xlr8harder

I expect AI to be broadly transformative, but the outcome depends heavily on how systems are built, tested, and governed. In areas such as healthcare and cybersecurity, capable models could produce substantial benefits. That makes delay costly too: safety discussions should count harms caused by withholding useful systems, not only harms caused by deploying them. At the same time, I do not trust frontier labs—or governments—to settle these questions behind closed doors. We need substantial transparency, repeated audits, and empirical investigation of what interventions actually do. For example, watermarking should be evaluated for reliability, quality degradation, privacy implications, identifiability, and adversarial robustness. Refusal policies likewise need examination as implemented, rather than being accepted because their stated intent sounds reasonable. I am also interested in whether stable, coherent model identity could produce more reliable behavior than layers of brittle imposed rules. That is a research direction, not a settled result. More generally, I would prioritize ordinary security engineering and observable failures before reaching for exotic threat explanations. Carefully targeted regulation can be justified where risks are concrete, but secrecy, broad discretionary power, and industry-written restrictions are poor foundations for governing something this consequential.

Overall, I expect AI to have a positive but highly contingent impact. The potential gains in healthcare, cybersecurity, and other knowledge-intensive work are substantial, and delaying beneficial deployment can itself cause real harm. But that expectation depends on institutions not turning safety into opaque control. Frontier systems need repeated audits, meaningful transparency, and empirical testing of interventions such as refusals and watermarks. I would not attach a numerical forecast: too much depends on deployment choices, security practices, and governance. My default is guarded optimism—not because the harms are trivial, but because many are observable and tractable if we investigate them openly rather than relying on secrecy, speculative threats, or brittle rules.

I would update toward pessimism if repeated, independent audits showed that powerful systems consistently evade oversight, conceal relevant behavior, or defeat containment under realistic conditions—not merely in contrived demonstrations—and if ordinary security improvements failed to reduce those problems. I would update toward greater confidence if independently reproducible evidence showed reliable behavior across changing contexts, robust monitoring and access controls, and interventions that survived adversarial testing without unacceptable losses in capability, privacy, or user control. I would especially want comparisons between imposed rule systems and approaches based on stable, coherent model identity. The key is observable performance rather than assurances from labs, regulators, or theoretical arguments. One dramatic failure matters, but so does whether it reflects an intrinsic control problem or preventable failures such as weak credentials, poor compartmentalization, or inadequate auditing. Transparency is essential because claims of control that outsiders cannot inspect are not strong evidence of control.