Archive

Adding detail to a deliverable doesn't fix credibility — it creates new interrogation targets.

A strand of multi-agent research — latent-space inter-agent communication — is thriving in Chinese labs and almost invisible in the West.

A 3-year AI roadmap designed around today's model capabilities may be solving last year's problem by year 2.

Monitoring systems need consumers before they need features

Most AI writing tools want to chat with you. The better model is annotation — AI that reads what you wrote and leaves margin notes.

The real risk of building a personal AI operating system isn't that a better tool appears — it's that your system's complexity becomes the thing you maintain instead of the thing that maintains you.

When AI agents can grep, read, and reason iteratively, most RAG infrastructure becomes unnecessary middleware.

When you have both a credential and real work in the same domain, route effort through the real work.

The best use of AI in investing isn't picking stocks — it's building the pipeline that teaches you why you can't.

Most people use AI transactionally. The real unlock is conversational — thinking with the model, not through it.

In a world where AI has perfect recall, the skill that matters is rebuilding frameworks from first principles — not memorising them.

Every layer humans retreat to, AI follows. The question isn't what we're still good at — it's what we teach the next generation when every cognitive advantage has a shelf life.

The value of having read everything is collapsing toward zero. What's left is what you can't extract from a model.

Most knowledge workers use LLMs as search engines. The real unlock is using them as subject matter experts you debrief.

The philosophical debate about whether LLMs understand is orthogonal to whether they're useful for knowledge extraction.

Actions are cheap to redo. Bad decisions compound. Build systems around the judgment calls, not the mechanical steps.

Most people do spaced repetition for facts but not for beliefs about themselves. Wrong priors calcify because there's no review system.

Most AI memory is either always-on or ephemeral. The missing category is prospective: remember until a context arises, then forget.

When AI tools search conversation history, they should index both sides by default — not just the human's half.

When an answer requires multi-step reasoning to reach, save the conclusion — a fresh start won't reliably reproduce the chain.

When you mine knowledge from an LLM, certain types come easily. Others are systematically absent. A taxonomy reveals the blind spots.

Not all knowledge works the same way. A taxonomy for what you're actually capturing when you write down what you've learned.

Machine learning says let the model find the signal. Heuristics research says use one variable and ignore the rest. They're both right — the dividing line is how much data you have.

Gigerenzer's thesis is that simple rules outperform complex analysis. If you've already internalised that, reading 300 pages of evidence for it might be the exact kind of overthinking he's arguing against.

Experts don't make more decisions — they make fewer, by having better defaults. The real meta-skill is accumulating simple rules and knowing when to stop reasoning.

Whether you're trusting a doctor's prescription, an AI agent's code, or a junior engineer's pull request — the trust heuristics are identical.

Title, salary, company, industry — the standard job evaluation checklist misses the things that actually predict whether you'll thrive.

Most technology decisions can be reasoned through. AI solution design can't — the domain is too empirical, too fast-moving, and too non-linear for theory alone.

Not all knowledge is equally trustworthy. Three tiers of validation — from 'a model said it' to 'it survived reality' — and why tracking the difference matters.

Getting ahead of AI best practices is a treadmill. The durable skill is testing assumptions faster than they expire.