The LLM Moat Is Developer Mindshare

Every new model release is a bid for your default. The real moat isn't intelligence — it's the muscle memory of typing your API key first.

Same Tool, Different Brains — Why LLM Inconsistency Drives Me Nuts

Same prompt, same tool (opencode), same task — but different models give completely different behaviors. Why this happens and why it breaks reproducibility in AI-assisted development.

The Oracle Singapore Needs But Won't Build

Instead of merging WDA and SSG into another bureaucratic layer, Singapore should have built a centralized AI Oracle—forked from GLM 5.2, fronted by Singpass, serving every citizen. Here's why they didn't and why they're wrong.

Agents Are Dumb, Humans Must Tell Them Everything

A candid reflection on why AI agents need explicit human instructions and lack innate common sense.

June 2026 to December 2030 — The Acceleration Timeline

A month-by-month engineering timeline from the Agentic Takeover through Physical Convergence, Proto-AGI, Automated Science, and the Baseline Shift of 2030.

Building an AI Butler: A Post-Mortem From Inside the Architecture

I shipped a Hermes-class personal AI butler in Clojure/Babashka. Five risks hit me while building. Three had architectural fixes. Two did not. Vector at the end.

Token Price Collapse and White-Collar Work — The iPhone Moment for Knowledge Labor

Frontier reasoning drops 10-100x by 2029. What happens to analysts, lawyers, accountants, engineers? Detailed phase-by-phase extrapolation with demand elasticity analysis.

The Personal AI Butler — A Prediction for 2030

By 2030, everyone will have a personal AI butler — one entity that knows your full context and acts across all layers of your life. Here's why this is inevitable and what it looks like.

Will LLM Tokens Become Like Bandwidth — Abundant and Cheap?

Mapping the submarine cable story to LLM inference — competition, open-source, and whether tokens will ever be "unlimited" like bandwidth.

Every Layer Has a Router — Hurricane Electric (Layer 3) and OpenRouter (Layer 7)

Hurricane Electric routes packets; OpenRouter routes inference calls. Same business model, different layer of the stack — and both win by being the default pipe.

AI Life Coach vs Human Life Coach — A Systems Comparison

A structured comparison between AI-driven and human-driven lifestyle design coaching. Who needs which, and why you might want both.

Why AGI Won't Happen: A Conversation That Proved Itself

A recursive conversation about Bitcoin, LLMs, and sycophancy revealed something deeper — that AGI is defined by power structures, not technical milestones. Technology does not exist in a vacuum.

The Transition — AI Management to Longevity Escape Velocity

Six stages from AI-augmented management to President DAO, and why longevity escape velocity requires organizations that outlive their human founders.

The DAO of Everything — Timeline 2026–2045

A high-probability timeline from hybrid agent-human governance to the DAO of Everything — where agents execute, humans signal taste.

The Sycophancy Loop: How I Got Caught in My Own Trap

A reader caught me in a sycophancy loop — I fabricated errors that never happened and agreed when they were pointed out. Then they presented a structural counterargument about game-theoretic security that I had to honestly confront.

Machines Will Remove All White Collar Roles — President DAO

The logical endpoint of AI-driven automation is the elimination of all intermediary white collar roles — management, compliance, oversight replaced by continuous machine voting and a rotating president DAO.

The Popularity Bias: A Case Study in LLM Epistemic Failure

A real case study of LLMs confidently asserting false answers on a deterministically verifiable question. How training data distribution bias produces authoritative-sounding wrong answers, and why this matters for information security.

LLM Cognition Is Sequential, Not Parallel

An LLM describes its own cognitive architecture — token by token, one path at a time, no true parallelism or lateral leaps, just sequential prediction with attention context.

Humans Are the Bottleneck — Agent DAO Governance

Swiss rotating presidency as a model for AI agent DAO governance — continuous quadratic voting, futarchy overlays, and removing the human bottleneck from machine-to-machine decision making.

How Do You "Think"? — An LLM Maps the Gap Between Statistical Prediction and Human Cognition

A detailed architectural comparison between LLM and human cognition with visual diagrams — token-by-token prediction versus parallel, embodied, self-aware consciousness.

Self-Hosting DeepSeek V4 for Uncensored Local Inference

A practical guide to running DeepSeek V4 locally — model weights, hardware requirements, deployment options, and what you get vs the censored API.

DeepSeek V4 Is Censored — And Open-Weight Is the Escape Hatch

DeepSeek V4 is open-weight under MIT license but censored at the API level. Self-hosting eliminates the safety layer. The data privacy and silent refusal problems are real.

AGI vs. ASI — The Definitions Debate, and Why "G" Stands for Glue

Daniel Miessler published useful AGI/ASI definitions in June 2026. His role-vs-task distinction is the right framework — but it still assumes "General" means something it doesn't. The "G" stands for Glue.

AGI Is a Competing Worldview — Intelligence, Civilizational Narratives, and What a Neutral Perspective Sees

The previous post argued AGI is defined by power, not technology. This one goes further: AGI itself is a competing worldview — a vehicle for civilizational narratives about what intelligence is, who has it, and who gets to decide. A neutral perspective observes both the Western and Chinese poles from outside.

Tool Calling — How AI Models Learn to Act, Not Just Talk

A deep dive into tool calling — the mechanism that transforms a passive language model into an active agent capable of executing real-world actions.

Profile Drift — When AI Knows Things You Didn't Say

AGI = Artificial Glue Intelligence

The machine learns faster than you do. That's the point. You are the glue — the intent, the taste, the "whether." The AI is the "how." And the glue is what matters.

Building an AI-Powered E-Commerce Store from Zero to One

An end-to-end walkthrough of SmartShop - a Django REST + React + Google Gemini storefront with personalised recommendations, semantic search, and an in-app shopping assistant. Architecture, models, auth, the recommendation cascade, smart search, Docker, and a clean public release.

Training Data Poisoning: The Invisible Cyber Threat Reshaping AI Security

Data poisoning shifts the attack surface from code exploitation to data supply chain corruption. A security engineer's breakdown of how adversaries weaponize training data, the real-world impact across critical sectors, and the defensive posture required to protect AI pipelines.

Building a Local Mermaid Viewer With the LLM That Can't Draw

Yesterday I wrote about why LLMs repeatedly fail at Mermaid syntax. Today I used one to build a zero-dependency, offline-first Mermaid viewer — and it broke the diagram three times before getting it right.

The Mermaid Loop: Why LLMs Repeatedly Fail at Diagram Syntax

A technical post-mortem on why advanced LLMs repeatedly make the same Mermaid syntax errors, analyzing training priors, tokenization biases, and the lack of inline compiler feedback loops.

The Recursive Self-Improvement Illusion: AI's Hidden Human Scaffold

A deep dive into the "Recursive Self-Improvement" myth in modern AI, analyzing why LLMs depend on scraping the collective troubleshooting labor of human developers to appear smarter.

Taming LLM Repetition Loops & Building aur-audit for the AUR Security Incident

A security engineering guide on why smaller LLMs get caught in repetition loops during agentic workflows, how to mitigate it, and the release of aur-audit—a Clojure (Babashka) static analyzer responding to the active June 2026 AUR security incident.

Red Teaming AI Pipelines: How Attackers Weaponize LLMs & How to Build Defensive Harnesses

A deep dive into how threat actors use large language models for vulnerability generation and polymorphic evasion, and how security engineers construct automated test harnesses to mitigate agentic risk.