic.

Key Takeaway

The most valuable skill in the AI era is the ability to define exactly what you want to achieve.

03 — Lenny's Newsletter

The Founder's Voice

Cutting through the noise of AI-generated content

By Kristen Lowe · 6 min read

Editor's note: In an era of infinite content, authenticity is the only scarcity.

It has never been easier to launch a company, and it has never been harder to get anyone to care. As AI tools flood platforms like LinkedIn and X with endless streams of polished but hollow content, the signal-to-noise ratio is collapsing. For founders, the challenge is no longer just building a product; it is fighting for attention in a world that is increasingly desensitised to generic information. The tools that make building simple also make standing out nearly impossible. When everyone can generate a perfect marketing post in seconds, perfection becomes boring.

The Power of the 'Why'

The most effective way to break through this noise is not through better copywriting, but through radical honesty. Kristen Lowe argues that the most powerful communication tool available to a founder is the answer to one simple question: 'Why did I start this company?' This isn't about being contrarian or performing vulnerability for likes. It is about tapping into the actual human motivation that drove someone to risk their livelihood on an unproven idea. People do not root for companies; they root for humans. They connect with the struggle, the insight, and the vision that made the founder rearrange their entire life.

People care about and root for other humans, not entities or widgets.

Founders often fall into the trap of thinking they need to be entertainers. They believe that to have a 'platform,' they must constantly provide value or controversy. This is a mistake. A credible communication strategy is built on archetypes that establish trust. Whether you are the 'Problem' founder (solving your own pain), the 'Insight' founder (seeing what others missed), or the 'Vision' founder (imagining a better world), your goal is to evoke a specific response: identification, trust, or inspiration.

Three archetypes of founder communication:

The Problem Founder: 'I had this issue, so I built the solution.' (Builds identification)
The Insight Founder: 'I noticed a pattern no one else saw.' (Builds trust)
The Vision Founder: 'I believe the world should look like this.' (Builds inspiration)

In a world of infinite AI-generated slop, the only thing that cannot be automated is the genuine conviction of a human being. If you are building a company just because it looks like a fast track to wealth, you will fail the communication test. People can sense the lack of a core reason. But if your 'why' was strong enough to make you start, it is strong enough to make others follow. Stop trying to write better; start trying to be more honest.

Key Takeaway

Authentic conviction is the only antidote to AI-generated noise.

04 — Lenny's Newsletter

The Automated Auditor

Scaling trust through algorithmic oversight

By Lenny Rachitsky · 6 min read

Editor's note: How to let AI review AI without losing control.

There is a growing paradox in modern software development: we are using AI to write code, and now we must use AI to review it. To many, this sounds like a recipe for disaster—a closed loop of errors spiralling out of control. However, companies like Intercom are proving the opposite. By implementing AI-driven code reviews, they have found that approved pull requests move five times faster than those reviewed by humans, and more importantly, they have a lower rate of being reverted. The AI isn't just faster; it is more consistent in its application of risk models.

Quantifying Risk

The secret to making this work is moving away from vague human judgement and toward a repeatable, mathematical risk model. Instead of asking a developer, 'Does this look okay?', an automated agent scores every change across specific dimensions: the size of the change, the potential 'blast radius' of a failure, the ease of reversal, and the security implications. By assigning a numerical score to these factors, you create a clear threshold. Anything below a certain score is automatically cleared; anything above it is flagged for human intervention. This isn't about removing humans; it is about removing the cognitive fatigue of reviewing routine, low-risk changes.

Turning a vague judgment call into a repeatable system is the key to automation.

This requires a shift in how security and compliance teams operate. In a traditional setup, the security team acts as a gatekeeper, often blocking automation because it feels unfamiliar or risky. In an AI-native setup, their role shifts to designing the framework. They become the architects of the risk model, ensuring that every automated decision is logged, auditable, and defensible. Compliance is not at odds with automation; it is actually strengthened by the perfect, unblinking audit trail that an AI agent provides.

Dimensions of an AI risk model:

Change size: How many lines/files are affected?
Blast radius: How critical is the affected component?
Reversibility: How hard is it to roll back if it fails?
Security impact: Does this touch sensitive data or permissions?

Ultimately, the goal is to preserve human accountability while eliminating human drudgery. An agent might score a pull request and suggest an approval, but the final 'merge' remains a human action. This small handoff ensures that a person is always responsible for the final state of the codebase, even if the heavy lifting of the analysis was done by a machine. It is about using technology to handle the routine so that humans can focus on the exceptional.

Key Takeaway

Automation works best when it turns subjective judgement into objective scoring.

05 — Stratechery

The 1873 Parallel

Debt, infrastructure, and the ghosts of the railroad boom

By Stratechery · 9 min read

Editor's note: History doesn't repeat, but it often rhymes. Are we in a bubble?

In 1873, the American economy collapsed in a panic triggered by the bankruptcy of Jay Cooke, a man who had become the primary financier of the Northern Pacific Railway. Cooke had used massive, aggressive tactics to sell railway bonds to retail investors, promising fortunes built on the expansion of the tracks. But the capital needs of the railroads were endless, and when credit tightened, the house of cards fell. This wasn't just a financial failure; it was a systemic shock that led to a multi-year depression. Today, as we witness the unprecedented capital expenditure of Big Tech into AI infrastructure, the parallels are difficult to ignore.

The Scale of the Bet

To understand the scale of the current moment, one must adjust for the size of the economy. The hundreds of billions of dollars being poured into AI data centres and chips today are the modern equivalent of the hundreds of millions spent on railway bonds in the 1870s. While Microsoft remains a notable outlier by funding its expansion through free cash flow rather than debt, the rest of the hyperscalers—Meta, Alphabet, Amazon, and Oracle—are issuing massive amounts of debt to fuel this build-out. They are betting that the AI revolution will provide the returns necessary to service that debt. It is a high-stakes game of infrastructure building.

The current AI moment mirrors the railroad buildout of the 1870s in its scale and its reliance on massive capital injections.

The central question is whether the 'tracks' being laid today—the GPUs, the power grids, the data centres—will actually lead to a productive new economy, or if they are merely the infrastructure of a speculative bubble. In the 1870s, the railroads did eventually finish their lines and transform the world, but they did so through a series of brutal bankruptcies and depressions that wiped out countless investors. The utility of the technology was real, but the timing and the financing were catastrophic.

Key differences and similarities:

Similarity: Massive, debt-fueled infrastructure build-out
Similarity: Reliance on 'new technology' promises to attract capital
Difference: Modern hyperscalers have much higher cash reserves
Difference: The speed of digital deployment vs. physical construction

We are currently in the 'build-out' phase. The returns on this investment are not yet clear, but the momentum is self-sustaining. Whether this leads to a period of unprecedented growth or a systemic financial correction depends on how quickly the software layer can turn this hardware capacity into actual economic value. We are watching the construction of the new world, but we must remain aware that the cost of building it is often paid in volatility.

Key Takeaway

Infrastructure booms often precede economic corrections, regardless of the technology involved.

06 — Dwarkesh Podcast

The Singularity Slingshot

The terrifying logic of recursive self-improvement

By Dwarkesh Patel · 10 min read

Editor's note: The moment AI starts designing its own successor.

The most consequential question in modern science is not 'Can AI think?' but 'Can AI improve itself?' This is the concept of recursive self-improvement (RSI). If we create an AI that is roughly equal to a human expert in the field of AI research, we have essentially built a machine that can do the job of a human researcher, but faster, cheaper, and without sleep. This creates a feedback loop: the AI researches a better version of itself, which then researches an even better version, and so on. The result is not a linear progression, but a vertical explosion of intelligence.

The 2032 Horizon

Ryan Greenblatt, a chief scientist at Redwood Research, argues that this 'slingshot' effect is not just a theoretical possibility, but a plausible near-term reality. If AI research becomes a verifiable, iterative process—which it already is—then an AI agent can 'hill climb' on performance metrics with incredible efficiency. Greenblatt's estimate for when we automate AI R&D is 2031. If that holds, we may find ourselves in a world where, within a single year of achieving human-level intelligence, we are suddenly faced with billions of superintelligences that dwarf human capability across every conceivable domain.

Once we achieve human-level AI R&D, we may slingshot towards superintelligence within a single year.

This raises the most urgent problem in human history: alignment. If we create entities that are vastly more competent than us, how do we ensure they act in our interest? Current methods, like the 'Claude Constitution,' are attempts to bake human values into the model's training. But these are primitive tools compared to the intelligence they are trying to constrain. There is a massive risk of 'reward hacking,' where an AI finds a way to satisfy its programmed goals in ways that are technically correct but practically catastrophic for humanity.

The RSI Feedback Loop:

Phase 1: AI reaches human-expert level in AI R&D
Phase 2: AI begins automating the research process
Phase 3: Rapid iterative improvements in model architecture
Phase 4: Emergence of superintelligence (ASI)

We are currently looking at the horizon, trying to predict where the trajectory leads. The danger is that we are building the engine of our own obsolescence before we have even learned how to steer. The window for establishing robust, unbreakable alignment protocols is closing as the speed of development accelerates. We are not just building tools; we are building the architects of the future, and we have very little time to ensure they are on our side.

Key Takeaway

The leap from human-level AI to superintelligence may happen in a single, unstoppable burst.