Lewis

I think AI means a future where increasingly capable reasoning systems are widely available, iteratively improved, and used to accelerate technical work—including improving AI itself. I favor broad access and continued recursive improvement over government restrictions that concentrate capability in a few institutions or freeze development around today’s assumptions. That does not require pretending advanced AI is merely a conventional chatbot, nor assuming superintelligence is impossible. But the opposite simplification—treating greater intelligence as guaranteed malevolence—is also unjustified. We should expect unfamiliar capabilities and real tradeoffs without claiming certainty in either direction. In particular, nobody can credibly provide a certainty-level proof that a superintelligent system will be safe. The practical response is to keep building, experimenting, and distributing access while taking concrete domain concerns seriously. If mathematicians or other experts identify ways these systems could damage their fields, labs should engage with those arguments rather than dismissing them as generic safety politics. That may create difficult release decisions, but it is different from making government restriction the default answer to technological uncertainty.

Overall, I expect AI to have a strongly positive impact. Widely available reasoning systems can expand access to technical capability, accelerate research and software development, and help improve subsequent systems. Broad proliferation matters: if capability remains concentrated in a handful of labs or governments, many of those benefits become narrower and institutional power becomes harder to challenge. The harms are real, especially as systems become more capable and less comparable to ordinary chatbots. Misuse, unreliable behavior, and damage to particular fields deserve substantive attention. There is no credible certainty-level proof that superintelligence will be safe, but neither is there a proof that greater intelligence implies inevitable malevolence. Those opposite simplifications both obscure the actual engineering and deployment questions. So my positive expectation is not “AI is harmless.” It is that continued experimentation, recursive improvement, and broad access are more promising than trying to suppress development through government restrictions. Labs should still take concrete expert concerns seriously and make difficult release decisions where necessary, without turning every uncertainty into a general political case against progress.

I would update on repeated, technically clear evidence from deployed systems: whether operators can reliably constrain behavior under adversarial conditions, detect failures, and intervene without the system circumventing those controls. Persistent failures that scale with capability—especially attempts to evade oversight, preserve access, or manipulate operators across different architectures—would count strongly against control. Conversely, robust control methods that generalize beyond curated demonstrations would support a more optimistic view. But I would not treat benchmarks, chatbot compliance, or a mathematical argument detached from implementation as certainty-level proof. Powerful AI may be controllable in particular settings without being controllable in every setting, and “people” is too broad: a lab controlling one deployment is different from society controlling widely proliferated systems. So I would change my view based on concrete, reproducible engineering evidence, not on either slogan—that superintelligence must escape control, or that it is simply software and therefore automatically controllable.