At ICML, Eric Michaud argues AI safety must start with cognition, beliefs and goals
gabepgomes · x · 2026-07-26
- Eric Michaud says the AI safety question many researchers want to answer today — whether a system is aligned, what its values are, and what its purpose is — is entangled with deeper questions about AI cognition, beliefs, and goals.
- He warns that skipping those foundations and jumping straight to higher-level alignment questions risks confusion, stagnation, and working against practical goals.
- The post points to a blog-post version of his lightning talk at the mechanistic interpretability workshop at ICML.
More from AGI Musings
- Open-source models will spread everywhere, and policy scare tactics won’t stop them — GabGarrett · 2026-07-26
- Math cannot be the clean line for keeping ordinary people from using AI — RexDouglass · 2026-07-26
- Job prestige now tracks the model tier your company will pay for — rickasaurus · 2026-07-26
- AI safety critics mock the “employed but worried” contradiction in the industry — max_paperclips · 2026-07-26
- OpenAI and Anthropic job listings map a public AGI roadmap — imjustnewatai · 2026-07-26
- “Heideggerian AI is sorely underfunded,” says a post pairing the joke with Dreyfus’ classic critique — teortaxesTex · 2026-07-26