The Deep Feed

01 — Lenny's Newsletter

The Taste Gap: Why Logic is No Longer the AI Frontier

Moving beyond benchmarks to the reality of model utility

By Claire Vo · 12 min read

Editor's note: Benchmarks are lying to you. The real winner in the AI race is the model that understands human intent, not just syntax.

The industry has become obsessed with the wrong metrics. We track parameter counts, token speeds, and mathematical accuracy as if these were the only indicators of value. But for a product manager or an agency owner, a model that is technically perfect but stylistically useless is a liability. This is the central tension in the latest release of OpenAI's GPT-5.6 Sol. While competitors like Claude Fable 5 are pushing the boundaries of precision and pedantry, Sol is winning on a different metric entirely: taste. In a recent benchmark, Sol outperformed its rivals not because it knew more facts, but because it understood how to build things that actually work for humans.

The Pedantry Problem

Claude Fable 5 represents a specific kind of excellence that is increasingly becoming a trap. It is precise. It is correct. It is, quite frankly, exhausting to work with. When tasked with collaborative design or PRD writing, Fable often gets stuck in a loop of correction, treating every minor deviation from a standard format as an error to be fixed. This pedantry creates a friction that slows down the creative process. If you are trying to move from a rough idea to a functional prototype, you do not need a critic; you need a partner. Sol breaks through this wall by knowing when to ignore the rules in favour of the objective.

A model that is technically perfect but stylistically useless is a liability.

This distinction is best captured by the 'Claire Weighted Index', a benchmark that prioritises human preference and utility over raw computational correctness. In this test, Sol's ability to handle complex, multi-step tasks—like building a gamified app in a single shot—made it the clear victor. It wasn't just about the code being functional; it was about the code being intuitive. Sol understands the 'vibe' of a product, a quality that is notoriously difficult to quantify but impossible to ignore when you are actually using the tool to build a business.

From Prompting to Automation

The true power of these models is not found in a chat interface, but in their ability to act as agents. We are seeing a shift from 'talking to AI' to 'deploying AI'. For example, using Codex alongside Chrome allows for browser automation that can handle hundreds of repetitive tasks—like LinkedIn outreach—while the user does nothing. This is the end of the 'prompt engineering' era and the beginning of the 'agentic automation' era. The value is no longer in how well you can phrase a request, but in how effectively you can integrate a model into a workflow that executes on your behalf.

Key Model Differentiators

Sol: High utility, high taste, superior for rapid prototyping.
Fable: High precision, high pedantry, better for strict compliance.
Sonnet 5: The current leader for agentic voice and conversational fluidity.

As we move forward, the competitive advantage for businesses will not be access to the smartest model, but the ability to deploy the most useful one. If your team is spending more time correcting an AI's pedantic errors than they are shipping products, you are using the wrong tool. The winner of the AI race will be the one that feels less like a calculator and more like a colleague.

Key Takeaway

Stop optimizing for accuracy and start optimizing for utility; a perfect model that is hard to collaborate with is a net loss.

02 — Lenny's Newsletter

The End of Prompting: Building the AI Harness

Why the future of agency lies in structured workflows, not chat boxes

By Claire Vo · 10 min read

Editor's note: If you are still typing 'please fix this bug' into a chat box, you are wasting your time. It is time to build harnesses.

There is a common refrain in the AI community: 'It is not the model, it is the harness.' Yet, few people actually define what that means. Most users are still stuck in the 'chat' paradigm, treating AI like a person they are sending emails to. They prompt, they wait, they correct, and they repeat. This is an inefficient way to run a business. A harness is not a conversation; it is a structured, automated environment that wraps an AI model in specific permissions, tools, and objectives to perform a repetitive task without human intervention.

The Architecture of Agency

To build a functional harness, you must move beyond the single prompt. A professional-grade harness, such as the Sentry bug-debugging tool built using the Claude Agent SDK, requires four distinct components: runs, tasks, tools, and artifacts. A 'run' is the execution of a workflow. A 'task' is the specific objective within that run. 'Tools' are the external APIs the agent can call—such as GitHub, Linear, or Vercel—and 'artifacts' are the tangible outputs, like a pull request or a bug report, that the agent leaves behind. This structure transforms the AI from a chatbot into a digital employee.

A harness transforms the AI from a chatbot into a digital employee.

The goal is to eliminate the need for 'dear agent, please fix this'. In a properly constructed harness, the trigger is external. A bug is reported in Sentry; the harness detects it, gathers the evidence, performs a root-cause analysis, and creates the necessary artifacts in GitHub and Linear. The human only enters the loop to approve the final result. This is how you scale expertise. You are not scaling your ability to prompt; you are scaling your ability to architect workflows.

When to Build vs. When to Buy

The temptation is to use general-purpose tools like Claude Code or Codex for everything. While these are excellent for ad-hoc tasks, they are not harnesses. A general tool requires a human to provide the context and the direction every single time. You should build a custom harness when a workflow is repetitive, structured, and requires access to your specific business data or internal tools. If you find yourself performing the same sequence of five actions every Tuesday, you should not be prompting; you should be engineering a harness.

The Harness Checklist

Does the task have a clear, repeatable trigger?
Does the agent require specific access to external tools (GitHub, Sentry, etc.)?
Can the output be formatted as a structured artifact for the team?
Does the workflow require a 'human-in-the-loop' for final approval?

Building these systems requires a shift in mindset from user to architect. You are no longer managing a conversation; you are managing a system. The complexity of the build is offset by the massive reduction in cognitive load. Once the harness is running, the work happens in the background, and your role shifts from doing the task to auditing the system that does the task.

Key Takeaway

Stop treating AI as a conversational partner and start treating it as a component in a structured, automated system.

03 — Lenny's Newsletter

The Generalist's Revenge

How AI is collapsing the traditional product team

By Lenny Rachitsky · 15 min read

Editor's note: The era of the hyper-specialist is ending. In its place, a new breed of generalist operator is rising.

For decades, the canonical product team structure has been a collection of specialists: the PM, the designer, the data scientist, and the researcher. Each held a distinct territory of expertise. But as AI begins to handle the heavy lifting of data analysis, wireframing, and even basic coding, those boundaries are dissolving. We are seeing the rise of the 'product staff'—a lean, highly capable generalist who can navigate the entire lifecycle of a feature. This is not just a change in headcount; it is a fundamental shift in how products are built.

The Death of the Specialist

Adam Mosseri, who leads Instagram, has watched this evolution from the inside. As functional boundaries blur, the traditional 'baker's dozen' of specialist teams is being replaced by small, agile pods of four to six generalists. In this new model, a single person might handle the product requirements, the initial design mocks, and the data tracking setup. AI acts as a force multiplier, providing the technical scaffolding that previously required a dedicated specialist. This makes the individual more powerful, but it also makes the traditional specialist role increasingly precarious.

The boundaries between PM, design, and data are dissolving into a single role: the generalist operator.

This shift does not mean designers are obsolete. On the contrary, it means the value of design is moving up the stack. When AI can generate a thousand variations of a button or a layout, the designer's job is no longer to draw the button, but to decide why the button should exist and how it fits into the broader user psychology. The skill is moving from execution to intent. The winners in this new era will be those who can combine high-level strategic thinking with the ability to use AI to execute the tactical details.

Authenticity in a Synthetic World

As AI-generated content becomes ubiquitous, a new tension emerges: the battle for authenticity. On platforms like Instagram, the rise of synthetic media could easily lead to a sea of mediocrity. However, Mosseri suggests that AI might actually act as a tailwind for authenticity. As the cost of creating 'perfect' content drops to zero, the value of 'real' content increases. Users will crave the unpolished, the human, and the genuinely unique. The algorithm, which is only now beginning to grasp the depth of human preference, will likely pivot to reward this human signal over synthetic noise.

Skills for the New Product Era

Strategic Intent: Knowing what to build, not just how.
AI Orchestration: Using models to bridge technical gaps in design and data.
Psychological Depth: Understanding user motivation in an automated environment.
Rapid Iteration: Moving from idea to prototype in hours, not weeks.

The transition will be painful for those wedded to their silos. The agency owner of 2026 will not be looking for a 'data scientist' or a 'UX designer' in the traditional sense. They will be looking for operators who can wield a suite of AI tools to deliver a complete product vision. The moat is no longer your ability to perform a specific task; it is your ability to manage the entire process.

Key Takeaway

Specialization is becoming a liability; the future belongs to the generalist who uses AI to master the entire product stack.

04 — Experimental History

The Price of Prosperity: Why Safety is Killing Innovation

The sociological cost of a world without deviance

By Adam Mastroianni · 14 min read

Editor's note: We have traded the chaos of the 1990s for a safer, more prosperous world. But in doing so, we may have accidentally engineered out the possibility of greatness.

There is a quiet, unintended consequence to the prosperity of the last thirty years: the decline of deviance. Since the 1990s, we have seen a massive drop in almost every measurable form of social risk-taking. Crime is down, smoking is down, drug use is down, and teenage pregnancy is down. On paper, this is an unmitigated victory. We have built a world that is safer, healthier, and more stable. But this stability comes with a hidden tax. Deviance—the act of breaking rules, taking risks, and acting outside the norm—is the primary engine of cultural and technological innovation. When you eliminate the risk of failure, you also eliminate the possibility of the radical breakthrough.

The Prosperity Paradox

This decline is not a result of targeted social engineering or anti-drug campaigns. It is a byproduct of wealth. As people become more prosperous, they have more to lose. The cost of a mistake—a criminal record, a ruined reputation, a lost opportunity—becomes higher. In a world of increasing stability, the rational choice is to conform. We are seeing a society that is effectively 'optimising for the middle'. We avoid the extremes of behavior because the downside of being 'weird' or 'deviant' is now too expensive to bear. We have traded the volatility of the 1990s for a predictable, but stagnant, present.

When you eliminate the risk of failure, you also eliminate the possibility of the radical breakthrough.

This phenomenon is not just social; it is cultural. We see it in the way media is produced and how trends are formed. In the past, culture was shaped by outliers—the artists and thinkers who operated on the fringes. Today, culture is increasingly shaped by the center. Algorithms curate our experiences toward the most broadly acceptable, least offensive, and most predictable content. This is what we might call 'cultural carcinization': the tendency for diverse, wild ideas to converge on a small, safe set of strategies that are known to work.

The Internet as an Accessory

It is tempting to blame the internet for this flattening of culture. The argument is that mass surveillance and algorithmic curation have created a digital panopticon where everyone is afraid to be weird. But the timing doesn't fit. The most significant declines in deviance began well before the smartphone era. The internet did not cause the decline; it merely accelerated it. It provided the tools for 'de-frictioning'—the removal of the small, messy, and unpredictable interactions that used to define human experience. It replaced the friction of real-world risk with the seamlessness of digital consumption.

The Three Drivers of Cultural Stagnation

Carcinization: The convergence of ideas into safe, predictable patterns.
De-frictioning: The removal of the messy, unpredictable elements of life.
The Flat-Earth Problem: The loss of extreme viewpoints in favour of algorithmic consensus.

The challenge for the next generation of creators and entrepreneurs is to find a way to reintroduce friction. If the world is becoming more predictable, the value of the unpredictable increases. The most successful actors in a stable world will be those who can navigate the edges without being destroyed by them. We must find ways to allow for deviance—to allow for the mistake, the outlier, and the radical idea—without sacrificing the safety that prosperity has provided. Otherwise, we are merely managing a slow decline into a perfectly safe, perfectly boring, and perfectly stagnant existence.

Key Takeaway

Stability is the enemy of innovation; a world without risk is a world without breakthroughs.