Fabian G. Williams aka Fabs

Fabian G. Williams

Principal Product Manager, Microsoft Subscribe to my YouTube.

Don't Worry About the Terminator. Worry About David.

When people talk about AI ending badly, they reach for Skynet and the Terminator: a machine that wakes up and turns on us. I think the better warning is David, the android in Prometheus, who did harm because of what his maker set out to do and taught him to value. Mo Gawdat makes the same point: AI is a magnification of whatever humanity is. So responsibility is shared. The labs owe us broader data, tighter release discipline and a business model that doesn't make companies rent back their own intelligence. The rest of us owe ourselves something too: stop giving our data away, catalog it locally, and build our own ontology that compounds. Here's what that looked like for me this weekend, grading my own local AI by hand.

Fabian Williams

11-Minute Read

A bar chart titled One weekend grading my local AI, by hand. My 3 real questions: 3 pass. 21 trick questions, traps and unknowns: 17 pass, 4 partial, 0 fail. Invented claims: 0 in the 3 real questions, 1 in the 21 trick questions.

BLUF (bottom line up front): I’m just starting too. My data has lived in my Obsidian vault, my 2nd brain, for a while, but the practices around it were fragmented. They’re now coming together as a whole, and that whole complements, and at times runs ahead of, what I focus on at work: observability, accountability, governance and ownership. My rule for all of it: if you can’t see it, you can’t observe it; if you can’t observe it, you can’t measure it; if you…

Limen: I Put a Camera at My Window and Asked Who Owns What It Sees

This week is Fix, Hack, Learn at Microsoft, and my project is a local-first street-vision system I am calling Limen, the Latin word for threshold. An Arduino board watches my front window, a vision model on my own MacBook Pro M3 Max names what crosses (person, vehicle, wildlife), and a human confirms the label to build a training set I actually own. No cloud, no subscription, no frame leaving the house. Today I ran a scripted 90-second proof of concept: I walked up, drove in, parked, and walked to the door, and every one of 180 frames was auto-labeled locally in about 8.4 seconds each. Here is what worked, what did not (plates are still unreadable), and the one hard constraint that shaped the whole architecture: my evals rig is sacred and this project was never allowed to touch it.

Fabian Williams

9-Minute Read

A view from a front window: a man walking up the path toward the door holding a bag and a bottle, boxed on device with a green person label, with his vehicle boxed and a blue vehicle label, and a caption reading Limen local VLM auto-label on-device no cloud

This week is Fix, Hack, Learn (FHL) at work (Microsoft), our internal week to chase a build we care about. Last FHL I worked on my agents and agentic loops, helping me scale myself and my work as a product manager. This time my project is personal, and it starts with a question I could not stop turning over: when a camera watches the front of my house, who owns what it sees?

Doug Was Right: I Swapped In The MoE, And The Concurrency Math Changed

Last post I measured the concurrency ceiling on a 27B dense model and closed with a promise. Doug Ware, who builds applied AI systems, had handed me the one caveat I did not test: a dense model is the hard case, and a mixture-of-experts model with only about 3B active parameters per token should leave real headroom for a second slot to pay off. So I ran the exact same 36-load matrix on an MLX MoE, changed one variable, and let the numbers settle it. The dense model flatlines. The MoE keeps climbing. And at 8 concurrent agents the dense model makes you wait 32 seconds for a first token while the MoE answers in under 1.

Fabian Williams

9-Minute Read

A line chart showing aggregate throughput staying flat near 20 tokens per second on a dense model while a mixture-of-experts model climbs from 58 to 159 tokens per second as concurrent agents rise from one to eight

Two days ago I posted 36 barrier-synchronized loads against 1 local model and let the numbers answer a question Reddit handed me. That post ended with a promise. A reader named Doug Ware, who builds applied AI systems for a living, had replied with the one caveat I had not tested, and I said out loud that measuring it was the next thing I would do. This is that measurement. I did not wait, I did not hand-wave it, and I kept every honest asterisk in.

Recent Posts

Categories

About

Fabian G. Williams aka Fabs Site