{"rewrite":{"id":"r_2cd268067481ddcca2c55a14","clusterId":"c_00b5a39f1443c44f5b347727","slug":"crankgpt-hand-cranked-ai-machine-runs-voice-chat-on-muscle-power","model":"deepseek-v4-flash","headline":"CrankGPT Hand-Cranked AI Machine Runs Voice Chat on Muscle Power","summary":"A new device called CrankGPT runs a local voice conversation AI entirely on electricity generated by a hand crank. The system uses a Raspberry Pi 5 with custom power and audio boards, and the crank's resistance increases during high-load AI processing, giving the user physical feedback.","whyItMatters":"CrankGPT inverts the usual AI power narrative by making the user supply the energy physically, turning inference into a tangible, muscle-powered act.","webCardHtml":"\u003cp\u003eThe hardware consists of a hand-crank generator connected to a Raspberry Pi 5 with 8GB of RAM, a cooling fan, and an audio output expansion board. The Pi runs DietPi Linux and a custom voice agent that combines speech recognition, a large language model, and speech synthesis. A custom power supply board handles instantaneous current spikes.\u003c/p\u003e\u003cp\u003eThe crank\u0026#39;s resistance changes with power demand, so the user physically feels when the system is doing heavy work such as running LLM inference and speech synthesis at the same time. Startup takes about 25 to 35 seconds total. Voice latency ranges from roughly 2.9 seconds with Gemma 3 1B to about 0.8 seconds with LF2.5-350M.\u003c/p\u003e","blueskyPost":"CrankGPT is a hand-cranked AI device that runs a voice conversation agent locally on a Raspberry Pi 5. The crank gets harder to turn during heavy inference. Muscle-powered AI is here.","twitterPost":"CrankGPT is a hand-cranked AI device that runs a voice conversation agent locally on a Raspberry Pi 5. The crank gets harder to turn during heavy inference. Muscle-powered AI is here.","threadsPost":null,"newsletterBlurb":"A new device called CrankGPT runs a local voice conversation AI entirely on electricity generated by a hand crank. The crank's resistance increases during high-load AI processing, giving the user physical feedback. It uses a Raspberry Pi 5 with custom power and audio boards.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260616-crankgpt-local-human-powered-ai/\",\"title\":\"Muscle-powered hand-cranked AI machine 'CrankGPT' appears, designed so that higher-load AI processing requires more physical strength\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4236,"outputTokens":529,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1781621879,"createdAt":"2026-06-16T14:46:01.000Z","publishedAt":"2026-06-16T14:49:57.000Z","updatedAt":"2026-06-16T14:49:57.000Z"},"cluster":{"id":"c_00b5a39f1443c44f5b347727","canonicalTitle":"力こそパワーな筋肉駆動の手回し式AIマシン「CrankGPT」が登場、負荷の高いAI処理ほど多くの筋力を求められる仕様","representativeArticleId":"a_9eda74f3b446c595b00eefe3","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"CrankGPT\"],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-06-16T14:00:00.000Z","lastSeenAt":"2026-06-16T14:00:00.000Z","updatedAt":"2026-06-16T14:49:57.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260616-crankgpt-local-human-powered-ai/","title":"力こそパワーな筋肉駆動の手回し式AIマシン「CrankGPT」が登場、負荷の高いAI処理ほど多くの筋力を求められる仕様"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["CrankGPT"],"studios":[],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":["CrankGPT runs a local voice conversation AI entirely on electricity generated by a hand crank.","The hardware uses a Raspberry Pi 5 with 8GB of RAM, a cooling fan, and an audio output expansion board.","The crank's resistance increases during high-load AI processing, giving the user physical feedback.","Startup takes about 25 to 35 seconds total.","Voice latency ranges from roughly 2.9 seconds with Gemma 3 1B to about 0.8 seconds with LF2.5-350M."]}