GOD MODE (ep. 1) is an interactive audiovisual performance set in a real-time 3D simulation running on a video-game engine. The simulation is similar to state of the art simulations used to generate synthetic data to train Artificial Intelligence models for computer vision tasks. In GOD MODE (ep. 1), an AI destined for deployment in a cashierless store is training to recognise items on shelves as it delivers a monologue cowritten with GPT-3. Throughout the performance, the AI grows frustrated with the difficulty of its training until it finds a bug to exploit in order to cheat its way out of the simulation.
Featuring live music by artist HERO IMAGE, facial motion capture and voice modulation to animate the AI on screen, and audience interaction via phone, the performance asks what kind of intelligence might emerge from using simulation in AI training. It addresses the claims of bias-free training and full automation behind the technology, the implications of learning from a virtual world in order to operate in the real world, and the limitations of deep-learning training methods using rewards to encourage AIs to repeat desirable actions, proposing that shortcuts to rewards, rather than concepts, are learned as a result.
Aspects of the deep-learning training method used in the simulation are turned into game mechanics that let the audience modify the environment via phone. As the AI responds, the link between rewards, shortcuts and the changing nature of the simulation is explored.
- 2022.3.30 (live)
- GOD MODE (ep. 1)HQI, london.......................................................................................................................................................................................................................................... A real-time rendered monologue about an AI cheating its way out of a simulation
- CPCCENTRE FOR POSTDIGITAL CULTURE
GOD MODE (ep. 1) is an interactive audiovisual performance set in a real-time 3D simulation running on a video-game engine. The simulation is similar to state of the art simulations used to generate synthetic data to train Artificial Intelligence models for computer vision tasks. In GOD MODE (ep. 1), an AI destined for deployment in a cashierless store is training to recognise items on shelves as it delivers a monologue cowritten with GPT-3. Throughout the performance, the AI grows frustrated with the difficulty of its training until it finds a bug to exploit in order to cheat its way out of the simulation.
Featuring live music by artist HERO IMAGE, facial motion capture and voice modulation to animate the AI on screen, and audience interaction via phone, the performance asks what kind of intelligence might emerge from using simulation in AI training. It addresses the claims of bias-free training and full automation behind the technology, the implications of learning from a virtual world in order to operate in the real world, and the limitations of deep-learning training methods using rewards to encourage AIs to repeat desirable actions, proposing that shortcuts to rewards, rather than concepts, are learned as a result.
Aspects of the deep-learning training method used in the simulation are turned into game mechanics that let the audience modify the environment via phone. As the AI responds, the link between rewards, shortcuts and the changing nature of the simulation is explored.
