Writing
A running list of things I'm thinking through, most of it pulled from my own notes and half-finished arguments. These are rough drafts, not finished essays. I'm posting the titles and the gist early, partly to make myself write them.
Work in progress · personal notesThree small labs shipped fine-tunes of NVIDIA's open Parakeet v3 in one week, each 156 to 253 MB and each claiming to match the 2.5 GB original. Accuracy at the top barely moved; the useful gains were names, speaker labels, turn-taking and models that fit on a phone. With an iPad PDF.
Local speech recognition on a base M5 Mac: ten models scored on 89 hours of open audio in 10 languages. Granite Speech 470M wins English, Whisper covers the Nordic languages, Qwen3-ASR leads Hindi. Two models still running. With an iPad PDF.
Open speech models are now within about a point of the closed services on word error rate. The real choice is timestamps, speaker labels, latency, price and licence. With an iPad PDF.
What moved in speech recognition from April to October 2026: small open models near the paid APIs, an end date for the Whisper API, speech recognition built on EmbeddingGemma 2, and Underdog Bark 3.0. With an iPad PDF.
Five papers and posts on shipping small models agree on something unglamorous: what the model is asked to output, how its labels were cleaned and how it is served matter as much as the model. With an iPad PDF.
A model gets smaller and more reliable each time you hand one of its jobs to plain code. And the metric can lie, so distrust single numbers. With an iPad PDF.
One week of frontier launches showed the top models converging on price. What’s scarce now is access, distribution and your own tests. With an iPad PDF.
Big labs are locking up their models’ reasoning. Small open models are winning narrow jobs anyway. What that means if you build with small models, with an iPad PDF.
A 2,500-year-old book that kept describing my job. Four passages on leading invisibly, starting small, subtracting daily, and teaching without words.
The highest-leverage thing we can do for education is make it easy for parents to learn alongside their kids. A case for tightening the loop between kids, parents, schools and communities, and for treating learning intent as the real asset.
A father's first-principles toolkit, reconciled across two value systems: a South Indian upbringing and the Swedish world my sons are growing up in. What I chose to carry over, what I let go of, and the questions I still haven't answered.
A build log. A sandboxed, browser-based coding classroom where each kid gets a container and a terminal in the tab, no setup and no IDE, with an AI tutor alongside. Why I built it, and what I learned watching them use it.
Two lenses for building learning tools. As a teacher: know the learner, assess gently, actually care that they progress. As a technologist: every early AI product is skeuomorphic, so you change one variable at a time to find what's genuinely new.
AGI's economic upheaval through the two lenses I actually hold: a parent of young kids, and an engineering director whose own role it reshapes. Less forecast, more "what does this mean for them, and for me?"
Teams don't move fast because they're handed autonomy. They move fast when information travels quickly between the people making decisions. A case for small bets in one direction, and for leaders who set direction instead of watching the work.
Context constrains the search space, clarifies intent, and cuts compute. For anything built on a general model, managing context about the user and the material is the real differentiator, not the model. Plus a note on stress-testing your own strategy before you commit to it.
Intelligent people are uniquely prone to confusing complexity with depth: the ten-page memo that should be one page, the system that costs twice as much and impresses everyone. Mastery is the elegant simplicity that cuts through.