Applied AI reviews, build logs and essays on systems, validation, agent workflows, decision-making and how ideas survive contact with reality.
A neat theoretical result suggests that different networks, trained on different data, can converge to the same internal representation, up to a rotation. The interesting part is that SGD itself may be responsible. I rea
Read essay ->
A 9B model challenging 27B-class models, an uncensored 27B running locally, and speculative decoding pushing another 27B model to 70 tokens per second on a Mac. The local-model stack is moving faster than the model-size
Read ->
An experiment in turning pictograms into a small, extensible protocol for communication across languages, objects, and machines. What if a T-shirt could work as a small dictionary? That was more or less the starting poin
Read ->
I want local models that can reason, use tools and work in the background without eating half my laptop’s memory. Three very different approaches suggest we’re getting surprisingly close. I like local inference. What I d
Read ->
Gisting turns a 6,000-token system prompt into 1,500 learned tokens, cutting latency by 38% without measured quality loss. Prompt…
Read ->
X has open-sourced enough of its recommendation stack to reveal what actually matters for reach in 2026. Some of the numbers are much stranger than the usual “post consistently and engage with your audience” advice. One
Read ->
Rambling changed how I work with AI agents. Not because speaking is smarter than typing, but because the machine can finally absorb the…
Read ->Agents’ Last Exam moves beyond coding puzzles into real professional workflows. On its hardest tasks, frontier agents are still nowhere near replacing humans. AI benchmarks have a rather unfortunate habit. We invent them
A wonderful 2016 paper found Linux leaving CPU cores idle while runnable work waited elsewhere. The deeper lesson is about what happens when you understand a mature system well enough to stop treating its defaults as sac
A practical 12-principle constitution for turning “we use AI” into something measurable: shared state, company memory, explicit agent authority, verification, executable rules, and controlled self-improvement. Almost eve
