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 demonstrates an inductive loop around a GPU power supply feeding an oscilloscope. A forward pass, backward pass, and break are visibly distinct; tiny spikes at each layer even reveal a 12-layer model. At data-center scale, synchronous training produces large, canonical power swings.
Tom Milton estimates labs have single-digit verification staff and maybe 50 people globally work on compute verification. He contrasts this with Project Vela, where the US spent billions in today's money on seismology and satellites to detect nuclear tests. A talented researcher could quickly become one of the field's most useful people.
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.
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.
Tom Milton compares model weights to chemical-industry IP. The OPCW's mass spectrometers have closed modes that return pass/fail rather than full spectra, protecting confidential formulas. He suggests AI verification needs similar systems that confirm treaty compliance without revealing weights or tokens.
Tom Milton shows a Fable-designed PCB that measures GPU power and works, but has curved traces, odd component placement, and a rat's-nest layout. The AI could do a task that took an engineer 10 to 30 hours, but a human would need to rework it. AI hardware time horizon is 10-100 hours at a quality trade-off.
Tom Milton had Fable access cut off overnight and says he lost a skill set. He worries the US will face a token crunch, prioritize domestic users, and restrict model weights unless allies like the UK offer high-security sovereign compute. That could hold back UK science and companies.
Tom Milton describes Emoto as a fit-for-purpose prime for biosecurity and AI verification, similar to Skunk Works for jets or BBN for ARPANET. He says DARPA had no lab benches; it needed performers. New philanthropic funders will likewise need engineering teams that can deliver physical R&D, not just reports.
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.
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 he is not convinced verification is strictly necessary. Nuclear deals relied partly on trust, satellites counting bombers, and spying; AI deals could also use auditors, whistleblowers, or unilateral slowdowns. But as more lab leaders and governments call for coordination, verification becomes valuable optionality.
Tom Milton cites Formula 1 teams redesigning a wing overnight and flying it to Monaco. Emoto wants the same response time for science, biosecurity, and AI verification hardware. This is a bet that fast physical iteration remains a bottleneck even as AI improves design.
Tom Milton describes a mechanical engineer who needed a differential pressure sensor PCB. Instead of asking an electronics engineer for a couple hours of work, he used Fable to design the board himself. Tom expects engineers to take on more disciplines outside their training, making small teams more ambitious.
