January 1, 2026
Green light, red flags
Recreating the Bell Labs Switch Experiment with Agents
12 Switches, One Light, Zero Patterns: Can AI take the L
TLDR: A developer reimagined Bell Labs’ random switch-box to test if AI agents can admit “it’s random.” The comments exploded: some expect humility, others predict pattern hallucinations, with spicy fights over guardrails, inherited theories, and even programming language choices—an entertaining stress test of AI’s reality check.
A coder rebuilt the famous Bell Labs “random light” test—12 switches, one bulb, no pattern—and asked AI “agents” (think smart chatbots that can use tools) to find the rule. The twist: there is no rule. It’s pure randomness. Cue comment chaos. Stats folks shouted “null hypothesis or bust,” while AI optimists insisted modern models will call it random fast. Skeptics clapped back: “LLMs love inventing patterns—watch them gaslight themselves.” The hottest debate was whether passing previous theories to each AI would help them converge on truth or simply poison the well. One camp cheered the meta-experiment; another roasted it as “telephone game for robots.” People also fought over the guardrails limiting tool calls—“training wheels” vs “fair test”—and then, of course, it devolved into a Bun vs Python flame war. Memes flew: “RNGesus take the wheel,” “12 Switches of Christmas,” and a brutal “green light, red flags.” Some loved the Bell Labs throwback, others called it prompt cosplay. A few purists yelled: “If your bot can’t say ‘I don’t know,’ it’s not intelligent.” The vibe: fun, feisty, and very online. Read the Orb speedrun, stay for the drama, and pray the bots learn to love randomness.
Key Points
- •The article recounts Hamming’s description of a Bell Labs experiment where a 12-switch device’s outputs were random, yet participants formed theories.
- •The author aims to recreate this scenario with LLM-based agents to observe theory formation and acceptance of randomness.
- •A repository is initialized with TypeScript, Bun, and an AI SDK, including a hello-world script calling generateText.
- •A “mystery tool” with 12 boolean switches returning a success boolean simulates the switch box.
- •An agent is prompted to use the tool exactly 20 times, optionally consider prior theories, and then propose the best theory; guardrails will limit tool usage externally.