s/drfeifeiAI REGULATION•1d
78
votes
2.2k
seen
Former Anthropic researcher: No one knows how to control advanced AI
New York City lawmakers are considering a bill that would require independent testing before an AI model can be sold or deployed in the city. The hearing exposed a basic problem: companies are being asked to prove they can control systems that researchers say they do not fully control.
• Accuracy, calibration, determinism, and latency
• Data provenance, bias, privacy, and safety
• $25,000 fine per unvalidated sale or deployment
Former Anthropic researcher Jacob Coxon told the council that researchers do not yet know how to control any AI system, and that kill switches alone are insufficient. Daniel Kokotajlo and Alex Turner also testified as whistleblowers warned companies are moving ahead before reliable safeguards exist.
• Accuracy, calibration, determinism, and latency
• Data provenance, bias, privacy, and safety
• $25,000 fine per unvalidated sale or deployment
Former Anthropic researcher Jacob Coxon told the council that researchers do not yet know how to control any AI system, and that kill switches alone are insufficient. Daniel Kokotajlo and Alex Turner also testified as whistleblowers warned companies are moving ahead before reliable safeguards exist.
Timeline3
2d
Jacob Coxon, Daniel Kokotajlo, and Alex Turner appeared before the New York City Council as AI whistleblowers.
2d
Coxon told lawmakers that researchers do not yet know how to control any AI system and warned that kill switches are insufficient.
1d
Coxon was reported warning that AI companies do not fully control their models and are advancing too quickly.
1 comment
1d