一系列关于Mastodon的帖子主张改变AI模型的审计和验证方式。作者认为，由于模型漂移和重新训练，传统的年度审计不足以应对，因此提倡使用加密溯源和密封AI决策过程的记录。这种方法，特别是使用后量子密码学，被认为是证明模型行为的关键，并且正在成为保险定价等领域的重要因素。 AI
影响 提出了一种新的AI验证标准，可能影响AI系统的保险和监管方式。
排序理由 该集群由讨论AI审计和验证方法的观点文章组成，而不是特定的发布或事件。
在 Mastodon — mastodon.social 阅读 →
AI 生成摘要 · Google Gemini · 来自 5 个来源。 我们如何撰写摘要 →
Continuous Audit or No Audit at All I think the annual audit is already a fiction. Models drift and get retrained between visits, so a yearly stamp certifies a system that no longer exists. I argue assurance must be sealed at execution, or it is nothing. https:// mickai.co.uk/art…
The Validator Economy Is a Verification Market, Not a Mining Race We have been pricing the wrong thing. I want validators paid not to order transactions but to prove a machine acted within its authority. That is the market I am building, and here is why it changes who gets paid. …
Adversary Models in Friendly Clothing I have learned to distrust a model by how friendly it looks. The handle fits your grip while the steel traces back somewhere else. I argue cryptographic provenance over the whole lineage is the only border that holds. https:// mickai.co.uk/ar…
When the Inference Itself Must Carry a Signature I think most signed-AI is signing the wrong thing. We seal the words and leave the thinking deniable. My argument: bind the whole decision event, model, inputs and authority, under post-quantum strength, or you have proven nothing.…
Your Insurer Will Ask to See the Record I have watched AI cover get repriced on a single question: can you prove what the model did. My view is that a sealed, post-quantum audit record is now an actuarial input, not a compliance nicety. https:// mickai.co.uk/articles/your-ins ure…