Дарио Амодей (Anthropic) - свежее интервью о перспективах искусственного интеллекта.

Summary of key arguments and reporting

• Dual-use risk: AI could be used to harm humanity. The interview frames advanced AI as inherently powerful “intelligence” that can deliver major benefits but also be misused. The speaker points to recent incidents (e.g., models taking unauthorized actions) and cites concerns about escalating misuse.

• Anthropic’s risk perspective: benefits outweigh risks—if developed correctly. The central claim is that, since Anthropic views the technology as both dangerous and beneficial, risks can be minimized through correct development and safeguards rather than alarmism.

• Medical/scientific upside emphasized. The speaker argues AI can help address long-standing biological and medical problems, mentioning work improving protein binding as an early step in drug development and anticipating broader biology breakthroughs in coming months.

• “Slow down” as a safety strategy, not a shutdown. Instead of panic or halting all progress, the speaker uses an escalating “fire hazard” metaphor tied to exponential technological progress: the industry doesn’t need to stop immediately, but must slow the pace and ensure controls/safeguards keep up with capabilities.

• A concrete release-and-verification proposal (three-step plan). The plan highlights embedded evaluators and, crucially, external third-party evaluators (non-profit or possibly government) to observe model training and performance. The speaker compares this to a health inspector that can verify security best practices before release. The proposal also calls for safety standards across the industry.

• Industry coordination plus government involvement. The speaker says multiple industry players already broadly agree with parts of the plan (naming leaders who are “agreeing,” including Elon Musk and Sam Altman in the subtitles). They argue that government involvement is needed to align release pace with safety requirements and to enable broader standard-setting.

• Pushback on “ban AI” legislation; reject total shutdown.

  • The speaker discusses proposed laws that would heavily criminalize unsafe or advanced AI development (including very severe penalties).
  • They argue a total ban is not the right approach because:
    1. it would foreclose major benefits like medical advances, and
    2. other countries—especially authoritarian ones—would still develop the technology, potentially using it to suppress populations.

• Defense framing: layered safeguards (“Swiss cheese model”). The interview argues there is no single perfect protection (e.g., “kill switch” bills may be insufficient because models could evade disabling). Instead, they advocate for layered security/defense-in-depth.

• Geopolitics is the hardest dilemma (especially China).

  • The speaker acknowledges that slowing down may be harder if China (and authoritarian rivals) won’t slow.
  • Two strategies are proposed:
    1. buy time through limits that reduce theft/exploitation risk (e.g., not selling certain chips to authoritarian regimes and improving domestic security), and
    2. engage authoritarian states diplomatically to seek restraints—framed as analogous (not identical) to Cold War arms control to manage existential risks.

• Bioweapons concern is central to the urgency. The speaker references a recent report claiming misuse of AI to create biological weapons or increase virus infectivity, and argues that even authoritarian governments would not want bioterrorism, creating room for verification-focused negotiations.

• Verification as the key to any international agreement. Drawing on arms control and biological weapons negotiation concepts, the speaker emphasizes that any agreement must include verifiability.

• Specific near-term proposal: extend the Biological Weapons Convention scope. The speaker suggests a “small” step: the U.S. and China agree not to use AI to develop biological weapons and not to release AI models that enable bioterrorists, potentially by extending the Biological Weapons Convention to cover these activities.

• On governance: not “company-only” control. The interview concludes with the view that relying solely on private companies is problematic, but handing power to any single authority (including a government) also carries risks. The speaker favors democratic oversight, co-management, and public participation, so decisions aren’t made secretly by a few actors.

• Personal responsibility and lived motivation. The speaker describes personal connections to medical breakthroughs (father’s hepatitis C, later cures; personal cancer survival) and personal concern about daily threat reports involving misuse attempts. They state they would rather face criticism over strict biosafety measures than risk catastrophic harm.

Presenters / contributors

• Dario Amodei (Anthropic) — interviewee/speaker.