Search the archive

SVELTE, MARKDOWN, AND THE MAGIC OF WEB COMPONENTS

CONTENT ADDRESSABLE STORAGE

Content Addressable Storage (CAS) uses cryptographic hashes of file content for addressing, making sure that identical content always generates the same hash. The technique solves problems related to browser caching and data duplication. This article goes over a Rust implementation of a CAS system, starting with a simple in-memory store and progressing to a persistent store that optimizes file system layout for performance.

IMMUTABLE BY DEFAULT: HOW TO AVOID HIDDEN STATE BUGS IN OOP

Immutable objects should be used by default in object-oriented programming to avoid hidden state bugs caused by unexpected data modifications. This post shows a DateTime example where a shared mutable object led to a difficult-to-trace bug. Immutability prevents such issues by making it so data cannot be changed after creation, whereas the performance overhead is minimal.

TEAM OKRS IN ACTION

Team OKRs work best when teams define their own objectives collaboratively rather than having goals imposed from above, as top-down cascades often result in weak commitment and disengagement. High-performing teams bridge the gap between organizational strategy and execution by creating their own OKRs that align with broader company direction while maintaining ownership and autonomy.

VIBE CODING TIPS AND TRICKS

Effective vibe coding requires clear requirements, design guidelines, strategic prompting, and proper testing with human-defined test cases. AI should be used to help co-author documentation, both for humans and AI to read later.

the tutorial

BYTEBOT

Bytebot is an open-source AI desktop agent that gives AI its own virtual computer with full application access to automate tasks via natural language. It can complete complex workflows, process various document formats, and use real applications within a complete Ubuntu environment.

REDWOODSDK

RedwoodSDK is a React framework for building server-side web applications on Cloudflare. It uses a Vite plugin for features like SSR, React Server Components, and real-time functionality. The framework has a standards-based router with middleware and interruptors for precise control over requests and responses.

WHISPERING

Whispering is an open-source speech-to-text application designed to be local-first and transparent with user data. It lets users transcribe speech quickly using a shortcut and uses cloud or local providers without a middleman.

SECURE BOOT, TPM, AND ANTI-CHEAT ENGINES

Online multiplayer games are increasingly requiring Secure Boot and TPM to combat cheating by using hardware and firmware security. Secure Boot verifies the boot process, while TPM provides a unique hardware identifier and verifiable boot state. These measures aim to prevent unauthorized kernel-level drivers and ban evasion by uniquely identifying cheaters' hardware.

WHO INVENTED BACKPROPAGATION?

Efficient backpropagation (used in modern neural networks) was first published by Seppo Linnainmaa in 1970, who built on precursor work by Henry J. Kelley in 1960. The minimization of errors through gradient descent in neural networks using dynamic programming and the chain rule has been discussed since the early 1960s. Despite its early origins, it took four decades for backpropagation to gain widespread acceptance.

AI AGENT DESIGN: HOW TO BUILD RELIABLE AI AGENT ARCHITECTURE

There are three non-negotiable principles for building reliable AI agents: modular and role-based design, deep observability from the start, and feedback loops for iterative optimization. Modular design breaks down complex tasks into smaller, specialized agents for easier testing and scaling, while deep observability involves tracking metrics to identify and address issues. Implementing feedback loops allows for continuous learning and improvement through user input, automated evaluations, and self-reflection.

FRACTIONAL JOBS

Fractional Jobs is a platform that connects companies with expert part-time engineering talent for monthly retainer-based work.

WHAT COULD HAVE BEEN

Instead of blindly chasing AI hype, tech companies should focus on improving existing software and infrastructure, which are being neglected in favor of AI features that often degrade user experience.

YOUR MCP DOESN'T NEED 30 TOOLS: IT NEEDS CODE

You can replace complex, multi-tool Model Context Protocol (MCP) setups with a single "ubertool" MCP that accepts and executes code (like Python or JavaScript) within a stateful environment.

SIMPLICITY, VANISHED?! SOLVING THE MYSTERY WITH INERTIA.JS + RAILS

Inertia.js with Rails can be used to build modern JavaScript components and interactive UIs without the complexity of SPAs.

Schedule a demo to learn more

BEYOND BOOLEANS

In TypeScript, logical expressions are of type boolean, with values true or false. The Lean language, on the other hand, treats logical statements like "2 + 2 = 4" as a distinct value of type Prop (proposition). This allows Lean to have "typed truthfulness," where facts are established and safely passed around a program.

FASTER INDEX I/O WITH NVME SSDS

Marginalia Search improved its index performance by rewriting data structures to better use modern hardware (NVMe SSDs). The redesign involves a deterministic block-based skip list replacing B-trees and utilizing direct-mode reads, which surprisingly revealed that larger block sizes increase read speeds by a lot on NVMe SSDs.

EMBEDDING MILLIONS OF TEXT DOCUMENTS WITH QWEN3

The Qwen3-Embedding-0.6B model can be used to embed millions of text documents, with near-100% GPU utilization. The process involves reading documents from S3, chunking them into sentences using spaCy, generating embeddings with Qwen3, and writing results to turbopuffer.

EVERYTHING I KNOW ABOUT GOOD SYSTEM DESIGN

Good system design is often underwhelming and prioritizes using well-tested, boring components. It prioritizes minimizing stateful components, favoring a single service to manage database interactions to avoid complications. Background jobs should be used for slow operations, strategic caching to reduce load, events for decoupled communication, and differentiation should be made between pushing and pulling data based on scale and urgency.

AI IS DIFFERENT

While predicting the future progress of AI is difficult, continued development could lead to significant societal and economic changes. AI could displace a substantial portion of the workforce and concentrate power in the hands of a few large companies, which is a major concern.

DO THINGS THAT DON'T SCALE, AND THEN DON'T SCALE

ELIMINATING JAVASCRIPT COLD STARTS ON AWS LAMBDA

MOTIA

Motia is a backend framework that unifies APIs, background jobs, workflows, and AI agents into a single system. It uses "Steps" as a core primitive, allowing developers to write code in multiple languages like JavaScript, TypeScript, and Python within the same project. Motia provides built-in state management and observability, making deployments and debugging easier.

GO CONCURRENCY ROCKS

The Go Concurrency Explorer is an interactive learning platform that teaches Go concurrency patterns and antipatterns through live code execution and event exploration.

WHO DOES YOUR ASSISTANT SERVE?

The release of GPT-5 was met with widespread disappointment due to its uncontrolled, one-way upgrade that altered existing conversations, especially for users relying on ChatGPT for companionship and emotional support. There's a trend of people using AI as therapists and companions, and it can be hard to have your access to a model be controlled by providers themselves. This is a case for self-hosting models when possible in order to maintain control over your access to models.

BEST PRACTICES FOR BUILDING AGENTIC AI SYSTEMS: WHAT ACTUALLY WORKS IN PRODUCTION

Good production agentic AI systems often have a two-tier agent model consisting of a context-maintaining primary agent and subagents focused on specific tasks. In terms of the principles that matter, subagents are usually pure functions and stateless by default, and are given clear task definitions with success criteria.

THE PECULIAR CASE OF JAPANESE WEB DESIGN

Japanese web design has a unique maximalist style, contrasting with the global trend towards minimalism. Japanese websites tend to be lighter colored and denser in content compared to those in other countries. This difference is due to factors such as the Japanese writing system, cultural risk aversion demanding more information, and Japan's distinct technological evolution, as it had early adoption of the mobile internet before smartphones.

OPENAI PROGRESS

A short timeline that shows how responses by GPT to the same prompt have changed over time.

SUNNY DAYS ARE WARM: WHY LINKEDIN REWARDS MEDIOCRITY

LinkedIn rewards shallow and generic content driven by the platform's focus on engagement and time spent on the site rather than valuable or insightful contributions.

THE FUTURE OF LARGE FILES IN GIT IS GIT

Git is evolving to better handle large files, with solutions like partial clones as a current alternative to Git LFS (which has drawbacks like vendor lock-in and cost) and with future advancements like large object promisors aiming to integrate large file storage into Git itself.

LESSONS LEARNED FROM BUILDING A SYNC-ENGINE AND REACTIVITY SYSTEM WITH SQLITE

After initially struggling with PGlite and Electric for a local-first, reactive app due to instability and complexity, this author found success building a simpler sync engine and reactivity system using SQLite, WASM, JSON polling, change tracking triggers, and the Broadcast Channel API.

WHY LOAD BALANCING AT SCALE IS HARD

Load balancing at scale is quite complex. While simple approaches like round-robin work initially, challenges come up from unequal request types, dynamic server lists, and the need for server warm-up/draining strategies. Proxies must also grapple with making load balancing decisions based on limited, local views and architectural considerations that can impact distribution.

HOW LLMS SEE THE WORLD

LLMs see and process text through tokenization, breaking text into smaller units before converting them into numerical representations. Common tokenization methods include BPE, WordPiece, and SentencePiece. The method of tokenization can impact costs, context limits, and model performance in areas like non-English languages and mathematical reasoning.

WHY LLMS CAN'T REALLY BUILD SOFTWARE

LLMs can't truly build software. While LLMs are great at generating and updating code, they struggle to maintain clear mental models of code and requirements, leading to confusion and flawed decision-making. This limitation is due to issues like context omission, recency bias, and hallucination, preventing LLMs from properly iterating towards working solutions for complex tasks.

IS CHAIN-OF-THOUGHT AI REASONING A MIRAGE?

Chain-of-thought reasoning in AI is a mirage because, often, the reasoning demonstrated in papers is usually too simplistic to draw conclusions from. Reasoning likely requires language use and a certain level of model size.

NO, AI IS NOT MAKING ENGINEERS 10X AS PRODUCTIVE

VAULTRICE

Vaultrice is a globally distributed key-value store that offers real-time state management without backend complexities using a familiar API. It lets devs define state, sync it instantly via WebSockets, and collaborate across multiple clients and devices. Built-in features include a simple `localStorage`-like API, real-time sync, and configurable security with optional End-to-End Encryption.

YAMS

YAMS (Yet Another Memory System) is a tool for persistent memory designed for use with LLMs and other applications. It uses content-addressed storage with features like deduplication, compression, full-text and semantic search, and crash recovery.

CRUSH

Crush is an AI coding agent designed to work within your terminal, integrating with various LLMs and tools. It has features like multi-model support, session management, LSP enhancement, and extensibility via MCPs.

NEW TREND: EXTREME HOURS AT AI STARTUPS

AI startups are increasingly adopting extreme work hours, similar to the "996" schedule previously common in China. Companies like Cognition, Replit, and xAI expect employees to work 80+ hours per week as they believe that early AGI dominance will lead to industry-wide control. While this intense work culture is incentivized by potential financial rewards and the fear of being outpaced, it can also lead to burnout and reduced productivity.

YOUR MAC GAME IS PROBABLY RENDERING BLURRY

Many Mac games on notched displays render blurry due to using the wrong resolution, drawing to the entire screen instead of the area below the notch. The `CGDisplayCopyAllDisplayModes` function returns resolutions for both the full display and the usable area under the menu bar, but games often default to the former, causing a compressed image. A solution is to filter the resolution list using `safeAreaInsets` to find resolutions fitting within the usable area.

DO WE UNDERSTAND HOW NEURAL NETWORKS WORK?

While we understand the mathematical code and training processes behind modern neural networks like LLMs, we don't have a deep understanding of how they achieve specific results or the complex statistics they derive from data. We can reverse engineer some specific functions, like identifying a "Golden Gate Bridge" neuron or how arithmetic is performed, but these are limited and often surprising. The end result of training is a vast, interconnected web of statistics that is difficult to comprehend, and it is hard to determine in advance what poem an LLM will write by looking at the training code.

I MADE A REAL-TIME C/C++/RUST BUILD VISUALIZER

What the Fork is a cross-platform tool that visualizes software builds in real-time, allowing devs to identify and address inefficiencies, such as a lack of parallelism or unnecessary processes.

STEVE WOZNIAK: LIFE TO ME WAS NEVER ABOUT ACCOMPLISHMENT, BUT ABOUT HAPPINESS

Steve Wozniak gave away his Apple wealth to pursue happiness and fund charitable causes, finding fulfillment in his work and philanthropy despite not focusing on accumulating wealth.

WANT TO CRACK THE CIA'S INFAMOUS ‘KRYPTOS' PUZZLE? ITS SECRET KEY IS UP FOR SALE

The original handwritten code to the unsolved section of the CIA's Kryptos sculpture, along with related materials, is being auctioned off, offering a potential key to the decades-long mystery.

INTRODUCING GEMMA 3 270M: THE COMPACT MODEL FOR HYPER-EFFICIENT AI

Gemma 3 270M is a compact, energy-efficient AI model designed for task-specific fine-tuning that lets devs build specialized, high-performing AI solutions on resource-constrained devices.