AI Safety Research

We are researching the Why, What and How to design and deploy Safe AI Systems.

Why building AI Safety is hard ?

We did not engineer AI to behave in the way it is behaving today. These capabilities emerged from massive neural networks trained on enormous datasets. We are exploring the business impact of reward misspecification and goal misgeneralisation.

Why Evals cannot be trusted ?

Evals are still a nascent field. They lack the scientific rigour thus they merely allow us to reduce uncertainty. We are exploring the true robustness that will complement Evals.

How Neural Networks learn?

Neural Networks do not learn from Prompts. Prompts are merely human attempt to guide a Model. We are therefore exploring Features and Circuits that connect the input to it behaviour.

What AI Controls bring to table?

AI Controls are temporary solutions. Super Intelligence can produce outputs which are too complicated for humans to evaluate. We are exploring the right complement for AI Controls towards truly safe AI.