Building with AI

AI is core to what we build and how we work. Leveraging it isn't optional and we expect everyone to use it aggressively. But power without judgment is a liability; human judgment is more critical than ever. This is how we think about using AI well.

AI is powerful, and it's never perfect

AI will never be 100% accurate, even on easy tasks. That's not a reason to avoid it, it's a design constraint. We build fault tolerance into everything we do, and we assume failure modes exist whether or not we've found them yet.

Know which context you're in

Some systems must be absolutely correct. Others can tolerate faults. The two demand completely different levels of human oversight, and the most important thing you can do is know which one you're working in at any given moment. Be deliberate about where each use case falls, and design the level of human checking to match.

You still own the decisions

AI augments your thinking, it doesn't replace it. It's no substitute for understanding how your own system works: if you can't explain what your system does without the AI, you don't understand it well enough. And you own what you ship. Be prepared to justify why you chose the right risk/reward tradeoff for your application. The decision to trust an output, and at what level of oversight, is yours.

Measure everything you can

Quantify AI output wherever possible. Evals are the only way we continuously improve the system, so they aren't an afterthought, they're how we know if we're getting better. If it matters, measure it.

Master your tools, keep improving

Understand your AI tools like any other tool of the craft, and always work to get better with them. The technology is developing fast, so we stay flexible and continuously evaluate where we are and where we need to improve. We're all learning here, and we learn faster together than alone: share what works, share what breaks, and trust each other's judgment. We encourage experimentation, and we trust you to exercise good judgment about when and where.