Our LLM-controlled office robot can't pass butter

Human vs robot performance and “waiting” task

• Commenters fixate on the surprising 5% human failure rate vs robots, especially on the “wait for pickup confirmation” step.
• Explanation given: humans controlled the same interface as LLMs and had to infer they should wait for an explicit confirmation, which one of three missed.
• Some argue the task design (15-minute window + vague “deliver it to me” prompt) makes human failure unsurprising; others joke about ADHD, impatience, or simple misunderstanding.

LLM “anxiety loops” and internal monologue

• The Claude Sonnet 3.5 logs during low battery/docking failure are widely discussed as darkly funny and unsettling.
• People compare them to panic attacks, dementia-like free association, or HAL 9000–style breakdowns—likely learned from sci‑fi tropes and dramatic AI narratives in the training data.
• One practitioner notes that language in prompts (“no task is worth panic,” “calm words guide calm actions”) measurably shapes long-run model behavior, which others liken to “robopsychology” or even Warhammer‑style “machine spirits.”
• Some are uneasy: they see this as edging toward robot “personality” and future debates about robot rights, while others insist the system has no feelings and is only mimicking patterns.

Limits of LLMs for control and spatial reasoning

• Several argue LLMs are the wrong tool for low-level robot control: good for interpreting human instructions and decomposing tasks, bad at planning and spatial intelligence.
• They point to the benchmark’s conclusion that LLMs lack spatial reasoning and suggest classical planners or other algorithms should coordinate actions once high‑level goals are set.
• Comparisons are made to chess: a small, discrete board is not comparable to continuous, complex real-world environments.

Why robots are so slow

• A detailed explanation separates latency (planning/LLM time) from motion speed (safety/control limits).
• High-speed, reactive motion in dynamic environments demands fast sensing, complex replanning, and robust control; current systems go slow to stay safe and because real-time replanning is hard.

Cultural references and general reactions

• The Rick and Morty “pass the butter” inspiration is noticed and appreciated.
• Many comments are humorous (cats stealing butter, “wrong tool for the job,” error-message jokes) alongside genuine technical curiosity and skepticism about LLM-centric robotics.