Why the RESI acronym?
IRS was already taken, and RESI (for REsponsible SuperIntelligence) is also the beginning of RESIlience.
What does success mean for RESI?
The ultimate success would be new safe-by-design approaches adopted in frontier AI systems: breakthroughs comparable to foundational cryptographic primitives. Progress toward that goal includes rigorous definitions of safety objectives, maps of what is and is not achievable under explicit assumptions, new mechanisms with useful guarantees, results on how guarantees compose, and working demonstrations in real systems. Success also means influencing policy and regulation in ways that support human flourishing.
How will RESI’s ideas become adopted?
Through design, people, and incentives. RESI designs with practical constraints in mind and tests promising constructions in real systems. It is connected to frontier labs through the experience of its researchers and through regular visitors and speakers, allowing ideas to travel with the people who carry them. Meanwhile, frontier models already face release barriers due to safety considerations. As capabilities increase, developers will face stronger safety, reliability, and deployment requirements. Approaches that provide greater assurance while preserving usefulness can reduce deployment risk and offer incentives for adoption.
How is RESI different from other AI safety initiatives?
RESI focuses on the formal and design-oriented foundations of AI safety: precise safety objectives, explicit assumptions, mechanisms and architectures with analyzable guarantees, composition of those guarantees, and impossibility results. The goal is to develop approaches intended to remain meaningful as systems become superintelligent. This complements work on evaluation, interpretability, alignment training, control, and safeguards for existing models. RESI is also distinctive in who is building it: it is led by researchers with a track record of turning vague concepts into precise definitions and usable mechanisms, from cryptography and other fields, and it is organized by those researchers.
How will RESI stay up to date with frontier research?
RESI combines researchers with recent frontier-lab experience, regular engagement with researchers at leading labs, a continuing visitor and lecture program, and practical testing of promising ideas on current systems. Its research agenda will evolve as capabilities, architectures, and deployment practices change.
Is this approach fast enough?
It has to be, and there is reason for optimism. Foundational fields can move quickly once the right people converge on the right questions, as happened repeatedly in modern cryptography, and RESI is built to concentrate that process and accelerate it with frontier AI tools. Nor does progress depend on one all-or-nothing breakthrough: definitions, impossibility results, and mechanisms can inform current systems along the way. Foundations take time to mature, but they are hardest to build once a crisis has arrived. The best time to begin was long ago; the second best is now.
What does superintelligence mean?
Superintelligence means systems that greatly exceed individual humans, and potentially expert teams, across a broad range of cognitive tasks.
How does superintelligence differ from AGI?
Artificial General Intelligence usually means matching humans across a broad range of tasks; superintelligence means substantially exceeding that level. Current AI is jagged, already superhuman in some domains and subhuman in others. RESI’s practical concern begins when systems exceed our ability to test and oversee them, because past that point safety must be designed in. In some domains, that point is already here.
How can science study something that doesn’t yet exist?
The principles and constraints that any system has to satisfy can be studied and built upon. Much of computer science and engineering falls within what Herbert Simon called the “sciences of the artificial,” which ask not only how existing systems behave, but what systems can be designed. Modern cryptography did not emerge merely from studying the ciphers of the time, but from analyzing what can and cannot be done.
How can I get involved?
For other scientists and practitioners local to the Boston/Cambridge, sign up for notifications on our reading group, events, and seminars here. RESI will host a small number of researchers and visitors from outside the area. We welcome inquiries from scientists whose tools could bear on these problems, whether or not they have worked on AI safety before. Email us with your ideas at hello@resi.org. If you are a local student, we consider interns from our partner organization, Cambridge/Boston Alignment Initiative (CBAI).
