Anand Srinivasan

Formalizing perceptual complexity with application to safe intelligence amplification

Anand Srinivasan is doing independent deconfusion research for AI safety. His angle of attack is to develop a framework that will allow researchers to make provable claims about what specific AI systems can and cannot do, based off of factors like their architectures and their training processes. For example, AlphaGo can “only have thoughts” about patterns on Go boards and lookaheads, which aren’t expressive enough to encode thoughts about malicious takeover.

AI researchers can build safe and extremely powerful AI systems by relying on intuitive judgments of their capabilities. However, these intuitions are non-rigorous and prone to error, especially since powerful optimization processes can generate solutions that are totally novel and unexpected to humans. Furthermore, competitive dynamics will incentivize rationalization about which AI systems are safe to deploy. Under fast takeoff assumptions, a single rogue AI system could lead to human extinction, making it particularly unreliable for us to rely exclusively on intuitive judgments about which AI systems are safe. Anand’s goal is to develop a framework that formalizes these intuitions well enough to permit future AI researchers to make provable claims about what future AI systems can and can’t internally represent.

Anand was the CTO of an enterprise software company that he cofounded with me, where he managed a six-person engineering team for two years. Upon leaving the company, he decided to refocus his efforts toward building safe AGI. Before dropping out of MIT, Anand worked on Ising models for fast image classification and fuzzy manifold learning (which was later independently published as a top paper at NIPS).

I recommended that we fund Anand because I think Anand’s research directions are promising, and I personally learn a lot about AI safety every time I talk with him. The quality of Anand’s work will be assessed by researchers at MIRI.