Questioning the assumptions of AI Safety.

One way of viewing AI Safety is through the question: how do we make future generally intelligent systems work for the betterment of humanity and society?

Put that way, it is plainly not only a computer science question. It asks what a general intelligence is. It asks what it would mean for a system to be moral. And it asks what betterment amounts to for a species that does not agree with itself about it. Those are questions about life, about mind, and about value, and they are much older than machine learning.

Meanwhile the work is urgent in a way philosophy usually is not. We are building systems that plan, model themselves, and act in the world, and we are deploying them faster than we can say clearly what they are.

Serious effort goes into keeping them safe: evaluations, oversight, interpretability. Almost all of it inherits a vocabulary of agents, goals, rewards, and optimization, assembled from economics, control theory, and machine learning. Those concepts were built for other purposes, and they strain here. Problems and unpredictabilities arise from this perspective: how do you continuously evaluate a continual learning agent? How do we find the right primitives for describing general agency in the first place?

The research tradition closest to this concern is agent foundations, which treats agency, goals, and reasoning as open problems rather than settled primitives. AFFINE takes its name from the tradition: the Agent Foundations Field Network.

The current safety work might be built on the wrong concepts, and careful work aimed through an unexamined concept can be precise and still miss. We believe that the field is pre-paradigmatic in Kuhn’s sense: it has not yet settled the concepts that would let it tell good questions from dead ends.

We would rather find that out early. If the frame is sound, this upstream work costs a few careful people some years. If it is not, this upstream work is the only thing that helps.

Most research holds its concepts fixed and iterates within them. When the concepts are sound, that is exactly right, and it is how fields make steady progress.

Sometimes research happens on underdefined concepts, which can lead to the degradation of science as people dig themselves into deeper pits than they were in before. If this is the case in AI Safety, it implies that we need to start digging in new places. This is difficult to do: what determines progress for a concept that does not yet exist? That thinking is slow, hard to measure, and easy to get wrong. How would you make this happen?

We believe that running it well requires two things that are difficult to fund: unhurried time, and people who will tell you that your framing is wrong. So that is what we’re trying to build. A setting where existing concepts can be taken apart carefully, by people who have read enough to know what has already been tried.

AFFINE grew out of a month-long residential seminar held in Czechia in May 2026, and is now being built into something lasting: recurring seminars, fellowships, and retreats that bring researchers across fields into the same room for long enough to disagree properly.

The most useful thing an outside researcher can do for a young field is tell it where it is confused. We mean that literally, and we have designed the invitation around it.