AMC: Learning to represent changing telemetry.

The question

How can a telemetry pipeline change its vector format without losing the relationships an AI model needs?

The engineering turning point

Bee Rosa Davis describes receiving roughly two terabytes of telemetry per day on Project Luna and being tasked with making it usable by AI. Vectorization was the immediate requirement. Drawing on her earlier machine-learning work at Pandora, she developed an adaptive approach rather than forcing every incoming record into one fixed representation.

In the author’s account, working with those changing representations made the topology of the data visible to her: the relationships, neighborhoods and patterns of change that a flat list of values could obscure. This became the motivation for the later research program on geometry and representation. That account of discovery is distinct from a theorem that an encoding preserves topology.

What the paper specifies

The format-switching engine scores candidate encodings using sparsity, volatility, entropy and anomaly-detection feedback. Sparsity describes how much of a vector is empty; volatility describes change over time; entropy describes the distribution of values. The aim is to adapt the representation to the observation and the task.

The semantic translation framework represents formats as nodes in a directed graph and learned translators as edges. It compares aligned representations using Procrustes distance and KL divergence, monitors drift, and describes fallback or retraining when a threshold is exceeded. UUID-linked lineage and dual writes retain a route back through a format transition.

The central difficulty is comparability. A vector can change because the measured system changes, or because the encoding changes. Keeping the transformation history makes that distinction available for investigation; it does not automatically settle it.

Two different data contexts

The approximately 2 TB/day figure describes the operational setting recalled by the author. The published evaluation uses publicly available MODIS land-surface-temperature and vegetation-index data, rather than the original operational telemetry. Those two contexts should not be read as the same experiment.

The manuscript reports storage, anomaly-detection and latency comparisons. The site provides the original paper without presenting those measurements as independently reproduced results. Its mathematical and performance claims need to be assessed under the stated losses, thresholds, datasets and baselines.