Harnesses are everything. Here's how to optimize yours. — Alex Ker đź”
Why read: A practical guide to configuring AI agent harnesses to maximize output quality and avoid context bloat.
Summary: The key to maximizing an LLM's utility is the harness managing its context and tool calls. A bloated global system prompt forces the model to hallucinate by exceeding its "instruction budget." Instead, adopt Progressive Disclosure by keeping configuration files lean and letting the agent discover tools only when needed. Use the R.P.I. framework to structure prompts so models approach problems like a staff engineer. Lastly, deploy subagents to handle isolated tasks, keeping the main context window clean and orthogonally aligning the harness with high-quality output.
Agents share environments, not data — Hunter Leath
Why read: A conceptual shift in how we should architect multi-agent systems, moving away from uploading data to sharing entire environments.
Summary: In the 2010s, microservices scaled by sharing pointers to data (e.g., S3 URLs) rather than the data itself. Today's AI agents face a similar scaling bottleneck, but with a twist: agents don't just operate on single files; they operate on an entire context or environment. Attempting to zip and transfer files between specialized agents is inefficient. Instead, we must build systems where agents share the entire environment—the disk, binaries, documents, and SQLite tables. This allows specialized micro-agents to seamlessly hand off deep contextual understanding without the latency of data transfers.
The Moat Is Time, Not CUDA — Rethinking Nvidia Export Controls — Jukan
Why read: A contrarian take on Nvidia's export controls, arguing that massive, unified compute clusters are humanity's premier strategic asset.
Summary: Nvidia's Jensen Huang recently downplayed the compute needed for large AI models, calling it "mundane." However, this is likely a defense mechanism to prevent GPUs from being classified as highly regulated strategic materiel. Training massive models requires hundreds of thousands of GPUs physically bound together, dealing with constant hardware failures and immense memory synchronization challenges. This scale of compute is not trivial or easily replicated. By framing it as ordinary, Nvidia protects its margins and autonomy, obscuring the reality that time and massive infrastructure—not just CUDA—are the true moats in the AI arms race.
Thoughts and Feelings around Claude Design — sam henri gold
Why read: A critique of Figma's complex design primitives and why the future of design tooling is shifting back to code.
Summary: Figma won the design tool war by creating its own proprietary primitives (components, styles, variables) that loosely approximated code. However, this locked-down format excluded Figma from the training data of modern LLMs, which are fluent in actual code. As AI makes it easier for designers to write and manipulate code directly, the baroque infrastructure of design systems looks increasingly archaic. Tools like Claude Design embrace "truth to materials" by working natively in the medium where the product will live. Ultimately, the source of truth is naturally migrating from lossy design approximations back to the code itself.
In August I wrote a thesis I never published — Brad Lyons
Why read: An analysis of why software multiples are crashing and how regulatory moats plus AI represent the new winning formula for SaaS.
Summary: Traditional SaaS moats based on engineering complexity have been decimated by AI, allowing small teams to replicate years of roadmap in weeks. This democratization of product development has triggered a collapse in software multiples. The remaining defensible moats are distribution, proprietary data, workflow breadth, and regulatory insulation. Regulatory complexity—such as compliance, audit trails, and certifications—creates high switching costs regardless of product quality. For vertical SaaS incumbents, the winning strategy is combining their captured regulatory surface area or proprietary data with functional AI before challengers can catch up.
We’ve identified four workflow categories: — Josh Schultz
Why read: A clear framework for categorizing workflows to optimize the deployment of AI agents and human effort.
Summary: Workflows can be divided into four categories: manual, assisted, delegated, and autonomous. Autonomous workflows are complex, token-expensive, and run continuously for table-stakes maintenance (e.g., SEO audits). Delegated workflows run once with human orchestration and validation. Assisted workflows use AI to help humans complete tasks faster. Finally, manual workflows rely on pure human effort for relationship-building and high-touch interactions. Selecting the appropriate AI level for each workflow is becoming a core strategic focus for modern organizations to maximize leverage.
Why read: An exploration of how generative AI transforms text from a static medium into a dynamic, living entity.
Summary: Historically, text was inherently nonliving—ink on paper or pixels on a screen. With the advent of LLMs, we are witnessing the birth of literal "living text." This text possesses protean dynamism, reshaping and regenerating itself in response to the environment and the reader's actions. Unlike human writing, which is produced serially in real-time, AI-generated text can be produced atemporally and infuse words with artificial life. This shift turns static documents into interactive entities, fundamentally changing the medium and message of written communication.
Deep Dive: The 60 Critical Minerals That Power the Modern World — Chamath Palihapitiya
Why read: A stark reminder of the geopolitical bottlenecks surrounding the physical inputs required for AI, energy, and defense.
Summary: The modern technology stack—from data centers to EV batteries—relies entirely on 60 critical minerals. China recognized this decades ago, securing dominance over the processing of rare earth elements through cheap labor and lax regulations, now controlling roughly 85% of processing. As AI and industrial demand explode, the US and its allies are facing severe vulnerabilities in supply chains. Winning the century requires solving cheap energy, and critical minerals are the fundamental input. This dynamic is driving a new geopolitical race and creating significant opportunities for private capital to rebuild capacity.
Why read: A foundational hiring philosophy from the CEO of Scale AI on avoiding the trap of becoming a credential-driven institution.
Summary: The most critical trait to screen for in early-stage hiring is whether a candidate genuinely cares about the company's mission and their work in general. As startups grow, they naturally attract top-of-funnel candidates who view the company as a prestigious credential rather than a cult-like mission. If left unchecked, this transforms the company into a university-like environment with smart but uninvested employees doing mediocre work. Recruiting must resemble courtship, searching for intensely obsessive individuals who have a history of working extraordinarily hard on things they care deeply about.
A 10-minute delay becomes a 24-hour delay by the end... — Alex Bouaziz
Why read: A powerful argument for why fast response times and unblocking others should be an executive's highest priority.
Summary: Small delays compound drastically across distributed teams and time zones. A simple 10-minute delay in responding to a query can easily cascade into a full lost day of productivity as work gets handed off across the globe. Fast response culture transforms company velocity by treating unblocking as the ultimate priority. Leaders must proactively identify and remove bottlenecks before they stall the chain, rather than waiting for perfect paths. As demonstrated by Elon Musk's intense design reviews, clearing hurdles on the frontline is the essence of moving fast.
Why read: Palantir's manifesto on the moral obligations of the tech elite and the necessity of hard power in the AI age.
Summary: Silicon Valley owes a moral debt to the nation that enabled its success, and its engineering elite must actively participate in national defense. The era of building consumer apps is yielding diminishing returns; true civilizational progress requires delivering economic growth and security. In an increasingly hostile world, the soaring rhetoric of soft power must be backed by hard power built on software and AI. Adversaries will not pause to debate the ethics of AI weapons, making it imperative for democratic societies to lead in their development to maintain global deterrence.
the neuroscience of visualization: why it works (but not how you think) — Jaynit
Why read: A scientific breakdown of why traditional manifestation fails and how elite performers use mental practice to physically rewire their brains.
Summary: Mainstream "outcome visualization" (imagining success) often saps motivation by tricking the brain into feeling rewarded prematurely. However, process visualization—vividly imagining the execution of a task—creates real neurological changes. The brain struggles to distinguish between real and deeply imagined experiences, firing the exact same neural pathways. Studies show that mental practice alone can increase muscle strength and expand the motor cortex. Visualization is a powerful tool for building skills, provided you focus on the rigorous practice of the action rather than the euphoria of the outcome.
These Are Leadership Ideas I Try To Apply Every Day — Ryan Holiday
Why read: Core leadership principles drawn from Stoicism and high-performance environments.
Summary: Effective leadership requires a codified philosophy to guide daily decision-making across an organization. A primary tenet is maintaining a "sense of urgency"—understanding that small delays have outsized downstream consequences and tackling tasks immediately. However, this must be balanced with the Latin concept of Festina Lente (make haste slowly), which emphasizes purposeful, measured exertion over frantic rushing. By establishing clear maxims and rules of engagement, leaders can empower their teams to operate autonomously while maintaining high standards and forward momentum.
Why read: A philosophical look at how AI's displacement of labor could return society to tradition-based, non-economic communities.
Summary: Modern society defines humans primarily as economic producers and consumers, organizing cities around labor markets. As AI consumes white-collar jobs and removes the economic incentive to migrate for work, our reason for gathering in multicultural urban centers will dissolve. This shift will force people to evaluate where they live based on shared cultural, religious, and aesthetic values. The result could be a return to tribalism and tradition, where communities form around shared ways of life and spiritual ideals rather than mere economic output.
.@tobi on why the right video game is a great... — David Senra
Why read: A reflection on how strategy games like StarCraft serve as the ultimate sandbox for learning entrepreneurship.
Summary: Complex simulation games teach high-agency actors how to manage resources, attention, and the consequences of their decisions. Shopify's Tobi LĂĽtke credits StarCraft with teaching him that there are no objectively right decisions, only contextually correct ones. It provides a rapid feedback loop for learning when to build infrastructure, when to invest, and when to reveal your hand. While business is ultimately a positive-sum, infinite game that contributes to humanity, the foundational strategic muscles required to succeed are effectively forged in the zero-sum crucible of these digital worlds.
The Evolution of Defensibility: In an AI-native world, traditional engineering moats are collapsing; the new moats are regulatory capture, proprietary data, and sheer infrastructural scale.
The Premium on Urgency and Agency: As technical execution becomes commoditized, operational velocity, intense care for the work, and unblocking team bottlenecks are the ultimate differentiators.
Rethinking Abstractions: From design tools to AI harnesses, layers of complexity are being stripped away in favor of lean, context-aware environments that operate closer to the ground truth.
Post-Labor Philosophy: Generative AI is prompting deep reflections on how society will reorganize itself when human economic output is no longer the central organizing principle.
