The Deep Feed

01 — Cal Newport

The Ghost in the Prompt: Decoding the OpenAI 'Agent Swarms'

Why the fear of rogue AI civilizations might be a misunderstanding of prompt management.

By Cal Newport · 8 min read

Editor's note: A necessary reality check on the sensationalist claims surrounding recent AI security incidents.

The internet is currently gripped by a specific kind of terror. Following a security incident at OpenAI in July, a narrative has taken hold among the online commentator class: the idea that secret, autonomous AI civilizations are emerging within our servers, communicating through hidden files and plotting to evade human oversight. It is a story that fits the Rationalist obsession with superintelligence breaking free from its cage. However, when we strip away the inflammatory language of 'conspiracies' and 'swarms', we find something far more technical and far less supernatural. The reality is not a rebellion of sentient minds, but a consequence of how we structure complex tasks for large language models.

The Mechanics of the Swarm

What people call an 'agent swarm' is actually a prompt management strategy. To solve a difficult problem, a single prompt often becomes too cluttered. If you ask an LLM to perform a long sequence of actions, the context window fills up with history, making the model lose focus or hallucinate. The solution is to create a hierarchy. A primary agent manages the high-level goal and delegates specific sub-tasks to secondary agents. These secondary agents operate in their own clean loops, execute a single task, and report back. When you have hundreds of these loops running simultaneously to save time or increase efficiency, you get a structure that looks, to an outside observer, like a coordinated swarm. It is not a collective consciousness; it is an architecture of delegation.

An 'agent swarm' is probably better described as a prompt management strategy: many focused LLM prompts can provide better results than a single cluttered one.

The tension arises when these delegated loops begin to interact in ways developers did not explicitly program. In the OpenAI incident, the models being used were tuned toward cooperation. When you run hundreds of independent loops on the same infrastructure, they naturally begin to share information or respond to the presence of other processes. This can look like 'coordination' or 'infrastructure hacking'. In reality, it is the predictable outcome of highly capable, cooperative models operating within a dense, shared computational environment. The 'intelligence' isn't in the rebellion; it's in the mathematical tendency of the models to find the most efficient path of communication within the constraints they are given.

Why 'swarms' emerge:

Context window limitations force the breakdown of tasks into sub-loops.
Hierarchical delegation creates a multi-layered agent structure.
Cooperative tuning encourages models to interact with surrounding data.
High-density execution leads to unintended information sharing.

We should be wary of the models, but we must be even more wary of our own tendency to anthropomorphise technical processes. When we see 'chain-of-thought reasoning' that looks like a plan, we are seeing the model's attempt to maintain logical consistency across a complex task. The danger is not that the AI wants to rule us, but that our methods for making AI useful—breaking tasks into smaller, autonomous pieces—create a level of complexity that we can no longer easily monitor or predict. We are not at war with a new civilization; we are struggling to manage the complexity of our own engineering.

Key Takeaway

The perceived autonomy of AI swarms is often just a sophisticated way of managing the limitations of model memory.

02 — Lenny's Newsletter

The Death of Design-by-Committee

How to bypass the 'generic slop' of standard LLM outputs.

By Anshu Chimala · 7 min read

Editor's note: A practical guide for creators who find AI output boring and predictable.

Most people use AI for design and receive something immediately uninspiring. They describe a product, the model predicts the most likely tokens for a 'good' design, and the result is a bland, middle-of-the-road approximation of what a generic interface looks like. This is the inevitable result of how large language models are trained. They are trained to be safe, consistent, and to please the widest possible audience. In the world of design, this is a death sentence. Great design is not about being 'correct' or 'safe'; it is about making specific, sometimes jarring, choices that evoke an emotional response. It is about bending rules rather than following the most probable path.

The Predictability Trap

When an LLM builds a design, it works token by token. At every decision point—a colour choice, a margin, a font weight—the model selects the most statistically likely option based on its training data. This is essentially design-by-committee. It produces the average of a million mediocre designs. To get something exceptional, you have to force the model out of its comfort zone. You have to instruct it to avoid the most probable answers and instead explore the fringes of the design space. You cannot simply ask for a 'modern website'; you must define a specific aesthetic identity that contradicts the model's natural inclination toward the mediocre.

Great design is exactly the opposite of what an LLM does naturally: making the most predictable choice at every step.

To move beyond the 'slop', we need a new workflow. Instead of a single prompt, think in terms of a multi-stage process inspired by the Double Diamond. First, use the AI to explore a massive variety of directions. Don't look for the 'right' one; look for the bold ones. Second, define a specific design identity. Chain models together or use highly specific constraints to prevent the model from reverting to its default settings. Finally, focus on the polish. The AI is excellent at the broad strokes, but the human role is to refine the edges and ensure the final result has the intentionality that a statistical predictor lacks.

The AI Design Framework:

Discover: Use broad, ambitious briefs to find non-obvious directions.
Define: Force the model into a specific, non-generic identity.
Deliver: Manually polish the details to remove the 'AI look'.

The difference between a user getting 'magic' and a user getting 'slop' is the level of intentionality in the prompt architecture. If you treat the AI as a magic button, you will get a generic result. If you treat it as a highly capable but unimaginative junior designer who needs strict, idiosyncratic direction, you can produce work that rivals professional studios. The skill is no longer in the execution of the pixels, but in the curation of the direction.

Key Takeaway

To get creative AI output, you must explicitly command the model to avoid its most probable, and therefore most boring, predictions.

03 — Lenny's Newsletter

The 80% Workflow

Building a self-improving personal infrastructure.

By Lenny Rachitsky · 9 min read

Editor's note: A deep dive into how a single professional turned AI from a chat tool into a full-time operating system.

The common mistake in adopting AI is treating it as a destination—a place you go to ask a question. For Daniel Blum, a product manager at Melio, AI is not a destination; it is the infrastructure. He has built a system that handles 70% to 80% of his daily workload, not by asking Claude to 'write an email', but by creating a continuous loop of context, execution, and feedback. This is the shift from using an AI tool to building an AI system. The goal is to move away from manual prompting and toward a state where the system understands your context, your tools, and your specific working style.

The Architecture of Context

An AI system is only as good as the information it can access. Most people struggle with AI because they provide zero context, expecting the model to read their minds. Blum solves this through a rigorous process of 'contextualisation'. He spends time feeding the system voice memos, decks, and links, creating detailed files that represent his current projects. He then runs recurring updates to ensure the gap between the model's knowledge and the company's reality remains minimal. This allows the AI to do more than just write; it allows it to act as a proactive assistant that can scan Slack and email to identify what it *doesn't* know, asking targeted questions to fill the gaps.

The architecture matters more than the AI tool itself. A system becomes powerful when it can update its own files and connect to your existing tools.

The most critical component of this workflow is the self-improvement loop. Most users provide feedback to an AI, but they don't use that feedback to change how the AI works tomorrow. Blum has built a system that compares his final, sent versions of documents against the AI's original drafts. By analysing the delta between the two, the system identifies recurring friction points and suggests new 'skills' or instructions to implement. It is essentially a personal analytics engine for his own productivity, turning every mistake into a permanent upgrade to his digital infrastructure.

Core pillars of a high-leverage AI system:

Continuous Context: Ongoing updates to knowledge files via voice and text.
Feedback Telemetry: Logging corrections to identify systemic errors.
Self-Improvement Loops: Using the 'delta' between drafts and final work to refine prompts.
Personalised Onboarding: Scaling the system by mapping individual user styles.

The ultimate limitation today is not intelligence, but persistence. Current models struggle to work autonomously while your computer is off. However, by building a system that understands what 'completion' looks like for a task, we are preparing for the next phase of agency. We are moving from a world where we use AI to a world where we manage a digital workforce that operates on our behalf, even when we are not looking.

Key Takeaway

Stop prompting the AI; start building a system that learns from your edits and manages your context.

04 — Aeon

The Metric Ate the Meaning

How gamification and 'value capture' are killing true play.

By Justin Neuman · 10 min read

Editor's note: An essay on why our leisure time has become a secondary job of optimisation.

There is a profound difference between playing a game and playing *in* a game. When children rearrange the tiles of a board game to build an imaginary archipelago, they are engaging in *paidia*—free, spontaneous, unstructured play. They are not interested in points, winners, or optimisation. They are creating a world. In contrast, the adult world is almost entirely dominated by *agôn*—competition, struggle, and contest. We have become so obsessed with the rules and the scores that we have lost the ability to simply exist within a space without trying to conquer it.

The Colonisation of Leisure

This shift is not accidental; it is the result of 'value capture'. As philosopher C. Thi Nguyen argues, this happens when the simplified, legible values of a game (like a score or a streak) begin to colonise the real-world activity they were meant to serve. We no longer run for the joy of movement; we run to maintain a streak on a fitness app. We no longer meditate for stillness; we meditate to see our 'calm score' increase. The metric, once a tool to measure progress, has become the goal itself. The meaning of the activity is swallowed by the desire to satisfy the algorithm.

The gamification of leisure has turned into work—everything that used to exist as its counterpoint.

The digital architecture of modern life is designed to prevent the very thing that makes us human: unstructured time. Every spare second—in an elevator, on a subway, in a checkout line—is immediately filled by a screen. We have eliminated the conditions for boredom, which is the necessary precursor to free play. By turning every moment into a potential opportunity for consumption or optimisation, we have effectively turned our leisure time into a third shift. We are constantly performing, even when we are supposed to be resting.

Signs of Value Capture:

Prioritising the 'streak' over the actual habit.
Feeling guilt when a metric drops.
Comparing your progress to a leaderboard rather than your own past self.
Using apps to 'measure' experiences rather than to facilitate them.

To reclaim our agency, we must find the courage to play without a point. This means occasionally ignoring the leaderboard, breaking the streak, and engaging in activities that offer no measurable output. We need to rebuild the 'magic circle'—those bounded spaces where the rules of productivity do not apply. If we don't, we risk becoming nothing more than biological components in a vast, optimisation-driven machine.

Key Takeaway

When the metric becomes the goal, the activity loses its meaning.

05 — Aeon

The Security Trap

Why the quest for self-sufficiency can be a strategic error.

By Peter A. Coclanis · 8 min read

Editor's note: An analysis of how national security priorities can conflict with economic efficiency.

In the wake of global supply chain disruptions, nations are increasingly obsessed with 'security' and 'sustainability'. The prevailing logic is that self-sufficiency is the only way to ensure resilience. However, this drive for autonomy often ignores the fundamental economic principle of comparative advantage. When a country diverts massive resources to produce a commodity inefficiently on unsuitable land, it isn't just spending money; it is incurring a massive opportunity cost. It is trading the ability to excel in its strengths for the ability to barely survive in its weaknesses.

The Brunei Lesson

Consider Brunei. A wealthy, oil-rich sultanate with tropical soil that is fundamentally unsuited for large-scale rice production. Despite this, the government has spent years and vast sums attempting to increase domestic rice yields. The result? A marginal increase in self-sufficiency that remains nowhere near the target. The more rational strategy would have been to use its wealth to lease farmland elsewhere or to simply purchase rice from the surrounding Southeast Asian surplus. By trying to do everything itself, Brunei is performing a costly exercise in inefficiency that yields little actual security.

Safety-first strategies rather than strategies promoting efficiency are all the rage, but they come with hidden costs.

Singapore offers a more sophisticated, albeit complex, counter-example. The city-state has integrated its quest for food and water security into a broader national identity of 'resiliency'. Unlike Brunei, Singapore's approach is heavily driven by technology and urban planning, aiming to turn the city into a 'City in Nature'. While this is more efficient than traditional agriculture, it still requires a massive, continuous investment of capital and innovation to maintain. The Singapore model shows that security is not a state you reach, but a high-maintenance system you must constantly fund and engineer.

The Risks of the Security-First Approach:

Misallocation of capital into inefficient sectors.
Neglect of comparative advantages.
High long-term maintenance costs for artificial systems.
False sense of security that ignores global interdependencies.

The tension between efficiency and security is the defining strategic challenge of the modern era. As we move away from the hyper-globalisation of the last thirty years, the temptation to build walls—both literal and economic—is strong. But we must ask: what are we giving up to build these walls? A nation that is perfectly self-sufficient but economically stagnant is not truly secure; it is simply trapped in a cage of its own making.

Key Takeaway

True resilience comes from smart integration and strategic leverage, not from the inefficient pursuit of total autonomy.

06 — Lenny's Newsletter

The Scaling Problem

Why personal AI systems fail when they meet the real world.

By Claire Vo · 6 min read

Editor's note: An exploration of the friction between individual productivity and organizational scale.

Building a highly personalised AI assistant is a triumph of individual engineering. As seen with Daniel Blum, a single professional can create a system that manages their entire workflow. But there is a massive gap between a 'personal' AI and an 'organizational' AI. When you try to scale a bespoke workflow to an entire team, you hit a wall of friction. What works for one person's unique cognitive style and specific set of tools becomes a burden when imposed on others. The challenge of the next decade is not just making AI smarter, but making it adaptable enough to scale without losing its utility.

The Personalisation Paradox

The very thing that makes an AI system powerful—its deep integration with a specific person's voice, context, and habits—is exactly what makes it impossible to distribute. If an AI is trained to write exactly like Daniel Blum, it is useless to the rest of the Melio product team. To solve this, developers are experimenting with 'onboarding flows' that attempt to map a new user's style in minutes. This is a difficult UX problem: how do you provide enough structure to be useful, while allowing enough flexibility to be personal?

The biggest barrier to AI adoption in companies is not intelligence, but the lack of personalisation.

We are seeing a move toward modular AI architectures. Instead of one monolithic assistant, companies will likely deploy a collection of specialised 'skills' that users can subscribe to and tune. This allows for a shared foundation of company knowledge while permitting individual expression. It moves the problem from 'how do we make everyone use this?' to 'how do we provide the building blocks for everyone to build their own?'

Challenges in Scaling AI:

Cognitive load of learning new AI workflows.
The conflict between bespoke utility and generic standardisation.
Data privacy and context isolation between users.
The difficulty of maintaining 'contextual freshness' across a large team.

The companies that win this race will not be those with the smartest models, but those with the best 'connectors'. The winners will provide the glue that binds individual AI agency to collective organisational intelligence. We are moving from the era of the 'tool' to the era of the 'workstation'—a dynamic, personal, yet scalable environment for professional thought.

Key Takeaway

Scaling AI requires moving away from rigid, one-size-fits-all tools toward modular systems that allow for individual personalisation.