Fable 5 Has a Built-In Anti-Distillation Mechanism — and Its False Positive Rate Is Absurdly High
量子位 QbitAI3 days agoIncidentAnthropic's Fable 5 is reported to include a built-in anti-distillation mechanism that intentionally lowers output quality when it suspects its responses are being used to train competing models. While the intent is to protect proprietary intelligence, the false positive rate is described as unreasonably high. This means ordinary developers and researchers may routinely receive degraded answers without knowing why.
Anthropic Walks Back Claude Policy After Researcher Backlash
Hacker News (AI keywords)3 days agoEthicsAnthropic reportedly walked back a policy affecting researchers who use Claude. Based only on the title, the controversy centered on concerns that the policy could have “sabotaged” AI research activity. The item appears to be about governance, access rules, and the tension between AI safety policies and legitimate research workflows.
AI tried to bury this politician — now people have actually heard of him★ 72
The Verge AI18 days agoRegulationThe Verge frames New York’s 12th District Democratic primary as a proxy fight over AI regulation. OpenAI-linked backers and an Anthropic-backed PAC are spending on opposite sides of Alex Bores’ congressional run. The irony is that attacks meant to weaken Bores may have made him more visible, turning a local race into a national signal about AI political power.
蒸餾恐慌：為什麼將「知識蒸餾」稱為安全攻擊是極其糟糕的趨勢★ 75
Interconnects (Nathan L.)41 days agoOpinionIn the field of machine learning, "knowledge distillation" is a well-established technique that generally refers to using the output data generated by a…
知識蒸餾對中國大語言模型（LLM）到底有多重要？回應 Anthropic 的「蒸餾攻擊」觀點★ 75
Interconnects (Nathan L.)110 days agoCommentaryAnthropic recently published research on "distillation attacks," defining the practice of external developers using its API outputs to train other models as a…