Lens 19: Scale Microscope. What changes at 10x? 100x? 1000x?

The scaling picture splits into three categories, but the dangerous-dimensions count drops from one to one-half once the Hailo-8 AI HAT+ on Pi 5 is activated as the L1 safety layer. Pre-Hailo, WiFi channel contention was a single undifferentiated cliff at 8 or more devices on the same 2.4 gigahertz channel — 802.11 CSMA CA's exponential backoff drove P95 latency from 80 milliseconds to over 200 milliseconds in a single-device increment, and that spike fell on both the obstacle-detection path and the semantic-query path simultaneously.

Post-Hailo, the cliff bifurcates. The 26 TOPS Hailo-8 NPU runs YOLOv8n locally on Pi 5 at 430 frames per second with under 10 milliseconds latency and zero WiFi dependency. Reactive obstacle avoidance — the path where a 200 millisecond spike could send the robot 20 centimeters past a decision point at 1 meter per second — now terminates inside the chassis. The superlinear cliff persists only for semantic queries ("where is the kitchen?", "is the path blocked by a glass door?") which still require the Gemma 4 E2B VLM on Panda over WiFi.

The bar chart now shows eight scaling dimensions, not seven. WiFi latency for semantic queries remains at 92 percent impact — the cliff persists for VLM paths. WiFi latency for the safety path has been demoted to 15 percent impact, in the favorable green zone, because Hailo-8 runs obstacle detection locally. VRAM pressure with SigLIP addition softens from 88 percent to 72 percent — still a step function, but with approximately 800 megabytes freed on Panda because obstacle detection moves off the GPU entirely. The new Hailo-8 power draw bar sits at 40 percent impact — strictly linear with inference load, no step functions, approximately 2 watts continuous.

VRAM pressure remains a step function, but Hailo-8 activation partially mitigates the ceiling on Panda. The current Panda configuration runs Gemma 4 E2B with roughly 4 to 5 gigabytes consumed against a 16 gigabyte practical ceiling. Adding SigLIP 2 ViT SO400M adds 800 megabytes in a single step; adding DINOv2 ViT-L adds another 1.2 gigabytes. Pre-Hailo, two models stacked alongside E2B crowded the ceiling. Post-Hailo, because obstacle detection runs on the Hailo NPU — separate silicon, separate memory, not a VRAM line-item — roughly 800 megabytes is freed from the Panda nav pipeline. Enough headroom to absorb the SigLIP step without qualitative pressure. The DINOv2 step is still binary, but now has breathing room.

Map area, embedding storage, and scene label vocabulary remain in the favorable zone. Scene labels plateau at 6 to 12 semantically distinct spaces per home. Map files are trivially small. Embedding storage at 60 kilobytes per session accumulates under 250 megabytes across a decade of daily use.

Nova. Hailo-8 activation neutralizes the superlinear WiFi cliff for the safety path. YOLOv8n runs locally on the 26 TOPS NPU at 430 FPS, under 10 milliseconds, zero WiFi dependency, approximately 2 watts continuous. Reactive obstacle avoidance no longer traverses the shared-medium channel. The 802.11 CSMA CA cliff persists only for semantic queries on Panda, not for safety-critical control. This is the single highest-leverage scaling improvement available to Annie, and it requires zero software rewrite — the NPU is already on the robot and currently idle.

Hailo-8 scales as a clean linear curve, not a step function. Power consumption rises smoothly with inference load, target approximately 2 watts continuous. VRAM is not a line-item — separate NPU silicon. No discontinuities, no cliffs. The new L1 safety layer adds capability without adding any of the dangerous scaling patterns present elsewhere in the stack.

VRAM step function is partially mitigated by Hailo offload. Moving obstacle detection to the Hailo NPU frees approximately 800 megabytes on Panda — roughly one SigLIP-sized addition of headroom against the 16 gigabyte ceiling. Each new model on Panda — SigLIP to DINOv2 — remains a fits-or-crashes decision, but one rung of the ladder is now wider. Session 270 silent-overflow discipline still applies. Hailo buys runway, not immunity.

The whole-house inflection point is the design horizon, and Hailo-8 moves it outward. With the safety layer decoupled from WiFi, the previous brick wall at 8 or more devices on 2.4 gigahertz becomes a soft degradation of semantic response time rather than a safety failure mode. Annie's architecture gains real headroom at whole-house scale without the WiFi brick-wall. Above multi-building campus scale, the architecture still requires structural change — shared inference, mesh networking, federated trust — but the sub-whole-house regime just got substantially more robust.