Notes

Agents are a third consumer of design systems, and they read structure rather than screens. Why the library isn't dying, why most AI misses are knowledge gaps, and what people do once the watching is delegated.

More agents means more output, not more progress. Four principles for telling real work from productivity theatre: count the human cost, explore before committing, set stop conditions, and budget your attention.

Your prompts stay on your machine, but that isn’t the risk any more. Once a model has tools, a stranger’s file can make it act. The five layers to check, and what to change in an afternoon.

AI is framed as a race with one winner. But intelligence is becoming ambient, the useful models are small, and the real risks are only manageable together.

Agents are moving from your terminal into pipelines that run on their own. What a software factory looks like, the pieces it needs, and why the engineer’s job becomes running the line.

REA lets a coding agent open up the apps on your computer and explain how they work. An X-ray for non-developers, an accelerator for developers, and a way to find the requirements you didn’t know you had.

As AI gets more capable, taste, agency, focus, resourcefulness and follow-through gain leverage. What each looks like now, and why it’s worth revisiting the projects you ruled out.

AI coding tools often buy a bigger queue, not a faster team. Six principles for turning individual speed into work that actually ships.

Coding agents keep invalidating the model’s cache, so every big prompt means another long wait. How oMLX keeps that cache on SSD, and how to set it up on a Mac.

Google bid $10 million for a bankrupt airline’s work emails to train agents. Why records of work aren’t the work, and who decides what “done” means.

A model at the top of a benchmark that almost nobody can use. Why customer obsession, not leaderboards, is the test that matters for AI products.

How MLX spreads models across several Macs over Thunderbolt: the stack, mesh versus ring, pipeline versus tensor parallelism, and faster local fine-tuning.

How MLX runs the whole agent loop on Apple silicon: the four-layer stack, a three-step setup, and what makes local agents fast enough for real work.

Agents make the app easier to leave and the agent harder to replace. How data, workflow and agent layers decide what software is still worth buying.

Price per token tells you little about what work costs. What Opus 5.5 shows about counting the whole job, steering, long runs, and testing a model on your own tasks.

Better agents can make AI bills grow fast. Remove the work that no longer needs to exist, then match what is left to the right model.

Jev is a fast and very cheap general-purpose classifier. Where it fits alongside code and LLMs, what it costs, and how to try it.

one ten hundred thousand million billion trillion quadrillion quintillion sextillion septillion octillion non-alion decillion undecilian duodacillion treata Cillian quadawardisillian quindicilian ceticilian…