- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,217
The AI box will answer any question even in the bunker... if you don't have a hand off.
While the technological giants compete in the power of data centers, two engineers have collected artificial intelligence, which only needs a human hand to work - literally.
Computer scientist Catherine Tomanek and former technical director of one of the divisions of Google Alex Kaufmann created a CrankGPT voice AI assistant living inside a 3D-printed hull with a rotary handle. A detailed video demonstration can be viewed YouTubeon YouTube.
The device answers questions and translates speech into other languages - but only for now someone rotates the handle. The built-in condenser unit of its own design stores the charge for about 20 seconds of offline operation, after which it again requires active actions on the part of the user. From the first turn to the start of the conversation takes about 30 seconds.
Inside the case is a single-board computer Raspberry Pi 5 with 8 GB of RAM, specialized expansion boards for audio input and output, as well as a serial 20 watt-in-watt manual generator, designed to charge USB devices in emergency situations.
The operating system is the most lightweight version of DietPi, which is loaded in three seconds. Voice recognition provides the Moonshine engine, speech synthesis - Piper. Local LLM-models Liquid LFM2 1.2B and Gemma 3 1B are responsible for semantic processing. All voice agent code is written from scratch and is already available on GitHub.
An unusual feature of the device: the resistance of the handle varies depending on the load - during the generation of the answer, it is noticeably heavier than in standby mode. In addition to text answers, on the same platform it was possible to generate images, write poems and code.
CrankGPT serves as a demonstration of the Squeez approach, which was founded by Tomanek and Kaufmann: compact, private and specialized language models that can work on cheap equipment without an Internet connection. Targeted scenarios are voice recognition, taking into account the accent or speech features of the user, as well as highly specialized local assistants, for example, on gardening or repair of cars.
Schemes and drawings for self-assembly, the authors plan to publish in the coming days. According to Kaufmann, due to the increase in memory prices, the cost of assembly increased to about $ 300 - against the 150 he spent when creating the first prototype a year earlier.
While the technological giants compete in the power of data centers, two engineers have collected artificial intelligence, which only needs a human hand to work - literally.
Computer scientist Catherine Tomanek and former technical director of one of the divisions of Google Alex Kaufmann created a CrankGPT voice AI assistant living inside a 3D-printed hull with a rotary handle. A detailed video demonstration can be viewed YouTubeon YouTube.
The device answers questions and translates speech into other languages - but only for now someone rotates the handle. The built-in condenser unit of its own design stores the charge for about 20 seconds of offline operation, after which it again requires active actions on the part of the user. From the first turn to the start of the conversation takes about 30 seconds.
Inside the case is a single-board computer Raspberry Pi 5 with 8 GB of RAM, specialized expansion boards for audio input and output, as well as a serial 20 watt-in-watt manual generator, designed to charge USB devices in emergency situations.
The operating system is the most lightweight version of DietPi, which is loaded in three seconds. Voice recognition provides the Moonshine engine, speech synthesis - Piper. Local LLM-models Liquid LFM2 1.2B and Gemma 3 1B are responsible for semantic processing. All voice agent code is written from scratch and is already available on GitHub.
An unusual feature of the device: the resistance of the handle varies depending on the load - during the generation of the answer, it is noticeably heavier than in standby mode. In addition to text answers, on the same platform it was possible to generate images, write poems and code.
CrankGPT serves as a demonstration of the Squeez approach, which was founded by Tomanek and Kaufmann: compact, private and specialized language models that can work on cheap equipment without an Internet connection. Targeted scenarios are voice recognition, taking into account the accent or speech features of the user, as well as highly specialized local assistants, for example, on gardening or repair of cars.
Schemes and drawings for self-assembly, the authors plan to publish in the coming days. According to Kaufmann, due to the increase in memory prices, the cost of assembly increased to about $ 300 - against the 150 he spent when creating the first prototype a year earlier.
