Writing

An experiment with AI deliberation revealed something uncomfortable: accumulating confident opinions feels like convergence on truth, but isn't.

I ran an experiment to find where multi-model deliberation adds value. The answer surprised me: it's the structured format, not the model diversity.

On the difference between eliminating friction and eliminating anxiety — and how to know when you've crossed the line.

Per-task model routing optimises cost per token. But at personal assistant scale, friction is the real cost.

In AI agent architecture, calling something a 'cross-cutting concern' without naming an owner and a gate is just a polite way of saying nobody owns it.

The planning/execution split is more useful than any benchmark comparison.

When red-teaming a document with multiple AI models, the second review — run on the edited version — consistently finds more than the first. Here's why, and what it means for how many rounds to run.

Evaluating LLM systems requires judgment, not statistics. That shifts who's qualified to do it — and where the gap is in most organisations.

When building MCP servers, search should return scannable summaries — not full content. Let the model decide what to read.

The retrieval pipeline was built for systems that couldn't reason about their own information needs. Agents can.

An open-source AI agent framework became the fastest-growing project in GitHub history — mostly in China. The pattern is the same as last year. So is the security panic.

Most enterprises are optimising the wrong variable. The gap between 5% and 40% agent adoption won't be closed by better models.

I spent an afternoon researching AI tools for personal life management. The conclusion was that I should stop looking.

One of the architects of modern deep learning just raised $1B on the thesis that token prediction can't reach real reasoning. Here's what he's proposing instead — and why it matters even if he's wrong.

An AI agent ran unsupervised for two days and found twenty improvements to another model's training. Not an AGI claim. A rate claim.

The shift from agents you summon to agents that watch. Enterprise AI workflows are becoming continuous loops — and the failure modes are different.

Not better function calling — decoupling. When tools expose MCP servers, any agent can compose any system freely. The heterogeneity problem becomes a configuration problem.

We called them language models and spent years confused about why they could reason. The name stuck to the interface, not the mechanism.

Ask an AI to write from scratch and you get the average of the corpus. Give it something rough and it amplifies what's already there. The workflow implications are significant.

Clients hire consultants to solve problems. What they're actually paying for is the reduction of a particular feeling. The distinction matters.

Most PKM tools are procrastination with better aesthetics. The problem isn't the software — it's that filing a note feels like understanding it.

Not productivity. Something stranger — the cognitive texture of days when the bottleneck shifts from execution to articulation.

The hardest phase of AI architecture isn't building the stack. It's the moment after the stack is built and eighteen teams start making independent decisions on top of it.

The comfort trap is about effort. This one is about epistemics — and it's harder to see.

The right test for any AI interaction isn't 'did it help me?' but 'am I more capable after it?'

AI coding agents didn't just make developers faster. They changed who gets to have a bespoke system.

The moment I stopped building scheduling into my tools, everything got simpler.

When building automation around a third-party app, the first question to answer is: what's the one thing this app does that nothing else can replicate? That feature becomes the tax you pay on everything else.

Running 10 real queries through 5 tools revealed that theoretical routing rules have systematic gaps — and the surprises were more useful than the confirmations.

AI agent results should be push, not pull. The feedback loop should close on mobile. Most tools miss all three — not from ignorance, but because dashboards photograph better.