Bibliothèque Nationale de France (BNF)

Conception 

For the exhibition Georges Perec: Modes d'emploi, the Bibliothèque nationale de France (BnF) commissioned TEKLIA and Réciproque to develop an interactive installation using generative artificial intelligence. The installation enables visitors to generate variations on Georges Perec’s texts by applying a set of predefined constraints. Its aim is to offer a playful experience that draws on the principles of Oulipo and constrained writing, extending these literary practices into the realm of AI-generated text. 

TEKLIA conducted a systematic evaluation of the Perec texts submitted to the AI system, as well as the literary constraints applied to them, in order to identify the most appropriate technical solutions for each use case. Particular attention was paid to the feasibility of running an AI model live and locally, as well as to the quality of the generated outputs.

The selection of constraints was based on an assessment of the relevance of approximately one hundred possible constraints, drawn from or inspired by Oulipo practices. These were evaluated according to the following criteria: 

• The ability to formulate an effective prompt to guide the tasks performed by the models.
• The capacity to generate meaningful, coherent text that can be readily understood by a broad audience.
• The possibility to present the generated text through a visual interface.

The AI model was selected following a comparative study assessing the creativity and relevance of the generated texts. The installation relies on a small language model (LLM), designed to operate with limited computing resources and pre-trained for real-time text generation. The model runs locally on the BnF’s premises.

By comparison:
Boiling one liter of water: 100 Wh (equivalent to 500 queries via the Perec device)
ChatGPT: 0.34 Wh per query  
Standard Google query: 0.3 to 3 Wh, depending on complexity

We have disabled the model's reflection mode to reduce energy consumption and the time required to generate results. 
Maximum power consumption of the selected model: 140 Wh