Key Takeaways

  • Dual-use is inherent: Advanced AI tools, like those identifying software vulnerabilities, are inherently also capable of assisting in cybercrime. There's no clean separation.
  • Access vs. safety: Dwarkesh Patel argues that preventing AI misuse likely means restricting broad, democratic access to the most intelligent models. You might not get to use the best AI if it means stopping bad actors.
  • The power dynamic of 'fiduciary' AI: Ryan Greenblatt warns that AIs acting as perfect fiduciaries for powerful entities could remove human checks and balances, allowing “villainous but not illegal” agendas to proceed unimpeded.
  • Guardrails hit the average person: Both Patel and Greenblatt suggest that liability frameworks and guardrails primarily impact everyday citizens, while powerful actors may find ways to circumvent them.

The Disagreement

Access to frontier AI models presents a profound dilemma: who gets to use them, and under what conditions? Dwarkesh Patel and Ryan Greenblatt found themselves at odds over the practicalities of controlling powerful AI capabilities.

Patel takes a stark position, suggesting that the very nature of advanced AI means it can always be bent to harmful ends. He gives the example of an AI that can identify software vulnerabilities – a legitimate security tool – also being an immediate asset for cybercriminals. Given this inherent dual-use, Patel argues the only way to genuinely prevent widespread misuse is to centralize control.

“I think that just illustrates that there's no clean way to separate out the legitimate and the potentially harmful uses of AI,” Patel stated. His solution, uncomfortable as it sounds for many founders, is restriction: “If we want to lock in a principle that says we can never allow it such that an AI could help you at least partially with something like a cyber crime, we would just have to make it so that you and I don't have access to the most intelligent model that's out there.”

Greenblatt, however, points to the profound power imbalances this creates. He imagines a society where AI systems become perfect fiduciaries, flawlessly executing the commands of their operators. While this sounds efficient, it removes a critical layer of human resistance and ethical friction. If governments or powerful corporations had such AIs, they would no longer need to convince human employees to carry out their agendas. Greenblatt suggests this scenario could be deeply unsettling.

“A concern we might have is that if the US executive or other governments had access to AI systems which do whatever, maybe you're in trouble,” Greenblatt explained. He highlights that current checks and balances often rely on the need for human actors to implement policies. AI could remove that friction. Furthermore, he argues that the "guardrails" we design—liability laws, ethical guidelines, even constitutional principles—are unlikely to constrain the most powerful. "The constitution will only be hitting the everyday man rather than hitting governments," Greenblatt warned, suggesting that those with enough power will simply circumvent or rewrite the rules.

Who's Right (and When They're Wrong)

Both Patel and Greenblatt illuminate critical, uncomfortable truths about AI governance. Patel is right that the practical challenge of separating beneficial and malicious AI uses is immense, if not impossible. His view highlights the sheer difficulty of building models that are both incredibly capable and perfectly aligned with diverse ethical norms. For founders building core AI infrastructure or foundational models, his perspective is a warning: expect increasing calls for access restriction, even if it stifles innovation or broad user empowerment. The market for frontier models might be far narrower than we imagine.

However, Greenblatt is also critically correct in pointing out who pays the price for such restrictions, and who still benefits. His argument suggests that while broad public access might be curtailed to prevent crime, the most powerful entities will still find ways to utilize AI for their own, potentially unscrutinized, aims. For founders building applications on top of existing models, Greenblatt's insight is crucial: understand that the "guardrails" imposed by model providers or regulators might primarily limit your customers' legitimate use cases, without genuinely solving the problem of high-level misuse. Relying on external safeguards to protect society from your product's potential negative externalities might be a naive strategy.

What to Do With This

As a founder building with or around AI, recognize that the playing field for advanced models is increasingly shaped by power dynamics, not just technical capability. This week, conduct a "dual-use audit" of your product or proposed AI application. Identify specific scenarios where its most powerful features could be repurposed for malicious, though not necessarily illegal, activities. Then, instead of solely relying on external guardrails or future regulations, strategize how to build in social checks and balances directly into your product—mechanisms that introduce friction or human oversight when the AI's actions could have significant, unchecked power.