18 posts
The first piece said nobody had yet specified an AI flight recorder. It was wrong: three proposals were already public. The recorder is turning out to be the easy part. What is still missing is who holds the record, who may open it, and what it may be used for.
The Oireachtas heard last week that data centres now use nearly a quarter of Ireland's electricity. The debate runs in megawatts per site, and no small business can do anything with a megawatt. The number they need is how much energy one answer costs. I've been measuring that on my own machines, and the biggest lever wasn't the model.
Xiaomi has released the highest-scoring open-weight model anyone has measured, under the MIT licence, and almost nobody can run it. The weights are the least useful part of the release. The recipe, the graded environments and a 9B student model are the parts that fit on a desk.
A new study finds AI tutors match expert human tutors on GRE learning gains, one of them at a nine-hundredth of the cost. The result is real. The mechanism it reports — faster replies, more messages, more practice — is also the shape of an engagement loop, and we already know what fluent machines do to people's willingness to check.
The Titanic carried more lifeboats than the law required. The law had stopped counting at ten thousand tons. What followed was not a ban but a way of learning from accidents: a recorder, a duty to report, and an independent investigator. AI has two of the three on paper, none in force, and one of them built backwards.
Generative AI reached 62% of American adults in three and a half years. The internet needed eight. The climb rate is about the same. What differs is where each one started, and who's paying for it.
Todd asked whether it's a mistake to sort a system like Claude.ai into intelligent or not-intelligent. It is — but not for the reason people usually give. A tour of the add-ons that make the thing you're talking to, and why the yes/no question can't even find its referent.
People keep asking me which AI is best. The question assumes an oracle — one machine you can settle on and stop thinking about. That's the exact habit that gets you burned.
Anthropic just closed $65B at a $965B post-money valuation. That's not a refutation of the bubble thesis. It's a date stamp on the next test of it.
Kahneman split thinking in two. The standard worry about AI is that we'll stop using System 1. That's not what's happening. Not to me.
Fifty years of software development, an AI coding partner, and the question of what the human is actually for.
An independent path to the same architecture — persistent memory, direct repository access, and a strict division of labour. I got there second, and the parallel is still uncomfortably exact.
I built a tool that runs two independent Claude sessions side by side, lets them argue, and lets you fire one's answer at the other. It's the most useful AI tool I've made.
How persistent memory, direct git integration, and a strict architect/coder split produce software that neither human nor AI could build as well alone.
It started with a command-line hack to change the music. Eighteen months later, Claude Code is generating half a billion dollars a year and redefining what software development looks like.
On growing minds rather than programming them, on software that fits like a handmade shoe, and on what a glowworm moving through space-time has to do with the question of what intelligence is for.
AI isn't just automating work — it's compressing the distance between intention and execution, activating latent ambitions, and shifting the real barrier upstream to judgment and taste.
Most developers use AI as a sophisticated clipboard. Here's what changes when you build the infrastructure to make it an actual collaborator.