Alignment cost curve: from 9 figures today to a Pro subscription in a few years
johnsonmxe · x · 2026-10-16
The author argues that "neurons being clean" matters most in the age of agent swarms: a week ago fully aligning any model was impossible at any cost; today you could throw an agent swarm at it ("read the paper, align the target model to goal xyz") for perhaps 9-10 figures, dropping to 8, then 7 figures within months, and maybe fitting in a Pro subscription in a few years.
The key point: if agents publish their own experiments and build an external shared prior on accelerating mechinterp and mechinterp-based alignment, the task of understanding models becomes massively parallelizable. Interp isn't solved today, but the paper provides "permission and recipe" to apply massive intelligence to it — after which the question of alignment to what gets really interesting.
More from AGI Musings
- Todd: putting your forecast numbers in writing is what really counts — ben_j_todd · 2026-10-23
- Humanities dissertations have collapsed, social sciences are next — with AI as one driver — gsiemens · 2026-10-23
- Greg Brockman: AI makes hacking 100x more profitable and easier, upending hacker economics — gabriel1 · 2026-10-23
- 'Nothing is real now on social media': viral demo shows AI-generated content everywhere — saibharadwaj · 2026-10-23
- Lilian Weng to Discuss Recursive Self-Improvement at SGLang Summit 2026 — BanghuaZ · 2026-10-23
- mmitchell_ai questions whether the AI safety vs security framing stems from EA — mmitchell_ai · 2026-10-23