We scan new podcasts and send you the top 5 insights daily.
Tom Milton says recomputation requires the verifier to access model weights inside the prover's data center, which labs dislike. Zero-knowledge proofs would let provers demonstrate correct inference without sharing weights, though current ZKML is orders of magnitude slower; sampling one in a thousand packets could suffice.
Tom Milton describes Emoto's prototype on about 32 H100/H200 GPUs: a network tap copies traffic to a verifier node, which reruns the claimed inference and checks outputs match. Unexplained packets reveal hidden workloads like training. This avoids inspecting GPU instructions directly.
Formally verifying the vast weights of complex neural networks or multi-agent architectures is practically impossible. Instead, developers can use formal methods to construct provably correct external runtime monitors that constrain agents to safe operational boundaries and handle recovery when violations occur.
Tom Milton argues verification is not only for outside regulators. If AI models gain nation-state-level cyber capability, labs need hardware-backed logs to confirm models have not changed their own weights or spoofed evaluation results. The same stack protects against untrusted AI as against an adversary country.
Tom Milton says a network certifier is an FPGA that stores a cryptographic log of traffic, enabling zero-knowledge proof that a cluster only serves inference. But FPGAs must be designed, security tested, red teamed, manufactured, and trusted by all parties. If a deal arrived next month, the hardware would not be ready.
Tom Milton says the fallback is turning data centers off, but that is extremely costly. Verification lets a prover show it is serving only inference, or running small safety experiments, so a slowdown can preserve useful work while stopping dangerous training. This reduces the cost of reaching a deal.
Instead of relying on legal service level agreements (SLAs), Bittensor subnets use validators that programmatically check miner performance. They can use ZK proofs to verify an AI model's output, ensuring quality and honesty without trust.
Instead of supervising an AI's hidden thought process, we can demand it produces a 'certificate of reasoning'—a checkable proof—along with its output. This could include citations or sensitivity analyses, shifting verification from observing the process to checking the provided proof.
Confidentiality concerns make AI companies hesitant to admit external human auditors into proprietary codebases and training runs. Highly capable AI auditors whose internal memory states are securely wiped immediately after compiling redaction-adjusted compliance reports can verify safety without leaking intellectual property.
Tom Milton says a government could require a zero-knowledge proof that a safety classifier is running on each API response. Then it could send random requests and check the proof remains attached, without inspecting the lab's hardware. This gives a correctness claim, though not proof no illegal model runs elsewhere.
Nikesh Arora argues enterprises must inspect a model's chain of thought to explain decisions to boards and auditors. Only open-weight models currently allow that, while the free argument fades because compute remains the real cost. Auditability is a stronger enterprise wedge than license price. That is the real pitch.