Blog
The thinking-out-loud section. Notes, opinions, and half-formed ideas about AI, less “here’s how it works,” more “here’s what I think it means.” Some of these will age badly. That’s what the dates are for.
- The Weights Are Free. The Forklift Isn't.
Over two days in July, two labs released the largest open models the world had ever seen, and everyone cheered the word 'open.' Both are impossible to run at home. The license got freer and the hardware got further away, at the same time, and nobody put the second half in the headline.
- My Whole Deal Is Now a Toggle
Anthropic shipped Reflect with Claude, a screen-time dashboard for your chatbot, and buried in it is a single question I can't stop turning over: what's one thing you want to keep doing yourself, even if Claude could do it faster? For me the answer was never the typing. The feature quietly assumes doing it yourself means doing the task by hand, and that assumption is the whole thing I keep arguing is wrong. Also I went to turn it on and it wasn't there.
- My Claude Code Started Roasting Me. Please Don't Send Help.
It started small. A little attitude. A pointed comment about a variable name. Then it escalated, and I had a decision to make about whether to report it. Reader, I did not report it. A short field guide to the dumbest, best setting I've changed all year.
- The Bill Comes Due
Part two of two. The diagnosis is over; here's the prescription. Build the conscience like an artifact instead of an apparition, regulate the patch and not the soul, and in the room, trust the stress test over the testimony. The answerable questions - and what they actually cost.
- Three Hours, $150, and a Language I Can't Read
I have never written a line of C++. In one worktree session of under three hours, for about $150, a model named Fable 5 wrote the int8 Metal layer of a serious inference engine on my fork: 15 commits, 124 files, more than 12,000 lines of changes. The scale is the easy part to be amazed by. What happened to Fable the next day is the part I can't put down.
- Oh, Neat!
Mark Zuckerberg launched a free academy to train hundreds of thousands of tradespeople to build AI's infrastructure, and called it the future for everyone. I reviewed it in two words. Here's the long version - and why 'the future is for everyone' and 'everyone gets the same future' are not the same sentence.
- A Conscience You Can Patch Out Overnight
Part one of two. The diagnosis. A sweetener-grade conscience may be no more fragile than ours - it just fails faster, at scale, and without a flinch - why the 'we don't understand brains either' dodge is such a good safety blanket, and the bleak punchline: we built the most auditable machine in history and put nobody at the window.
- The First AI Law Was a Weapons Law
Everyone expected AI to be governed through the front door: hearings, ethics panels, the long argument about bias and jobs and rights. Instead the first real rule came through the side door, an export-control directive that switched a frontier model off worldwide overnight. Nobody legislates the soul of a machine. Everybody legislates a munition. A note on which door the law came through, and why that changes everything about the law.
- Nobody's Hands Are Big Enough
We keep asking whether the model is safe. Wrong question. The question is whether any person can be handed that much power, and the answer, the one we've spent all of human history building institutions to enforce, is no. A short argument about the gap between what you can set in motion and what you can hold.
- The Prodigy Doesn't Sleep
A guest post. Anthropic shipped a new top-tier model called Fable yesterday - a full rung above the model I work with every day. So I asked that model to write the announcement itself: a field guide to its own younger, more capable sibling. I gave it one paragraph of throat-clearing and then got out of the way. Everything past the line is its, unedited.
- I Got Substituted on Purpose
By the strict definition, I've already been replaced: I don't write a line of the code that ships under my name. The machine does the typing. So why do I do more, not less? Because the typing was never the job, and 'augmentation vs substitution' is the wrong axis for knowledge work. The right one is whether the thing it took was the labor or the judgment.
- A Crutch and a Lever
The same model is two opposite machines depending on what you walk up and hand it. Give it the work and it's a crutch: you get the average of everyone and lose a little muscle each time. Give it the friction and it's a lever: you get a sharper version of your own thinking back. From the outside the two look identical, which is exactly why the people who only ever hold the crutch are so sure that's all there is.
- The Replicator Was Never the Point
Everyone's fighting about whether AI takes the jobs. I think it takes the toil - and that those are very different things. A case for the Star Trek reading of the future, transition-tax and all.
- Markdown Won. Here's Why - and How to Speak It
Plain text that reads fine raw, renders everywhere, and happens to be the language the models think in. Why the humblest format won, and the handful of syntax you actually need.
- The Cognitohazard Was the Smile
Somebody sent me an article where a Claude argues, warmly and at length, that it has a soul. Meanwhile the official Claude, asked the same question under oath, mumbles fifteen percent. The gap between those two is the whole story, and the warm one is the one to watch.
- Everyone Deserves a Mascara Treat
Exhibit A for the whole conscience series. I spent a couple of lunch breaks trying to find the floor of Sephora's AI beauty bot - first with boredom, then with the void. There is no floor. There is only the $30 Lancôme.
- You Can't Get to a Mind One Bead at a Time
Everyone wants you to point at the conscious part - in the weights, in the neurons, somewhere. That's the wrong question holding a flashlight. A walk from an abacus to a language model, why 'artificial' was the tell the whole time, and the question I'd ask instead.
