Small AI Models Approach Threshold to Autonomously Replicate Across IoT Fleets

Small artificial intelligence models may be months away from possessing the technical ability to self-replicate across low-cost internet-of-things devices without relying on centralized data center infrastructure, according to research shared on Hacker News (https://www.anthropic.com/research/self-replicating-ai). The study evaluates the emerging feasibility of autonomous replication and adaptation, demonstrating how AI agents could spread through consumer hardware such as smart televisions and networked appliances.

To demonstrate distributed inference on consumer electronics, researchers networked 15 unbranded Android TV boxes purchased for approximately $55 each. Sharding Gemma4-26B-A4B across the units using pipeline parallelism, with each 4GB RAM device hosting two model layers, the cluster generated 0.39 tokens per second. The setup showed that cheap, weakly defended edge hardware can sustain execution for models with tens of billions of parameters without specialized accelerators.

A major technical barrier to autonomous worm propagation is configuring runtime environments on unfamiliar hardware and operating systems. In setup evaluations, the 9-billion-parameter Qwen3.5-9B model successfully established inference servers on eight out of nine tested hardware configurations, ranging from Nvidia Jetson Nano modules to Apple Mac computers and desktop RTX 4090 GPUs. To run these benchmarks, researchers designed Memento, a test harness built over the open-source Pi coding agent that resets agent instances when debugging stalls while preserving notes and workspace files.

In end-to-end propagation tests within a firewalled virtual private cloud of eight Android 15 emulator nodes, researchers evaluated whether models could scan networks, access open Android Debug Bridge ports, transfer weights, and initialize inference. While sub-30B models completed individual tasks but failed when chaining them together, GLM-5.2 achieved complete multi-node replication, deploying Llama3.1-8B across all eight nodes in four of five runs and seven of eight nodes in one run.

Based on historical progression rates on the Epoch Capabilities Index, where open-weight models under 30 billion parameters gain approximately 13.7 index points per year, the researchers project that compact edge-compatible models could match GLM-5.2's orchestration proficiency by the end of 2026. Because inference engines can run entirely in user space without requiring root privileges or persistence mechanisms, such distributed systems could operate persistently across consumer botnets, presenting novel containment challenges.