The Deep Feed

01 — Lenny's Newsletter

The Review Bottleneck

How AI-generated code is breaking the pull request model

By Claire Vo · 12 min read

Editor's note: As AI writes more code, the human bottleneck has shifted from creation to verification.

The pull request queue is where engineering velocity goes to die. For decades, the bottleneck in software development was the act of writing code. It was a slow, deliberate process of thought and syntax. Now, that has changed. AI agents can generate functional, syntactically correct code in seconds. This creates a new, more dangerous problem: a flood of pull requests that no human team can possibly review. The senior engineers, who used to spend their time solving hard problems, are now drowning in a sea of automated changes. They are stuck in a cycle of endless verification, trying to spot the one subtle error in a hundred lines of AI-generated logic that will eventually crash the production environment.

The Failure of Manual Review

The traditional model of peer review assumes a manageable volume of changes. It relies on the idea that a human can look at a diff, understand the intent, and spot the risk. But when the volume of code increases by an order of magnitude, the quality of human review collapses. Engineers begin to skim. They approve changes because they lack the time to do anything else. This is how technical debt accumulates at an accelerated rate. We are building faster than we can think, creating a gap between our ability to ship and our ability to understand what we have actually shipped.

The bottleneck is no longer the act of writing; it is the act of verifying.

The solution is not to review everything more carefully. That is a mathematical impossibility. The solution is to stop treating all code as equal. We need to move toward a model of automated risk scoring. By categorising changes based on their potential impact, we can decide which ones require a human eye and which ones can be handled by an agent. This requires a shift in how we define trust in a software system. Trust should not be a blanket assumption; it should be a calculated metric based on the specific properties of a change.

The Six Dimensions of Risk

To automate the review process, you must first define what constitutes risk. A robust system does not just look for bugs; it looks for systemic impact. A sophisticated agent can score a pull request across several specific dimensions to determine its safety. This allows for a tiered approach to merging, where low-risk changes move through the pipeline automatically, and high-risk changes are escalated to human experts.

The Risk Scoring Framework

Blast Radius: How much of the system does this change touch?
Reversibility: How difficult is it to undo this change if it fails?
Data Security: Does this modification touch sensitive user information?
Ops Impact: Will this change affect system stability or latency?
Verification Gap: How much testing is required to prove this works?
Change Surface: How many files and lines of code are being altered?

This approach turns the review process into an audit rather than a manual inspection. As long as the process is queryable and follows a clear risk policy, it can remain compliant with standards like SOC 2. The goal is to clear the backlog by automating the mundane, leaving the humans to focus on the architectural decisions that actually matter.

Key Takeaway

Automate the low-risk verification to protect the high-value engineering time.

02 — Lenny's Newsletter

The Emergent Career

Why long-term planning is a myth for the successful

By Molly Graham · 15 min read

Editor's note: The obsession with the five-year plan creates anxiety; the science of emergence offers a better way.

The modern career advice industry is obsessed with the five-year plan. We are told to map out our ascent, to pick a destination, and to march toward it with grim determination. This approach assumes the world is static. It assumes that the opportunities of tomorrow will look like the goals of today. For many, this leads to a sense of failure when the path deviates. They feel broken because they lack a map. They actually need a compass.

The Myth of the Staircase

We are taught that careers should look like a set of stairs. You decide where you want to end up and you climb. When you look at successful people, their resumes often look like that from a distance. They have the C-level titles and the big-name companies. But if you look closer, the path is rarely a straight line. It is often a chaotic sequence of pivots, lucky breaks, and unexpected detours. The idea of a designed career is a retrospective illusion. We see the pattern after the fact, but the actual journey is far more unpredictable.

A career without a plan is not a failure; it is an opportunity for emergence.

This brings us to the concept of emergence. In biology, emergence occurs when individual parts follow simple rules to create a complex, intelligent whole. A murmuration of starlings has no choreographer. There is no leader bird directing the shape of the cloud. Instead, each bird follows three local rules: stay close to your neighbours, do not collide, and move in the same general direction. From these simple instructions, a breathtaking, shape-shifting entity appears. The complexity is not designed; it emerges from the interaction of the parts.

Principles Over Destinations

If you abandon the long-term plan, what do you use to guide you? The answer is principles. Instead of optimizing for a specific job title or salary, you optimize for a set of local rules. These might include following your energy, seeking out high-leverage problems, or maintaining a specific set of values. When you focus on the next step in front of you and adhere to your principles, you create the conditions for success to happen. You stop trying to force the world into a shape and start allowing the right opportunities to reveal themselves.

The Emergent Framework

Define local rules (principles) rather than global destinations.
Follow your energy to identify high-interest areas.
Focus on the immediate next step with high intensity.
Trust that consistent action within your rules creates complexity over time.

An emergent career is harder to manage than a planned one. It requires a high degree of faith and a tolerance for uncertainty. There will be stretches where you cannot see the destination. But that uncertainty is exactly what creates the room for the opportunities you could not have seen coming. The most meaningful parts of a career are rarely the ones you planned for; they are the ones that emerged from the life you actually lived.

Key Takeaway

Stop designing your destination and start refining your principles.

03 — Experimental History

The Incentive Trap

Why relying on rewards is a recipe for mediocrity

By Adam Mastroianni · 10 min read

Editor's note: The common diagnosis of 'bad incentives' ignores the most important human element: purpose.

We are obsessed with the idea that bad incentives cause bad behaviour. When a system fails, we immediately look for the reward structure. We say, 'People are doing the wrong thing because they are being rewarded for it.' It is a tidy, tautological explanation. It treats human beings like pigeons in a Skinner box, incapable of acting on anything other than a direct reward. But this way of thinking is fundamentally flawed. It assumes that incentives are something to be obeyed, rather than something to be transcended.

The Batting Helmet Metaphor

Think of incentives like a communal batting helmet in a local baseball league. The helmets are old, they smell of sweat, and they do not fit properly. Most kids use them because they are what is provided. They accept the discomfort and the poor fit because it is the path of least resistance. But the serious players bring their own helmets. They bring gear that is custom-fitted and high-quality. They do not rely on the provided system; they bring their own standards to the game. Incentives are the same. If you only act to satisfy the provided reward system, you are playing by someone else's rules.

Doing something simply because an arbitrary reward system gives you a treat is for pigeons, not people.

The most important part of human development is the acquisition of purpose. This is the process of refining your own sense of right and wrong and figuring out how you can be useful. Once you have a purpose, you no longer need to chase the 'lil treats' of prestige, money, or status. A sense of purpose is both prosocial and protective. It provides a stable core that remains intact even when the external incentive structures are broken or corrupt.

The Void of Purpose

When a person lacks a sense of purpose, they develop a void. This cavity in their personality is almost always filled by the very things they claim to despise: the pursuit of power, fame, and status. These are the people who are most insistent that the incentives must be obeyed. They must swear that the currency is real, because if it isn't, their entire life's work is a fraud. They are the ladder-climbers who reach the top only to find they have no convictions to hold them there.

Signs of an Incentive-Driven Life

Prioritising external validation over internal standards.
Viewing every decision through the lens of 'what is the reward?'
Reacting to systemic failures by demanding stricter rules rather than better purpose.
Lacking a consistent moral compass outside of professional requirements.

To avoid this trap, you must bring your own helmet. You must develop a set of internal standards that exist independently of the rewards offered by your job, your industry, or your society. The goal is to move from being a responder to incentives to being a person driven by a mission. Only then can you navigate a broken system without becoming part of the problem.

Key Takeaway

Purpose is the only thing that survives a broken incentive structure.

04 — Simon Willison

The Reasoning Interface

Moving beyond the text-in, text-out era of AI

By Simon Willison · 10 min read

Editor's note: The next generation of LLM tools will focus on transparency and tool-use rather than just chat.

The era of the simple chatbot is ending. For the last two years, our interaction with large language models has been a game of text-in, text-out. We type a prompt, and a black box returns a response. We have treated these models as magic oracles, hoping the right sequence of words would trigger the right answer. But the technology is moving toward something more transparent and more capable: reasoning traces and tool-use. We are moving from prompting to orchestration.

Seeing the Thought

One of the most significant shifts in recent LLM releases is the support for visible reasoning traces. Previously, when a model 'thought' through a problem, that process was hidden from the user. You only saw the final output. This made it difficult to debug errors or understand the logic behind a response. New interfaces now allow us to see the reasoning process as it happens. This transparency changes the relationship between the human and the machine. The model is no longer a black box; it is a collaborator whose logic can be inspected and corrected.

We are moving from talking to machines to supervising their thought processes.

This shift is essential for high-stakes applications. If an AI is helping you write code or analyse a legal document, you cannot afford to trust it blindly. You need to see the steps it took to reach its conclusion. By exposing the reasoning trace, we can identify where the logic diverged from reality. This turns the interaction from a leap of faith into a rigorous process of verification.

The Rise of the Agentic Toolset

Beyond reasoning, the next frontier is tool-use. We are seeing a move toward models that can interact with the world through server-side tools. This includes code execution environments, web search capabilities, and the ability to call external APIs. An LLM is no longer just a text generator; it is an agent that can use a calculator, browse the internet, or query a database to find the truth. The model becomes the brain, and the tools become its hands.

Key Capabilities of the New Interface

Reasoning Traces: Visibility into the model's internal logic.
Server-Side Tools: The ability to execute code and search the web.
Content-Addressable Logging: A Git-like approach to managing conversation history.
Streaming Events: Handling a mix of reasoning, text, and tool calls in real-time.

For developers, this means the workflow is changing. We are no longer just writing prompts; we are designing environments where models can act. The complexity of these interactions is increasing, but so is their utility. The goal is to build systems where the human provides the intent and the machine provides the execution, supported by a transparent and verifiable reasoning process.

Key Takeaway

The future of AI is not about better chat, but about better reasoning and tool integration.

05 — Stratechery

The Capex Arms Race

Why massive infrastructure spending is the new AI moat

By Stratechery · 8 min read

Editor's note: The massive capital expenditure of tech giants is a defensive necessity in the race for the frontier.

The recent earnings reports from Google and Amazon have revealed a staggering truth: the cost of staying relevant in the AI era is measured in tens of billions of dollars. To the casual observer, the massive capital expenditure (capex) looks like waste. It looks like a spending spree driven by hype. But for the leaders of these companies, this spending is a defensive necessity. It is the price of admission to the frontier.

The Frontier Hedge

The race for artificial intelligence is not just a race for better algorithms; it is a race for the physical capacity to compute. The models that will define the next decade require an amount of silicon and electricity that only the largest corporations can provide. This creates a 'frontier hedge.' Companies like Google and Amazon are investing heavily in infrastructure today to ensure they are not left behind when the next leap in model capability occurs. They are betting that the cost of being too small will be far higher than the cost of being too big.

In the AI era, capital is the primary weapon of competition.

This investment is closely tied to the rise of frontier model providers like Anthropic. The relationship between the cloud providers and the model builders is symbiotic. The cloud providers provide the massive compute required to train the models, and in return, they secure a massive, permanent customer in the model builders themselves. This creates a virtuous cycle of spending that reinforces the dominance of the incumbents.

The Infrastructure Moat

As the barrier to entry rises, the moat around the big tech companies deepens. It is no longer enough to have the best engineers; you must also have the best data centres, the most efficient power agreements, and the most advanced custom silicon. This shifts the competitive landscape from one of pure software innovation to one of industrial-scale resource management. The winners will be those who can most effectively turn capital into compute.

Drivers of AI Capex

Training the next generation of frontier models.
Scaling inference capacity for millions of users.
Developing custom AI-specific silicon (ASICs).
Securing energy and data centre real estate.

We are witnessing the industrialisation of intelligence. The era of the garage startup building the next great AI is being replaced by an era of massive, capital-intensive infrastructure projects. The question is not whether the spending is justified, but who will be able to sustain it long enough to reach the next level of capability.

Key Takeaway

Infrastructure is the new battlefield of the AI era.

06 — Stratechery

The Efficiency Payoff

How Microsoft is turning AI hype into bottom-line reality

By Stratechery · 8 min read

Editor's note: Microsoft's strategy focuses on application and efficiency rather than pure R&D.

While much of the tech industry is caught in a debate about the long-term value of AI, Microsoft is quietly reaping the efficiency payoff. Their recent earnings demonstrate a clarity of strategy that sets them apart from their peers. They are not just building models; they are embedding them into the workflows that drive the global economy. This is the difference between a research project and a product.

Strategy vs. Hype

The distinction between Microsoft and companies like Meta lies in their application of the technology. Meta's AI efforts are often focused on improving engagement and advertising efficiency—essential, but secondary to the core product. Microsoft, however, is integrating AI into the very fabric of productivity. By embedding AI into Office, Azure, and Windows, they are making the technology indispensable to the enterprise. They are not just selling a new feature; they are selling a new way of working.

Real value comes from application, not just innovation.

This approach allows Microsoft to demonstrate a tangible return on investment. When a company adopts Microsoft's AI-enabled tools, they see immediate gains in efficiency and output. This creates a feedback loop of adoption and revenue that is much more stable than the speculative hype surrounding pure-play AI companies. Microsoft is turning the promise of AI into a measurable economic reality.

The Implementation Edge

The ultimate winner in the AI era will not necessarily be the company with the most advanced model. It will be the company that can most effectively implement that model into existing, high-value workflows. Microsoft has a massive advantage here: they already own the workflows. They own the desktop, the cloud, and the collaboration tools. They do not need to find new customers; they only need to make their existing customers more productive.

Microsoft's Competitive Advantages

Deep integration into existing enterprise workflows.
A massive, established customer base.
The ability to turn AI capabilities into subscription revenue.
A focus on practical application over speculative R&D.

The lesson for the rest of the industry is clear: innovation is only half the battle. The other half is integration. The companies that will dominate the next decade are those that can take the raw power of AI and turn it into a seamless, indispensable part of the human experience.

Key Takeaway

The winner of the AI race is the one who integrates it best, not the one who builds it best.