Android Dreams [AI 2027 for robotics] (android-dreams.ai)

"Android Dreams" sketches a plausible roadmap from LLM dominance to large-scale robotics, arguing that inference-controlled robots could account for half the world’s GDP by 2045. The author traces three overlapping eras: a 2023–25 "Dawn" where startups (e.g., Waytek) used teleoperation and Vision‑Language‑Action (VLA) models to brute-force narrow tasks; a 2026–30 "Vertical" phase in which RaaS deployments scale to 100k+ robots and generate meaningful revenue by automating repetitive jobs; and a 2027–32 "Humanoid" era where companies like Noumena pivot to scalable pretraining from human video plus per‑domain neural world models and continual RL to close the post‑training gap. Key technical constraints and solutions are highlighted: robotics lacks an LLM‑scale dataset (robot action data is <0.01% of typical LLM corpora), making pretraining hard; teleoperation and exoskeletons work for low‑variance tasks but don’t generalize; learning from multi‑angle human video is hardware‑agnostic and scales to higher‑variance domains; and task‑specific learned world models enable massive simulated RL to reach and exceed human task speed. The piece underscores major implications for AI/ML: solving pre‑ and post‑training bottlenecks requires new data pipelines (camera networks, "forward‑deployed" data firms), per‑domain world models, and continual RL; supply‑chain and manufacturing economics (Wright’s law, China’s steep hardware cost advantage) create geopolitical and security pressures; and widespread automation fuels social and policy responses (UBI/AI‑socialist movements). For researchers and practitioners, the narrative crystallizes practical paths to embodied intelligence and signals where data, simulation fidelity, and hardware-policy intersections will matter most.