You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Postdoc work on building affordable zero-knowledge protocols for proving properties of AI training, such as FLOP counts
The goal is to design protocols to enable zero-knowledge verification for frontier AI training (and inference), without requiring too much hardware R&D. The first goal is compute [measured in FLOPs] bounds for pretraining, but there are many extensions after that.
The concrete outputs would be research papers, and formalizations in languages such as Lean and Rocq. My work would mainly be on the theory side, designing and formalizing protocols, studying their properties, red-teaming, etc.
The implementation team will work on running experiments, building prototypes, and optimizing.
The funding will cover my postdoc wage, expenses, and overhead: about 6k USD, resp. 1.5k USD, resp. 0.5k USD, per month, so about 8k USD total per month. The minimum value would thus fund ~4 months, and the maximum ~12 months.
I'm working with the people mentioned in https://gpaipolicylab.org/verification, in particular Pierre Peigné, Ky Nguyen and Dimitris Kolonelos. With Pierre Peigné and Ky Nguyen, we produced a first preprint on the question (https://arxiv.org/abs/2606.05433). Dmitris Kolonelos and Ky Nguyen have substantial experience in cryptography and zero-knowledge proofs, and I expect my work to benefit a lot from discussing with them. I would be based at LIP6, Sorbonne Université, in the ALMASTY team with Ky Nguyen.
The most likely failure mode is finding only robust protocols that are too expensive, or affordable protocols that are vulnerable to attacks. In that case, I expect we will be able to produce a comprehensive analysis of the cost-robustness frontier, possibly leading to formalized impossibility results. But even in that case, actual costs could still be lowered by optimizing the implementation.
About 60k USD, from Coefficient Giving's Career Development and Transition program, though not for this project.