Alignment Is Not Unsolvable, It Just Lacks Algorithmic Research, Practitioners Say

A practitioner and the MillionInt team argue that AI alignment is not intractable but an algorithmic problem largely abandoned by the ML community. Using Asimov's Three Laws as an example, they note the challenge lies in computing gradients for goals like 'do no harm,' which RL cannot directly optimize.

2026-09-14 ~ 2026-09-14 · 2 related posts

• Alignment Is an Algorithm Problem: Why RL Can't Optimize "Don't Harm Humans" — MillionInt · 2026-09-14
• 'Alignment is an algorithmic problem the ML community stopped working on' — hughbzhang · 2026-09-14