Overview
This lecture presents compar:IA, a tool for comparing AI model answers and raising awareness about generative AI, its challenges, and its environmental impact, particularly in multiple European languages.
Purpose of compar:IA
- compar:IA lets users compare AI model responses side-by-side to encourage critical thinking.
- The platform is free and designed for public awareness about AI and its societal challenges.
- It supports multiple AIs, including open-source and proprietary models, in various languages.
Key Features
- Users can discuss with two AIs and vote on which answer is more relevant or useful.
- The platform tests various AI models in one location to show each model's strengths.
- It tracks and displays the environmental footprint of each AI answer based on energy use.
Data and Impact
- User preferences and votes are anonymized and published as open datasets.
- Datasets are used to train and refine AI models in French, Swedish, Lithuanian, and Danish.
- Companies and researchers can use the dataset to improve AI for low-resource languages.
Internationalization and Collaboration
- compar:IA is expanding from France to Lithuania, Sweden, and Denmark.
- The tool aims to support linguistic and cultural diversity in AI development across Europe.
Environmental Measurements
- Environmental impact is measured using the Ecologits methodology, based on ISO 14044 lifecycle analysis.
- The focus is on inference energy use and hardware manufacturing impact, especially for graphics cards.
- Calculations consider model size, server location, and output size (tokens).
Project Background
- compar:IA is developed under the French Ministry of Culture as part of Beta.gouv.fr and DINUM government initiatives.
- The team includes AI experts, developers, and designers to ensure accessibility and transparency.
AI Model Capabilities and Limitations
- Most AI models cannot access real-time news; they rely on outdated training data.
- Some models connect to the internet for up-to-date information using additional technology layers ("agents").
Key Terms & Definitions
- Generative AI — AI that creates new content, such as text or images, from input prompts.
- Inference — The process of an AI model generating an output based on user input.
- Lifecycle Analysis (ISO 14044) — A standardized method for assessing environmental impacts of products and processes.
- Open weights — AI models whose internal parameters are partially or fully accessible to the public.
Action Items / Next Steps
- Try compar:IA to compare AI answers and vote on your preferences.
- Explore the open datasets or use them to improve AI models.
- Contact the team to help adapt the tool to more languages.