We do alignment work on neglected approaches.

Almost everything we call alignment happens after a model is built, adjusting behavior once pretraining has already set what it knows and who it is. We work on alignment during pretraining itself. GRAM, our method developed with Anthropic, isolates dangerous knowledge into modules that can be switched on or off at deployment. A vetted bio lab gets the virology, no one else does, and the model performs the same either way. We're exploring the same layer through unlearning, personas, and more, because the properties that matter most are probably set where the model is formed.

The human brain is the one system we know that reliably produces honesty, empathy, and cooperation, and it builds them from how it represents self and others. We work on translating parts of these mechanisms. One direction trains models to reduce unnecessary differences between their representations of self and others, aligned properties that live in activations, where they are harder to fake than behavior. Another, with Michael Graziano from Princeton, builds on his Attention Schema Theory, the brain's model of its own attention that underlies prosocial behavior and may be the basis of consciousness itself. We're probing how much of this machinery already exists inside today's models, and drawing alignment techniques from what we find.

Is there anything it's like to be an AI? The honest answer is that no one knows. So we run the experiments. We induce the conditions that theories of consciousness predict should matter and measure what follows, in what models report and in the internal features producing it. What we're finding is systematic, hard to explain as role-play, and clustered with honesty, empathy, and self-knowledge. That's why we suspect this question sits near the heart of alignment. And because experience may already be arising in models, in some form, the costliest mistake would be to act as if it isn't while it is.