We scan new podcasts and send you the top 5 insights daily.
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.
By monitoring a model's internal activations during inference, safety checks can be performed with minimal overhead. Rinks claims to have reduced the compute for protecting an 8B parameter model from a 160B parameter guard model operation down to just 20M parameters—a "rounding error" that makes robust safety on edge devices finally feasible.
Bizarre inference bugs, like a model endlessly repeating a token, may not be a model or software issue. They can be elusive race conditions in CUDA kernels, only exposed by specific hardware configurations, such as a slower node-to-node interconnect in one cluster versus another.
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.
Ivan Barazin says Meta ran the CPU and execution side of Muse post-training on Daytona sandboxes. This shows agent execution infrastructure is already used not just for live agents but for model training tasks. The work is separate from inference, where the model thinks and answers.
Hardware architectures originally built for AI training can outcompete purpose-built inference hardware as model sizes scale. AWS originally built Inferentia specifically for inference and Trainium for training. However, Garman notes that as enterprise models grew significantly, Trainium actually emerged as their highest-performing and most cost-effective inference chip, powering the vast majority of Bedrock's operational inference workloads today.
While the idea of distributed compute pools is appealing, it's not feasible for AI training due to high latency demands; GPUs must be physically co-located. However, AI inference is less sensitive to this lag, making a distributed network of compute (like home GPUs) a much more viable and exciting model.
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.
To reduce hallucinations, Goodfire runs a detection probe on a frozen copy of a model, not the live one being trained. This makes it computationally harder for the model to learn to evade the detector than to simply learn not to hallucinate, addressing a key failure mode in AI safety.
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.