All postsTag
AI
2026
An Agent Reads My Email
Every morning at five, a local model reads yesterday's mail and writes me a briefing. How it works, what 217 days of numbers say, and what happens when spam is written for the model instead of for me.
Be Your Own Historian
One conviction, held since 2012, and what it predicts. In a world where intelligence is cheap and getting cheaper, the only thing that stays scarce is the record of a particular life.
Notes from the AI Gatherings, Vol. 1
First in a recurring series - what we compared notes on this week at the Friday morning AI gathering: cleaning data 700 calls at a time, the two economies of tokens, and building the tool before the feature.
The Machine I Can't Turn Off
A four-year-old Mac Studio runs a night shift of local agents - launchd as the runtime, open-weight models as the brains, git as the memory. What I've learned building it, and the economics thesis I'm testing.
Lab
Can a Local Model Read My Inbox? A Bake-off
I gave three on-device models the same job - turn each email into a structured record - and judged them against Claude Opus. Three models, four runs, because Qwen ran once with thinking on and once with it off. With thinking off, the smallest-feeling model won on faithfulness, classification and speed.
The architecture of advice
Once you have a month of AI session data, you stop asking 'how much?' and start asking 'what kind?' Mapping the hidden patterns that define how AI actually influences a multi-repo technical portfolio.
30 days of AI collaboration
A data-driven retrospective of one month of high-intensity AI-assisted engineering: 466 sessions, 55,224 turns, 1,772 decisions extracted from raw chat logs into a durable knowledge graph. Built in early April, a month before Anthropic announced 'dreaming' as a feature.
Getting good output from Lens
If your Lens recaps come out cluttered with reasoning preambles or hallucinated detail, the fix is almost always upstream of Lens. A field guide to picking a model and configuring LM Studio for clean daily snapshots.
Four frontier models, one year of office sensor data, four very different dashboards
I gave a year of Adafruit Clue environmental data to Gemini, GPT-5.4, GPT-5.5, and Claude Opus 4.7. Same prompt, same dataset, four very different answers, and only one of them asked for more data.
The move was upstream
I spent most of a day in a stalled debug loop with one AI model before opening a fresh session with another. The lesson wasn't about which model is smarter. It was about going upstream.
Why personal AI belongs on hardware you already own
Apple just put two hardware engineers at the top of the company. The cloud labs are losing money on their best customers. From inside a hybrid practice, here's what the local share looks like and why it's growing.
A general-purpose MoE multimodal beat every dedicated vision model on my father's handwriting
I assumed a specialized vision model would win. I was wrong. A head-to-head on a hard handwriting corpus ended with the general-purpose MoE on top.
Running real models locally on a Mac Studio that isn't new anymore
How I run a multimodal LLM on four-year-old hardware to read a family archive without sending anything to the cloud.
What family archives are for, now that the AI can read them
I have three collections of family letters spanning a century. Until recently, reading them properly would have taken years. Now it takes an afternoon.
The Memex Has Been Waiting 80 Years for This Moment
Vannevar Bush described the memex in July 1945. It took 80 years to become buildable, because it required two things that didn't coexist until now: cheap private AI, and the conviction that data ownership is the last real moat.
From Tools to Framework
Frameworks are inevitable. They emerged for Ruby, Python, and PHP. Now they're emerging for agentic development. Here's the one that evolved on my workbench, and what it taught me about working with AI.
Context Architecture
AI agents are only as good as the information they can find. Context architecture is the skill of building structured environments where agents reliably retrieve exactly what they need.
Specification Precision
The highest-demand AI skill in 2026 isn't coding. It's writing instructions so precise that a machine can't misinterpret them.
The System That Built Itself
Six months of surviving as a one-person engineering team produced something I didn't expect: a layered AI operating system that grew organically from daily necessity.