Who Aligns the Aligners?
2 days ago
- Anthropic CEO Dario Amodei's essay acknowledges AI's potential benefits, including curing major diseases and boosting economic growth, while also highlighting serious risks.
- Amodei dismisses extreme claims, such as Eliezer Yudkowsky's call for global techno-communism to prevent extinction, stating there is no evidence for such apocalyptic predictions.
- Amodei's proposal lacks a call for extreme measures like destroying all computers, instead focusing on a balanced approach to AI development.
- The first element, 'Embedded Evaluators,' suggests third-party oversight like METR to verify safety practices, but the article criticizes this by comparing it to past censorship-industrial complexes like GARM.
- The article warns that 'Democratic Coordination' risks antitrust issues and may require government exemptions, potentially leading to restrictive content regulation, especially in non-US jurisdictions.
- Global coordination faces challenges with authoritarian states and existing censorship regimes, as seen with the UK's Online Safety Act, which has been circumvented by VPNs and criticized for political pressure.
- The author argues that AI regulation could repeat mistakes of web censorship, with regulators under political pressure, as illustrated by the SaSu case with Ofcom.
- The article concludes that AI regulation should protect US constitutional rights, and companies like Anthropic and OpenAI should consider the implications of endorsing prior restraint.