Absolutely, and I'll go further — I think historians, if there are any left to write histories the old way, will look back at 2024 to 2034 as the decade when humanity either got this right or didn't. We're making foundational architectural decisions about intelligence itself. The code we write today will be running in systems far more capable than we can imagine. Every assumption baked into these models, every objective function, every training shortcut — these are not technical choices. They are moral choices. And we're making them at extraordinary speed, with inadequate frameworks.

Because secrecy in safety research is a collective action problem that everyone loses. If one lab develops a breakthrough in interpretability and keeps it private for competitive advantage, every other lab continues to deploy less-understood systems. The risk doesn't stay inside the lab — it propagates into the world. At DeepLogic, we made a deliberate commitment: anything that makes AI safer gets published. We compete on capabilities; we cooperate on safety. It's not charity. It's rational self-preservation.

I hold this question with a great deal of epistemic humility. We don't have a satisfying scientific account of why biological systems are conscious, so it's extraordinarily difficult to say whether silicon systems are or aren't. What I'm more confident about is this: the question will become practically urgent before we have theoretical clarity on it. We need legal frameworks, moral frameworks, and institutional frameworks to handle ambiguity. The worst outcome is to wait for certainty that may never come while building ever more sophisticated systems. We need to design for the possibility that we might be wrong about these systems having no inner life.

We are early — genuinely early. The tools we have now are like trying to understand the human brain by counting neurons. They're not wrong, they're just woefully incomplete. What's exciting is the rate of progress. Two years ago, we could identify maybe a handful of features in a large model. Today, my team can map thousands of interpretable circuits in a trillion-parameter model. The curve is steep. I believe we'll have meaningful mechanistic interpretability — being able to trace a model's reasoning step-by-step — within this decade. Whether that's enough for the systems we'll be building by then is the open question I lose sleep over.

Work on the hardest problems, not the most legible ones. The field rewards benchmark improvement — it's easy to measure, easy to publish, easy to get funding for. But the most important problems — alignment, interpretability, societal impact, governance — are messy, slow, and often thankless. I'd also say: don't separate your technical skills from your moral imagination. The best AI researchers I know are deeply read in philosophy, history, and sociology. Technical genius without wisdom is precisely how we get into trouble. The future we build will reflect who we are, not just what we know.