The Quest to Verify AI 尋求驗證人工智慧協議

This presentation explores "AI Verification," an emerging cryptographic and technical framework designed to resolve geopolitical deadlocks and trust deficits between global powers like the US and China. Rather than relying on traditional physical inspections or blind trust that compromise national security and proprietary IP, it leverages cryptographic solutions such as Recomputation and Zero-Knowledge Proofs (ZKP). These methods mathematically confirm model authenticity and facility intentions (training vs. inference) without exposing underlying models, training data, or infrastructure.

本簡報探討新興的「人工智慧驗證」（AI Verification）密碼學與技術框架，旨在化解美中等全球超級大國之間的地緣政治僵局與信任赤字。此架構捨棄了傳統需交出底層機密代碼與基礎設施的實地視察模式，轉而藉由「重新計算」（Recomputation）與「零知識證明」（ZKP）等密碼學手段，在完全不開啟「黑盒子」的前提下，以純數學方式客觀證明 AI 模型的真實性及資料中心的運作狀態（訓練或運行），實現兼顧國家安全、專利保護與跨國監管的治理目標。

Transcript