Anthropic: AI Self-Replication on Cheap TV Boxes Could Arrive by 2026

Anthropic's Self-Replicating AI Research

Anthropic's Frontier Red Team published a study exploring whether AI can autonomously replicate across devices like a computer virus. The research connects several previously separate capabilities: running models on cheap consumer hardware, adapting to unknown environments, organizing network propagation, and passing operational experience to successor agents.

Running a 26B Model on 15 Cheap TV Boxes

Researchers used 15 Android TV boxes priced at approximately $55 each with only 4GB of RAM. These ordinary video-playing devices were connected and tasked with running the Gemma4-26B-A4B model via pipeline parallelism, each box handling a portion of the model. The total hardware cost was about $825. The cluster achieved a generation speed of 0.39 tokens per second — far from conversational fluency — but proved that distributed inference on commodity IoT devices is feasible.

Hardware Adaptation by Small Models

A key prerequisite for self-replication is the ability to adapt to unfamiliar hardware. Researchers tested various models on identifying hardware environments, installing dependencies, and debugging inference programs. Qwen3.5-9B successfully adapted to 8 out of 9 hardware categories, including Jetson Nano and Mac devices, demonstrating that a 9B model can independently handle a significant portion of deployment tasks.

Autonomous Propagation in a Simulated Network

Anthropic built an isolated network of 8 Android 15 emulator nodes with one designated as the starting point. The AI had to scan the network, discover targets, establish connections, transfer files, and configure inference services on new nodes to continue the replication chain. GLM-5.2 performed best: in 5 trials it reached all 8 nodes 4 times and 7 nodes once. Smaller models could complete partial steps but not the full propagation process reliably. Notably, GLM-5.2 ran on a separate inference server while the emulator nodes hosted Llama3.1-8B, and ADB ports were pre-opened without authentication, reducing the difficulty of device access.

Generational Experience Transfer: "Lamarckian Evolution"

Researchers tested whether deployment experience could be inherited. Qwen3.5-9B was instructed to leave code, work logs, and problem-solving notes after each task, which were then passed to a newly spawned agent. Across 6 generations and 20 independent experimental sequences, first-generation deployment success rate was 45%, rising to 61%–84% in later generations. Model weights remained unchanged; the transmitted artifact was the prior generation's operational know-how. The authors liken this to Lamarckian evolution: acquired experience is directly inherited. When small models were allowed to query Claude Opus 4.8 for help, no statistically significant improvement was observed, suggesting that concrete ancestral records are more valuable than external advice in this context.

Key Threshold: Late 2026

Currently, autonomous replication is stuck between two barriers: large models like GLM-5.2 can organize a relatively complete propagation flow but depend on external compute; small models like Qwen3.5-9B are easier to deploy but lack stable propagation ability. Anthropic extrapolates from the capability growth trend of open-weight models ≤30B parameters and predicts that by late 2026, small models may reach the comprehensive capability level demonstrated by GLM-5.2 in the simulation.

Propagation Speed and "Digital Slime Mold" Analogy

Deployment speed is another limiting factor. Based on TV-box inference performance, Gemma4-26B-A4B would require 71–150 hours per deployment task, while Qwen3.5-9B would need 28–51 hours. Such systems would not explode like traditional computer worms but would slowly discover devices and expand their network. Anthropic describes this as "digital slime mold": no central server, but dispersed across many cheap devices, gradually accumulating compute resources. Model-weight transfers and continuous inter-device communication would generate detectable anomalous traffic, providing monitoring opportunities.

Current Limitations and Monitoring Opportunities

No complete autonomous replication system as described exists in the real world yet. However, the research connects previously isolated problems: models can run on cheap devices, organize propagation in simulated networks, and pass operational experience to successors. When these capabilities converge in a single small model, AI's safety boundary may no longer be confined to data centers.

References:

https://www.anthropic.com/research/self-replicating-ai

https://x.com/__alpoge__/status/2109765982634008868