David's Archive
The article explores the ongoing debate between HTML and Markdown as output formats for AI agents, arguing that the optimal choice depends on the audience. Markdown, while cheaper and machine-readable for agent-to-agent communication, contributes to human information overload due to its lack of visual density. HTML, despite higher token costs and security challenges, excels in human-facing interactions by providing rich, interactive, and visually dense outputs that aid comprehension. Major enterprise AI platforms are increasingly adopting HTML or similar structured UI components for human-consumable agent outputs, signaling a shift towards formats that balance efficiency with human understanding.
This article introduces Andrej Karpathy's LLM Wiki, a system designed to transform raw documents into a structured markdown knowledge base. This personal knowledge base can be efficiently queried using Claude's capabilities. The guide provides practical instructions on setting up this system quickly, specifically highlighting its integration with Obsidian in just five minutes.
The EveryInc/compound-engineering-plugin introduces the "Compound Engineering" methodology, which aims to make each unit of engineering work easier by reducing technical debt and codifying knowledge. It inverts traditional development by focusing heavily on thorough planning and review (80%) before execution (20%).
The article argues that complex AI applications for enterprises require a "forward deployed engineer" approach, prioritizing hands-on implementation and deep integration over initial high gross margins. Drawing parallels with the success of Salesforce and ServiceNow, it suggests that AI startups should embrace professional services to become indispensable "systems of work," capturing valuable data and building a durable moat. This strategy, though initially sacrificing margin, leads to faster gross profit growth, greater market dominance, and eventually high margins as services become more scalable through AI automation.
This article describes how an LLM gateway optimizes model routing using an online learning approach. It specifically employs Thompson sampling over lognormal latency posteriors to dynamically select the fastest and most reliable LLM route for each incoming request. This strategy is designed to achieve cost-efficient operation by ensuring optimal performance and reliability in real-time.
Hamel’s guide to LLM evaluations emphasizes that effective evaluation is a continuous development process, beginning with rigorous error analysis performed by domain experts. It advocates for identifying and prioritizing real-world failure modes in application traces, using pragmatic binary pass/fail judgments. The article recommends integrating evaluation as a core part of development, focusing on understanding actual failures rather than solely optimizing for high pass rates.
In a piece presented by LangChain, Jensen Huang discusses the critical need for companies to adopt "open agent systems." His key argument focuses on why these systems are essential for modern businesses, likely exploring their benefits and strategic importance. This insight from Huang is highlighted through the LangChain channel.
AI agents are writing code faster than humans can review, creating a bottleneck in software development workflows. This article proposes four strategies used by PostHog to leverage AI for offloading code review and adjacent tasks, allowing human developers to review less code and focus on complex issues. These methods include making agents review each other's work, delegating PR babysitting, implementing an auto-stamper for low-risk PRs, and verifying agent-generated code through observation rather than reasoning. By adopting these approaches, teams can maintain quality, accelerate shipping, and reduce developer fatigue.
The `qa-swarm` is an AI-powered GitHub bot skill designed to provide comprehensive, multi-perspective Pull Request (PR) reviews. It orchestrates four independent AI agents—qa-team, paul-reviewer, xp-reviewer, and security-audit—to run in parallel, each bringing a unique review perspective. The skill synthesizes their findings, deduplicates them, and posts them as inline GitHub PR comments and an upserted summary comment with a calculated verdict. This ensures thorough code review while clearly marking all automated feedback.
