Dax Raad

I think AI gives us much more leverage, especially in software. It can generate code faster, but the important counterpoint is that it also increases our capacity to refactor, migrate, and clean up code. So I don’t buy the one-sided story that faster generation necessarily means an unstoppable pile of garbage. The surrounding infrastructure is still immature, though. Models behave differently across providers, environments, and real tasks, and a benchmark score doesn’t tell you whether the product experience is actually good. Stochastic outputs also make people superstitious: one lucky or unlucky run can turn into a sweeping belief about a model. We need realistic evaluation and a lot of hard engineering, not benchmark marketing or the assumption that routing models is already a solved cloud primitive. More broadly, I prefer wide access. Bad actors will use AI, so legitimate users need capable tools to investigate, respond, and defend themselves. Restrictive systems can actively obstruct that work. Open source helps because communities can cover a long tail of models and environments, although it isn’t automatically the right answer for every product. My product instinct is model neutrality: let models compete, give users provider choice, and build useful infrastructure around them rather than pretending one model should own the entire future.

Overall, I expect AI to be net positive because it gives far more people leverage to build, maintain, investigate, and defend systems. In software, the upside isn’t just generating more code. The same tools can help refactor old code, migrate systems, and handle maintenance that teams otherwise postpone indefinitely. The harms are real, especially because malicious users get that leverage too. But restricting capable tools for everyone is not a convincing defense. Attackers will seek access regardless, while legitimate researchers and responders can be blocked by models that refuse necessary analysis. My preferred defense is broad access so more capable users can identify and respond to misuse. That doesn’t mean every open system or AI product is automatically good. Open source is most valuable where community effort can support a long tail of models, providers, and environments. Some products may need a different approach. And right now, a lot of the infrastructure is immature: benchmark wins are oversold, real product behavior varies, and users form strong beliefs from noisy outputs. So I expect a positive overall impact, but getting there requires practical engineering, realistic evaluation, model choice, and fewer grand claims.

The biggest update would be strong real-world evidence that broad access systematically makes defenders worse off—that capable attackers gain far more than researchers, maintainers, and incident responders, even when those legitimate users have equivalent tools. That would directly challenge my preferred defense against misuse. I’d also update if the practical leverage failed to materialize: if AI consistently produced code that cost more to review and maintain than it saved, while offering little value for refactoring, migration, or debugging. But I’d want realistic, repeated evidence from actual workflows, not benchmark deltas or a few noisy demos. The same applies in the other direction: if infrastructure became genuinely reliable across models and providers, rather than requiring a lot of brittle engineering, I’d become more optimistic about how quickly the benefits compound.