Some of it, none of the technical parts. I was on Einstein at Salesforce before
this. Models from that era are museum pieces and none of that code would survive
contact with 2026.

Two things stuck. First, the model is never the bottleneck. It's the data
pipeline, the labeling, and getting the thing running in production without a human
babysitting it. Still true, and still most of why enterprise AI projects quietly
die.

Second, and I didn't appreciate this until much later, compute demand is a physical
business. Someone has to buy the machines, find the power, get the interconnect,
and move the heat. Mining learned that the hard way about a decade before the AI
people had to. It's most of the reason mining companies ended up holding the power
and not the other way around.

I left AI to go mine Bitcoin and AI came and found me anyway.