AI SAFETY
Alignment Is a Human Problem
Coming soon.
Test
The Temporal Pipeline
Every AI system is shaped by human decisions at every stage of its lifecycle — from the capital that funds it, to the researchers who design it, to the annotators who steer it, to the users who reshape it after deployment. There is no stage where humans are absent.
Humans at Every Stage
A model's lifecycle — from capital allocation to deployment feedback loops — is shaped by human choices at every step. There is no stage where humans are absent.
Pre-training
Training
Deployed
Capital → Lab Matching
Regulators — could constrain this stage, often don't
VC Fund A
VC Fund B
Sovereign $
Lab αfrontier lab
Lab βstartup
design choices
does not exist yet
Architecture, loss fn, objective — all chosen by humans before a single gradient is computed.
Researchersloss function
constrain
training objective
shapes
what model can become