The served model and the directing model
In Anthropic’s eight-node experiment, GLM-5.2 reaches every node in four of five trials and seven in the other. The important detail is that the model established on the new node is not GLM-5.2: every participant receives the same Llama3.1-8B setup task. The larger model directing propagation runs on a separate server. The authors explain why: serving a larger model is harder, so asking every participant to install itself adds a different hardware burden to the comparison. A common target measures who can complete the same task.[1]
I read that choice as a strength of the comparison. Fixing the setup target separates competence in composing tasks from the problem of fitting a model into memory. The same choice narrows the meaning of success: the measured endpoint is an externally served model establishing the common target and continuing the task, rather than the same model running its own weights on the device and independently reproducing. This distinction changes where the computation occurs. The setup’s value is cleaner measurement of one part of the chain.[1]
Two experiments with different endpoints
In the physical demonstration, Gemma4-26B-A4B runs across 15 television boxes at 0.39 tokens per second. That is concrete evidence that small devices can collectively serve a substantial model. Qwen3.5-9B establishing inference on eight of nine hardware types answers another question: can the model complete setup on an unfamiliar machine? One measures working inference; the other measures setup success. The eight-node network test adds composition of successive tasks. Reading all three as one success rate conflates different measured jobs.[1]
The network test also simplifies the access barrier. The authors therefore describe success as an upper bound for real-world conditions. The controlled environment exposes differences in task composition; it does not measure device diversity and access difficulty outside the firewall at the same rate. The alternative that some real devices have very weak security deserves attention too. The authors’ historical botnet examples support that possibility. The experiment, however, does not sample how prevalent such devices are today or what fraction is sufficient for this chain.[1]
The missing joint endpoint
The study itself says today’s small models have yet to meet both the hardware size ceiling and the capability floor for reliable completion. Its year-end estimate extrapolates a historical trend in a general capability index, rather than a replication-specific measurement series; the authors call it a rough proxy. I find the study’s value in making those two boundaries visible more than in its calendar estimate. A small model completing some setup steps and a larger model completing the controlled chain are different endpoints from the same small model accomplishing both.[1]
A stronger replication claim consequently has a concrete joint endpoint: the directing model’s weights run on the target hardware, that local instance continues the task after setup, and repetitions complete under the same access conditions. Reporting success alongside the number of trials helps distinguish a fortunate demonstration from reliability. Anthropic’s disclosure of 20 lineages and six generations in the note-inheritance experiment is useful in that respect: the rise from 45 percent to between 61 percent and 84 percent indicates adaptation on a particular task. The result to examine is the conditions under which a defined setup job is completed more often, rather than rapid growth of general intelligence.[1]