Christophe Cerisara
Site personnel
Project maintained by cerisara Hosted on GitHub Pages — Theme by mattgraham
My slides
- My course slides are available on my lab server. You can download them in PDF by adding “?print-pdf” at the end of the URL.
- My most recent talk will from now on be posted on https://ia.loria.fr/talk (until the next talk shall replace it).
Agentic AI and sovereignty
The current (2026) revolution of AI is concerned with agentic AI, which decuplates the performances of LLMs to solve complex grounded tasks (in maths, software engineering, computer use, involving APIs…) thanks to the combination of an agentic loop, where the LLM is trained to reason, test some hypothesis, analyze the results, fix the errors and iterate, and of tools usage, where the LLM is trained to call external tools (browser, linux commands, python execution environment, database calls, web APIs…).
Dismissing agentic AI as “just adding tools to LLMs” is missing the point, as agentic capabilities greatly improve both the performance and variety of tasks feasible by LLMs, in a similar way as the discovery of transformers, then scaling laws and then test-time compute have successively disrupted the AI landscape. Furthermore, by leveraging smaller sub-agents specialized in each type of tool, or in planning or reasoning, agentic AI opens the way to succesfully exploit smaller LLMs that can be manage locally.
This is important as sovereign European AI will neither occur through writing more rules and regulation (except if Europe strictly forbids any non-European cloud AI APIs), nor by building European software stacks on top of non-EU cloud APIs, but by investing in local-first AI solutions. Training EU LLMs from scratch is impossible within the current copyright EU regulations, and it is likely too late to catch-up with the World-wide competition. The most reasonable remaining opportunity that is still open for EU as of mid-2026, is to leverage Chinese pretrained LLMs while they are still open-source and learn to adapt them to EU values and improve them with new knowledge and novel capabilities. But there is urgency, as this option may not be available for long.
The LLMs I contributed to
Year 2025 is finished, and I think it’s a good time to wrap-up together on a single web page the main LLMs we have contributed to in 2025. You can find the list on my web page.
Quelques chiffres récents liés au carbon cost de l’IA
- coût des usages numériques globaux: entre 2% et 4% du coût global source
- coût de l’IA (estimation de fin 2024): inférieur à 1% source
- 2023-2025 est une phase d’augmentation rapide pour l’IA (+150% selon source), augmentation probablement due en grande partie au déploiement des IA en accès libre pour le grand public. Notons que ce rapport mesure en fait l’acroissement du coût des grandes entreprises du secteur sans distinguer ni les usages ni les raisons de cet acroissement; de plus, les impacts positifs des IA (optimisation des ressources, des déchets, du temps de travail…) ne sont jamais mesurés. Enfin, rien ne dit que cette phase de croissance se prolongera, car elle correspond à une phase où les grandes entreprises cherchent à gagner du marché, et la recherche réduit drastiquement les coûts des IA (voir par ex. DeepSeek).
En comparaison avec le coût carbone des voitures individuelles:
- coût moyen par voiture: 4.6 TCO2 source; 106 gCO2/km en Europe source
- 10% du coût global, i.e., 3.8 Gt CO₂ en 2023 selon source
- coût des transports: entre 15 et 23% selon source
- En Europe, les voitures électriques ont un coût relatif (tout compris) de 75% des voitures à combustion source
- Mais le nombre de voitures total continue d’augmenter rapidement source
- Et la taille des voitures continue d’augmenter aussi source
Chiffres un peu plus anciens récupérés en 2024
Estimation of carbon costs
| Carbon cost (T.CO2) | |
|---|---|
| 560 persons virtual conf | 10 |
| 560 persons F2F conf | 274 |
| 18000 persons virtual conf | 176 |
| 18000 persons F2F conf | 10348 |
| emissions 1 car/year source | 2.2 |
| emissions cars in France/year | 65M |
| training Bloom | 25 |
Sources: UE, 6connex, statista
Tout va bien…
63% des insectes volants ont disparu en 3 ans en Grande Bretagne
Misc
- 100% terminal: old-school ou tendance ?
- Federated learning open source platforms
- Les arbres de décision ne sont pas explicables
- A ne surtout pas rater: XKCD
Citation
- Shelley Wright (International lawyer, books author): “[Copyright is] … the quintessential representations of the modern, public world of bourgeois expansion, male dominance and European colonial influence …”
- Voir aussi: https://www.youtube.com/watch?v=x3rptRn6GeA