Field notes.

Whatever shiny model you are looking at, the bottleneck is going to be the data layer underneath it.

I sat with a long Andrej Karpathy video until I could finally explain what is happening when I type into ChatGPT.

I watched Erik Schluntz's vibe coding talk three times before the whole insight landed.

I do not use AI for where it is. I use it for where it will be in nine months.

What the second law of thermodynamics keeps teaching me about everything I care about.

I keep noticing people use the word 'AI' to mean two completely different things.

Why a lot of enterprise AI pilots seem to ship and a lot of production rollouts seem to die.

I have started thinking that the fastest way to test an AI strategy is to build the smallest version of it before the deck gets approved.

I spent a year nodding along in AI meetings before admitting I did not actually know how a language model works.