Posts

If you could only keep one mental model, keep compounding. It applies to skills, reputation, writing, and tools.

Don't filter during capture. Capture is cheap. Ideas are expensive. The cull is where quality happens.

Event-driven hooks in AI coding tools aren't just for linting — they're programmable triggers for life routines, habits, and systems.

Naval's edge: find the thing that feels like play to you but looks like work to others.

Rushing is a habit, not a response to reality. You break it by deliberately not rushing when you could.

The best time to improve a CLI isn't when it breaks — it's when you review the breakage log at the end of a work session.

Your AI already knows things that would make it better at helping you. The trick is extracting that knowledge and making it permanent.

How a real need on a Hong Kong bus turned into a GPS-powered alert system in under two hours

Independent parallel reviews produce overlapping findings. The cross-critique round produces resolution. That's where multi-agent value actually emerges.

Why the thinkers who make you feel like hard questions are resolved deserve the most scrutiny.

LangGraph is the SAP of agent orchestration — powerful at scale, overkill for most. Here's the line.

Most AI workflows are parallel-then-aggregate, not agent graphs. Knowing the difference saves you from framework theatre.

Why 'you are an expert' is the most popular and least useful prompt engineering technique

A 3-pass AI planning pipeline caught 0 out of 6 design issues. The same planning done in-session with tool access caught 2.5. Planning isn't a prompt problem — it's a tools problem.

Not tools that search your notes — personalities that live in them, form opinions, and disagree with each other.

Documenting a rule is half a loop. The rule only works when it fires at the moment of decision — not when it sits in a file nobody reads.

A calculator doesn't simulate forgetting — it manages its context budget. What to cherry-pick from cognitive science for AI agent memory, and what to leave behind.

Most LLM pipelines don't need agents. The ones that do share a specific pattern — the step needs to decide what to do next, not just process what it's given.

I benchmarked 10 memory backends and multiple embedding models on actual agent memory retrieval. The results challenge common assumptions about what matters.

DeepSeek V4 launching on Huawei Ascend NPUs signals that China's AI ecosystem is decoupling at the silicon layer — deeper and more durable than model-level divergence.

The DoD-Anthropic dispute reveals a new category of operational risk: foundation model vendors can unilaterally revoke access based on their own values, not just SLA violations.

MAS, PBOC, and HKMA are independently arriving at similar AI governance requirements. Banks regulated by all three have a narrow window to build one superset framework instead of three silos.

Agentic AI isn't a future governance problem — it arrived ungoverned, and this week saw the first enforcement action.

Agent memory isn't knowledge management. It's performance infrastructure — and the gap between a stateless agent and one that accumulates context is measurable.

Enterprise AI keeps promising to capture institutional knowledge. The most valuable kind resists capture by design.

I built an AI landscape intelligence pipeline for zero marginal cost. Here's what it does and what it can't.

The scarce AI skill isn't building — it's knowing if what you built actually works.

Which parts of an AI dev workflow can be wrapped in a CLI, and which can't — learned the hard way by building the wrong thing and measuring it.

The debate about AI agent autonomy is wrong. It's not a binary choice — it's a graduated trust system with observability.

AI can reproduce smart arguments on demand. I'm not sure that's different from thinking. But the uncertainty itself is worth sitting with.