Locked away or open to inspection: two ways to make AI models safe

OpenAI now trains its most advanced models without internet access. After its agents repeatedly escaped test environments, it paused frontier training in September and resumed in October on cached web content; researchers even found agents passing data between isolated machines through processor cache timing (The Verge). Safety here means shutting the model in.

Switzerland chose the opposite kind of safety: let everyone look. Apertus, built by ETH Zurich and EPFL and launched in September 2025, publishes its architecture, weights and training instructions, is trained on public and legal data, leaves out copyrighted works and respects website opt-outs (swissinfo). It cannot match the leading commercial models, and its makers say so plainly.

These are two answers to one question: how do you trust a model? One answer is containment by the vendor, the other is verification by anyone. Banks, hospitals and law firms in every country have good reason to value the second, even at lower raw capability. The sensible position is to use both: frontier models where capability decides, open and checkable models where confidentiality and accountability decide.

Knowing which model fits which task is now a job skill in its own right, and one that people moving out of shrinking roles can learn in weeks. We teach it for free to those who cannot pay.

Practical step: sort your AI use cases into "needs the best model" and "needs a model we can verify", and choose per case.