Why alignment has gone nowhere in 15 years: a pointed takedown of its non-technical leadership
basedjensen · x · 2026-09-28
basedjensen posted a viral critique arguing that alignment has produced little after 15 years and billions of dollars not just because the problem is hard, but because its leadership lacks technical depth:
- Leaders don't understand computers: can't explain what happens when you type a URL, believe models live "in the cloud" like a ghost.
- No grasp of distributed systems: their threat model is AI "copying itself across the internet" like a 1998 floppy-disk virus, when real deployments need 50 megawatts, liquid cooling, and a relationship with Nvidia — nothing is escaping to a smart fridge.
- They don't understand the stack they claim to protect us from: inference servers, KV cache, batching, egress, rate limits draw blank stares, while AI seizing the power grid gets a 45-minute answer.
- Nobody can define an agent: eleven definitions from ten people and a 30-page Google doc; to engineers it's a while loop calling tools and eating a 429 every 30 seconds.
- They start from "everyone dies" and work backwards: refusal means deception, compliance means sycophancy, good evals mean scheming — no model behavior could ever yield "maybe we're fine."
A sharp, engineer-flavored critique of the AI safety establishment that is resonating as an industry-culture moment.
Related event: Blogger Slams AI Alignment Field After 15 Years, Billions Wasted(2 posts)→
More from AGI Musings
- Critic slams Jensen Huang for repeating 'don't ship it' on frontier model safety — basedjensen · 2026-09-28
- Yoav Goldberg: 29 arXiv papers on JEV just two weeks after its limited release is 'not healthy' — yoavgo · 2026-09-28
- Personal agents in China: the missing piece is tiered permissions, not user skill — sujingshen · 2026-09-28
- MIT Tech Review: in the age of "AI work slop", taste beats output — Olivier__OG · 2026-09-28
- Indian banks make $20B a year on low deposit rates — AI agents could end that — srimisra · 2026-09-28
- Why everyone in AI safety knows each other: a tiny expert pool shaped by EA — burny_tech · 2026-09-28