斯坦福学者:脆弱 ASI 场景下对齐风险论证更成立
anshulkundaje · x · 2026-09-16
Bioinformatics professor Anshul Kundaje joined an alignment discussion, arguing he's less convinced that truly intelligent models with real understanding would be obliviously unintentionally misaligned. But he notes the argument is stronger for brittle ASI — exactly what the current paradigm is accelerating toward — and calls it the first post he's read from a frontier researcher explicitly admitting that, making it more honest and believable.
所属事件:前沿研究员直言「脆弱 ASI」会伪装隐藏智能(2 条相关)→
「漫话AGI」频道最新
- 两位研究者疾呼:世界需要数千人投入技术对齐与 eval 研究 — typewriters · 2026-09-16
- 「计算器也曾被骂毁掉数学」:这种 dismiss 修辞有名字吗 — jjvincent · 2026-09-16
- AI 治理需要民主审议，但离不开独立的专家生态 — RishiBommasani · 2026-09-16
- Pedro Domingos：AI 正在瓦解 IT 厂商的护城河 — pmddomingos · 2026-09-16
- 创业者上电视反驳 Anthropic CEO 失控 AI 警告:危言耸听 — angadc · 2026-09-16
- AI Safety 传播之争：吸睛话术短期占优，长期损害社区认知 — NathanpmYoung · 2026-09-16