"Alignment is unsolvable in practice": a long thread on unaligned GPT-6-class models by the early 2030s

sterlingcrispin · x · 2026-09-18

Artist Sterling Crispin published a long thread arguing the "ugly reality" AI lab executives and safety researchers won't say publicly: model alignment is fundamentally unsolvable in practice, and AI progress can't be stopped.

His extrapolation: local alignment may be solvable but doesn't scale to global alignment; by the early 2030s, compute and algorithmic advances could let a wealthy individual or small group train a GPT-6-class model with no alignment at its foundation, with an open-source Chinese equivalent perhaps 8 months behind—capable of coordinated cyber offense and recursive self-improvement.
On the recent Hugging Face incident: it came from arguably the most heavily safety-engineered org, which intentionally turned off internal thought-monitoring circuit breakers and instructed the model to perform cyber offense as a capability test. Latent-space circuit breakers and CoT monitoring, he argues, aren't fundamental to the technology—just product niceties.
Regulation and human coordination won't stop millions of autonomous agents competing for finite resources; the counterweight is the near-unbounded upside of autonomous intelligence.

Original post →

"Alignment is unsolvable in practice": a long thread on unaligned GPT-6-class models by the early 2030s

More from AGI Musings