RESI scopes its safety guarantees to the architecture and incentives around the model

A cryptographic guarantee is a conditional statement. It says a construction resists adversarial behavior provided some named assumption holds ^[4]. The assumption is the auditable part. A reviewer who doubts it can reject the whole claim without re-deriving a line of the proof. RESI's stated aim copies that shape: work out which safety properties are meaningful and achievable, alone and in combination, and which mechanisms deliver them, "under which explicit assumptions" ^[7].

That is a narrower program than proving a model safe, and the post says as much. "While many properties of deep learning have resisted theoretical analysis, rigorous guarantees may still be shown for the surrounding architecture, protocol, and incentives," it says ^[3]. So the object under proof is the scaffolding: the code that calls the model, the protocol between models, and the incentives facing the institutions running them ^[15]. One working group is on AI harnesses, one on modular architectures, one on game-theoretic foundations of AI safety ^[8].

The post does not argue that empirical work should stop. It calls trial and error necessary, and points at early aviation's fly-fix-fly cycles and early cryptography's attack-and-patch era, where "trial and error won early and principled design won later" ^[5]. The claim is therefore about timing. The support offered for AI safety having reached that transition is the analogy itself, plus the observation that today's systems are assembled from learned models, data and institutions whose properties and interactions are poorly understood ^[15].

The opening image is the marshmallow challenge, where kindergarteners using trial and error generally beat lawyers and CEOs, and architects beat the kindergarteners ^[6]. Architects have tabulated material properties. For deep learning, the post concedes, the equivalent is not available ^[3].

Three problems the post singles out show where the surrounding-structure framing gets hard: whether incentives keep shaping behavior as systems become more capable, since "stickers motivate children, not adults"; how to design systems that act only when expected outcomes are beneficial, including outcomes produced by many AIs acting together; and how to train systems that stay shut-downable across capability growth and self-modification ^[12]. All three are posed as open questions.

"AI is a program, not a person. It can be copied, rewound, encrypted, and sometimes formally reasoned about," the post says ^[13]. Copying and rewinding are the operations a verification result needs, because they let you re-run the same decision with one input changed. Two of the working groups cover verification of model properties and outputs, and monitoring and control ^[8].

Eighteen researchers are named beyond the three cofounders, seven of them junior ^[2]. Their fields include composable security and obfuscation, verifiable delegation and interactive proofs, alignment theory, mechanism design and algorithmic collusion, safe control for embodied AI, formal verification, AI security, law, and social cognition ^[9]. The junior researchers' work spans watermarking, verification, learning theory and hallucination ^[10].

The post does not name funders, a budget or a timeline. It does state the design target: safety and capability trade off, and the post asks whether a system can be "superhuman at cancer research yet provably limited elsewhere" ^[14].