Breakthroughs do not always come from within a field. Sometimes they come from people who are not bound by its assumptions and habits of thinking. Because of that distance, outsiders occasionally notice possibilities that insiders overlook. One famous example is George Green, a self-taught miller who, in 1828, introduced what later became known as Green’s functions — ideas that eventually became fundamental across physics and engineering.

That possibility feels especially relevant today in the context of AI alignment. As AI systems become more capable, the gap between what we can build and what we can reliably control keeps growing. I believe, that at this point, we cannot afford to ignore even the most unconventional approaches.

A friend of mine, Prof. Michael Zibulevsky (Technion – Israel Institute of Technology), has been thinking deeply about this problem for several years. His main field is optimization but alongside his academic work he has independently started to think about AI alignment. This intellectual exploration has resulted in an unconventional approach towards alignment that has been refined over multiple years.

Recently, his approach started generating surprisingly active discussions in informal circles. That obviously does not mean the ideas are correct, but sometimes strong engagement is at least a signal that something genuinely useful is being explored.

At this stage, honest feedback from people familiar with AI alignment or related areas would genuinely help. Even a short high-level feedback could clarify whether this direction is worth pursuing further. The article looks rather long but indeed it is not, since most of it is the appendix. Feel free to take a look at it here:

Moti’s Journey: Growing Aligned Superintelligence from Infancy