Dan Grover

i lost the lid to my coffee grinder and codex tracked down the part, found the dimensions, and made me a 3d print file that fit perfectly in one shot

I call this the Trashcan Method of AI Engineering: 1. Identify thing you need 2. Build it fast, comprehension be damned 3. Does anyone actually use it? How? Cool write that down it’s your new spec 4. Throw away all old code and rewrite from scratch Don’t feel pressured to 1-shot maintainable code. Code is cheap, throw it away more.

sf has maybe 30 physical AI startups right now and i'd bet real money half of them have never run their robot outside a controlled demo environment. you can tell who's real pretty fast. the real ones have scratched up hardware, engineers who smell like solder, and a warehouse lease south of market that's bleeding them dry. the vaporware ones have perfect demo videos, a pitch deck with "foundation model" on every slide, and a founding team that hasn't touched a motor controller since grad school. the talent pool for people who can actually bridge sim-to-real is maybe 200 people worldwide and VCs are bidding them up like it's 2021 crypto. if your robot only works in a video, you don't have a robotics company. you have a content studio.

Big news for AI on a budget. GLM 5.2 Colibri int4 is a Mixture of Experts model that runs entirely on your CPU. No GPU needed. It's fast, efficient, and opens up advanced language capabilities to anyone with a standard computer. This is a game changer for offline AI.

Introducing “Comeback City: A Love Letter to San Francisco.” Proud to call this place home.