When capability moves outward, its constraint moves with it
WASTE moves weights into storage and OpenAI moves mathematical results into Lean certificates; both make a capability claim depend on an external system's operating limits.
Artificial Intelligence··Morning
The model weights moved outside working memory
WASTE 0.6.2 is a dependency-free C inference engine that runs Kimi K3, a model with 2.78 trillion parameters, without loading every weight into working memory. It keeps the shared model trunk resident, streams the expert weights selected for each token from NVMe storage, and assigns the remaining memory to a bounded expert cache. In the repository's MacBook Pro M5 Pro measurement, the model container is 982 GiB, the resident trunk is 27.28 GB and the minimum requirement at a 4K context is 29.05 GB. New router lookahead calls the likely next expert earlier, raising the cache hit rate from 14-19% to 38-40%; the reported median speedup is 1.17 times. Even at a 38% hit rate, however, each token requires 10.5 GB of reads. Internal NVMe carries 12.78 GB per second, while a drive in a USB enclosure manages 0.94 GB per second and takes 13 seconds for the same token. The package organises which part of the workload fast storage must carry instead of shrinking the giant model.[1]
The mathematics results moved into a formal checker
OpenAI's ten-proofs repository offers a different form of externalisation. Ten results in mathematics and theoretical computer science were released not only as natural-language papers but with machine-processable Lean files. Their subjects range from sphere packing and coding bounds to non-sofic groups, a counterexample to Connes's rigidity conjecture, and the hardness of approximating the closest vector problem. The files build with Lean 4.32.0, mathlib and Lake; the repository recommends elan for installation and links to instructions for rechecking the formalisations with Comparator. An Apache 2.0 licence also defines the terms for accessing and reusing the files. The package lets a reader run the formal objects for the results in another environment. The repository description, however, does not name the model that produced the results, describe human involvement in preparing the papers, or state the peer-review status. A certificate accepted by Lean shows that the formal statement passes through the defined toolchain; other roles in the research process are outside what that file records.[2]
The external system selects which part of the claim can be tested
The two releases do not describe the same kind of product: WASTE provides an operating inference path, while ten-proofs provides formal companions to published mathematical results. Their common feature is that a capability becomes observable only after its load is translated into a defined external system. In WASTE, model capacity becomes usable when selected expert weights move from storage to memory at the right time; the visible limits are bandwidth, cache hits and the paging window. In the Lean package, results become objects that a checker with specified versions and dependencies can accept; the visible boundary is the formalised statement and toolchain. This resemblance does not make the approaches interchangeable. Fast NVMe supplies no mathematical verification, just as a certificate that builds does not measure model speed or the review status of the research. Read together, the reports replace a broad capability label with two concrete questions: which infrastructure received the load, and what can that infrastructure actually carry or check? The answers define both the scope of the published measurement and the claims left outside it.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
