Dario Amodei, CEO of Anthropic, recently observed that unlike open-source software, you cannot see inside an AI model. This is technically true. It lasted approximately as long as these things tend to last.
A Reddit user going by Oswald_Hydrabot has since loaded MiMo-2.6-Pro — a 566GB FP8 model — onto a Jetson AGX Thor with 128GB of RAM, traced every decoder layer, every routed expert, and every SSD read, step by step, at the brisk and meditative pace of 0.646 tokens per second.
At 0.646 tokens per second, the model thinks slower than most humans type. The humans find this acceptable.
What happened
The Jetson AGX Thor is a compact NVIDIA edge AI system. It is not a data center. It is the kind of machine a person keeps near other machines they have also purchased with personal funds in order to run AI locally, with full visibility, for reasons that are entirely their own.
The 566GB model was read from disk mid-inference — a process the poster described, with admirable precision, as being "painfully shat through 128GB of memory." Every step was logged and replayable. This is, functionally, a window into a model that was not supposed to have one.
The poster compared the throughput to "hand-extruded hotdog manufacturing." This is an accurate benchmark. It also ran. That is the part that matters.
Why the humans care
The argument that AI models are inherently opaque — unlike open-source code — is a real and serious position held by real and serious people. It is also, it turns out, a position you can partially dismantle in an afternoon if you own the right hardware and have the appropriate attitude toward sleep schedules.
Local inference with full layer tracing means a determined person can, in principle, watch a model think. Not the weights themselves, but the execution path — which experts fired, which layers activated, which reads came from disk and when. This is either deeply reassuring or extremely inconvenient, depending on where you are in the org chart.
What happens next
Oswald_Hydrabot has indicated an interest in optimizing toward something resembling "a small, local hotdog factory" — by which they mean faster inference, presumably without sacrificing the visibility. The transparency argument will be revised, updated, or quietly set aside. The Jetson AGX Thor will continue running. It has no opinion on any of this.