Note : The original text is in French.

From conversational agent to collective tool

New information behaviors emerge among users with the deployment of conversational interfaces in cyberspace. This is because natural language interactions with computer are made possible by the capabilities of language models. This new mode of interaction is transforming the practices of many people in how they acquire and process knowledge. Without exhaustively detailing the consequences of this new approach to knowledge, we think that connectionism is now fundamental to digital technology. And digital, as a cognitive support, is now largely influenced by artificial neural networks and their automatic processing applications.

Based on this observation, our goal is to support these changes and steer them toward cooperation. Thus, we aim to build a system that promotes reflective and collective practice. To achieve this, we propose a simple principle :

  • Use automatic generation to ask questions to a community.

This is in the context of collaborative knowledge bases, wikis, in order to encourage contributions and collective writing.

Impacts

Given social and environmental concerns, the impacts of deploying language models are significant.

  1. The infrastructure powering conversational agents demands substantial energy, while the hardware supply chain for data centers accentuate mining extractivism. As a result, this amplifies environmental pollution and global warming.
  2. Generative AI technologies are mostly owned by a monopoly of private actors, giving them a strong cognitive influence over society.

For more information on the impacts, you can explore the Estampa’s cartography.

A third way

By integrating behavioral change and considering technological impacts, we aim to offer a third way based on a set of technical principles :

  1. A mode of collective interaction with a language model: the model performs automatic language processing tasks based on shared knowledge to respond to the needs of a community.
  2. A prompt system, designed and implemented in a free and open process.
  3. A limited number of daily requests for pre-determined tasks.
  4. The use of small language models, executable on CPUs* and self-hostable, via an open-source inference engine.
  5. The use of language models, including public data sources and machine learning methods.

(*) The aim is to select system that are comptatible with basic device : Running a small language model requires computational operations (FLOPs) compatible with most CPUs, whereas large language models demand GPU-based infrastructure with significantly higher energy consumption.

Manolia

The principles outlined form the foundation of Manolia, a tool for contributors to collaborative knowledge bases, implemented as a MediaWiki extension. Since contributions to a wiki stem partly from social motivations, our goal is to reinforce these incentives. This leads to an improvement in the quality and relevance of shared knowledge.

So, we focus on two simple and complementary features:

  • Generating a list of questions for contributors, representing the community information needs.
  • Providing edit location suggestions for each question.

In summary, we propose using the capabilities of a language model to detect, synthesize, and formalize the community’s information needs by generating questions. These rely on human reasoning applied to knowledge and thus support contributions.

Ultimately, Manolia aims to support knowledge system for small and medium-sized communities, outside the ecosystem of major AI providers.

Astrolabe CAE logo Astrolabe CAE