Technical articles on AI, Web3, and software development
Hexagram 18 is a picture of worms bred in a vessel left standing too long. I came back to an autonomous content system I built and found exactly that: bugs that grew in the dark because no one was reading the logs. The talk I submitted to PyCon HK 2026 is the field report.
We benchmarked 25 LLMs on structured regulatory output. Across Anthropic, DeepSeek, and Google, newer models regressed. The reason is instructive — and the subagent trap is worse than you think.
Edward Shaughnessy's 1983 Stanford dissertation proved that the oldest layer of the Zhouyi encodes star charts, seasonal calendars, and astronomical myths — not abstract philosophy. The dragons of Qian are the Dragon constellation rising and setting over the course of a year. The ghosts and swine of Kui are lunar lodges. Here's what he found, and what it means.
Tobi Lutke built an AI agent that only works in public. It accidentally recreated the apprenticeship model I've been arguing for. Capability over functionality, taste over technique, and why the best AI rollouts look like a German Lehrwerkstatt.
Google discovered 273 of 12,936 URLs from my sitemap. Search Console said 'Success.' The problem wasn't content quality or backlinks — it was a directory path.
A deep dive into Claude Code's preprocessing pipeline — the 6-stage system that transforms raw user input into a carefully orchestrated message stream with context injection, file pre-reading, skill discovery, and message normalization. Based on source code analysis.
On woodblock prints, parallel subagents, and the point where OCR stops correcting and starts fabricating. Also: why downsampling saves you 75% and nobody tells you.
We ran a 7-state Warring States wargame with AI agents. The smallest state consulted I-Ching hexagrams before each move. The hexagram content didn't predict actions — but it doubled the agent's diplomacy rate and improved survival. Here's what that tells us about why every strategist in ancient China consulted the Changes.
Most AI governance frameworks collapse risk into a single score. ARA Eval decomposes it into 7 dimensions, applies deterministic gates, and tests whether LLMs can reproduce human judgment. The answer: only one model can. I open-sourced the framework.
Three Android phones, one weekend, zero successful WhatsApp transfers. The error is 'Failed to find provider info for com.whatsapp.provider.migrate.ios.' The fix is a shell script. The journey to get there involved ADB forensics, a 2014 APK from the Wayback Machine, and a manual database decryption pipeline I never needed.
12 years at Informatica taught me that every data migration fails at the same step: understanding what you have. AI is finally good enough to change that.
Adding site search to 8-Bit Oracle revealed that every SSG page was an empty div. The fix touched auth gates, static params, and hidden content — and changed how I think about what 'rendered' means.
Making provisioning software usable by supply officers on phones in a warehouse. 320px minimum, 20px text, light/dark themes, bilingual — and what 'mobile-first' means when your users don't think in software terms.
I spent 60% of my effort fighting a framework and 40% building domain logic. Then I looked at OpenClaw (500k lines) and NanoClaw (6k lines) and realized: a sturdy core plus skills that encode taste beats a large framework every time.
On turning Han dynasty oracular poetry into AI-generated ink paintings by encoding art direction as a reusable agent skill. Also: why content filters...
A raw brainstorm on whether initial model weights could predispose toward naturally aligned behavior - and why the Confucian junzi might be a better...
Counter-intuitive finding: forcing Cantonese (yue) on Whisper causes decoder collapse. Force Chinese (zh) instead for coherent, searchable...
Using Claude Code and OpenRouter infrastructure to rapidly build model evaluations with Claude Opus 4.5 as the judge
Extracting the deterministic core of 8-Bit Oracle into a standalone library, and why 'same seed, same hexagram, every time' matters more than you'd...
Upgrading 8-Bit Oracle's Hexagram 24: Return with Proust's famous involuntary memory scene, then systematically testing 12 diverse AI image models...
A three-day journey through Twitter API v2, v1.1, OAuth, rate limits, and the gradual acceptance that you can't actually hack your way around free-tier...
A technical post-mortem on why WordPress's REST API is theoretically elegant but practically unusable for CI/CD, how SiteGround's security theater...
How we learned to stop worrying and love the binary calendar. On deleting databases, removing cron jobs, and discovering that the oracle was encoded in...
What we actually ship in 8-Bit Oracle today: JSONB conversations, hexagram context, and Postgres views — plus why we didn't adopt Neo4j, MemMachine, or...
Complete technical comparison of MemMachine vs mem0, PostgreSQL graph alternatives (recursive CTEs, ltree, adjacency lists), and detailed...
On replacing Wilhelm with Unknown Pleasures, cultural anchoring vs keyword stuffing, and whether we're all just shouting into the void while ChatGPT...
Hard constraints beat soft self-regulation when an LLM's reasoning converges to a template.
Why creating ADRs during development, not before it, leads to more honest and valuable documentation of your architectural decisions.
A debugging session reveals why your gut feeling about 'two sets of items' was right - and how to guide AI toward simpler, better solutions.
Replace style copying with principled search: explore a multi-dimensional voice space and use Pareto-optimal selection to shape an original AI voice.
A technical deep-dive into how we avoided a risky database migration by choosing an additive strategy for adding Simplified Chinese support to our...
My experience with Claude Code's productivity gains, the double-edged sword of rapid development, and lessons learned from over-engineering features.
Exploring why Mem0's default LLM prompts are overly restrictive, causing the `infer=True` mode to filter out too much content. Includes root cause...
Quick guide to configure browser-tools-mcp AgentDeskAI with Claude Code - solving connection issues and getting your MCP server properly integrated.
The standard DSPy OpenRouter integration lacks failover and proper app attribution. Here's how to fix it with custom headers to get automatic model...
A step-by-step guide to integrating RevenueCat Web Billing in a plain React/Next.js app. Covers account setup, SDK usage, code samples, and key gotchas...
A dive into debugging service worker caching and update notifications for new pages in a Jekyll site, ensuring the 'Content Refresh' toast works as...
A technical exploration of 'Futures Thinking,' a potential experimental feature for 8-Bit Oracle, looking at a Triple Hex Mode and AI-driven scenario...
Building the Android frontend for FIDO2 passkeys: tackling CredentialManager, migrating DataStore, and the joys of debugging in Jetpack Compose.
A developer's journey through the challenges of implementing FIDO2 passkeys for an Android app using Python, covering assetlinks.json, data...
A developer's guide to deploying a CosmWasm smart contract on the Stargaze testnet, covering environment setup, starsd CLI usage, contract uploading...
Deep dive into recent Bonsol framework improvements: a new I-Ching sample dApp, RISC0_DEV_MODE integration for faster development, and custom S3 URL...
Integrated OriginTrail DKG with Eliza AI and a custom divination backend, enabling on-chain persistence of divinations via a Telegram bot.
An exploration of using the I-Ching's King Wen sequence as a potential framework for self-directed learning in AGI. Details the thought process, key...
Detailing significant enhancements to the ai16z Eliza AI project, focusing on local-first capabilities, model optimization, and backend stability.
An overview of Zustand as an efficient state management library for React, highlighting its benefits over traditional methods (refs, Context API) for...
A guide to minimizing re-renders and optimizing performance in React/Next.js applications with authentication. Covers combining state, memoizing...
A guide to conditionally rendering components, like a footer, in a Next.js 14 (App Router) application by bridging server and client components using...
A developer's experience integrating Shadcn UI components into a Next.js project. Covers project preparation, handling opinionated styling (dark theme...
A detailed guide to troubleshooting and configuring Fail2Ban with UFW for robust SSH security. Covers iptables rule order, jail.local setup, persistent...
A developer's journey through UI framework choices, from Bootstrap to Next.js with Shadcn/ui. Discusses the pros and cons of component libraries like...
Explores client-side vs. server-side authentication using Supabase with Next.js 14 (App Router). Covers Google OAuth2 (client-side flow with...
Details on setting up authentication and database for the 8-Bit Oracle project using Next.js, Supabase (with Auth-UI for OAuth2), and Tailwind CSS...
A developer's comparison of using Claude Opus and GPT-4 (via Cursor.sh) as AI coding assistants, particularly when working with Next.js App Router vs...
A developer's notes and tips on implementing a smooth, multi-paragraph typing effect in React with a skip animation feature. Covers state management...
An analysis of using Chainlink VRF for I-Ching divination, comparing it to traditional methods and other randomness generation techniques like PRNGs...
Troubleshooting common Webpack issues, including double script loading due to manual and Webpack config imports, and resolving Flash of Unstyled...
A quick developer tip: If your React useEffect hook seems to run twice in Next.js development mode, Hot Module Reloading (HMR) might be the reason...
A project idea to create Fortune-Telling NFTs by integrating the ancient wisdom of the I-Ching (Yin/Yang, Trigrams, Hexagrams) with modern blockchain...
My initial thoughts on designing an AI service that can analyze and critique uploaded photos. I explore a multi-stage approach using CNNs, CLIP, and...
A step-by-step guide to diagnosing and fixing network connection problems between a Dockerized Spring Boot app on AWS EC2 and an RDS Postgres database...
My developer notes on quickly getting to grips with JWT authentication and data persistence in Spring Boot 3. I cover my experiences with Spring...
Some brainstorming on using AI to transcribe family stories shared at gatherings. I'm thinking about detecting narrators, mapping complex family...
My experience implementing Retrieval Augmented Generation (RAG) and AI Agents with Haystack 1.x. I cover conversational memory, creating custom data...
My practical tips and experiences from working with the Haystack LLM framework. I cover version differences (1.x vs. 2.x beta), the benefits of...
A guide on dramatically improving GPT-J-6B inference speed. I'll show you how I converted a Hugging Face model from its standard format to GGML...
Exploring the difficulties in finding truly uncensored AI models. I discuss the influence of OpenAI's data and look at methods for fine-tuning base...
A breakdown of a helpful Reddit discussion on AI model formats like GGUF, GGML, and tools such as Llama.cpp and Hugging Face Transformers. I add my own...