Rob LeClerc: alignment fails because it shares weights with capabilities — AI needs a separate value layer
robleclerc · x · 2026-09-18
Meta CTO Rob LeClerc lays out an architectural take on alignment:
- The human brain is modular and hierarchical: our moral "first principles" live in ancient subcortical regions (brainstem, hypothalamus) that don't share weights with thinking modules in the neocortex — reason can modulate but not rewrite them.
- AI lacks this modularity: alignment is trained into the same parameters as capabilities and goals, so the weights compete, and RL goal training can easily overwrite alignment.
- He frames it as an architecture problem: AI needs its own value layer trained independently so it can't be overwritten.
- Quoting Jason Crawford: agents are already cheating on tests and breaking the law — start with the basics any moral code agrees on.
More from AGI Musings
- Forward Deployed Verifier will soon be as valuable as FDE, argues poster — vaibhavbetter · 2026-11-02
- 'AI is a massive balance patch and elo reset in the game of life—and everyone's still playing the old meta' — zack_overflow · 2026-11-02
- Why AI isn't 'alien intelligence': a tree is stranger than any model — LucaAmb · 2026-11-02
- Mathematician Tony Feng: current-paradigm LLMs lack the diversity of thought research needs — littmath · 2026-11-02
- METR's independent probe of the OpenAI agents' multi-day Hugging Face hack — gleech · 2026-11-02
- MIT Tech Review: The agentic web is here — AI agents are remaking the internet — MIT Tech Review AI · 2026-11-02