The last day of Devoxx France delivered a series of dense conferences, mixing technological prospective, geopolitical infrastructure and questioning of the founding myths of artificial intelligence. Five major interventions are distinguished, offering technical directors a useful reading grid for thinking about their Roadmap 2025-2028.
1. Demystifying AI: a probabilistic machine, not an autonomous entity
Dr. Luc Julia, co-creator of Siri and scientific director of Renault, opened the session with an unambiguous position: “Artificial intelligence does not exist. »» For him, AI remains a technology ofStatistical optimizationincapable of understanding in the human sense. Even the most efficient systems, such as LLMS, “reach only two -thirds relevance”.
Julia urges the technical directions not to give in to alarmist or messianic stories. The stake, according to him, is to reintegrate humans into the loopby favoring transparent, explainable systems, and at the service of concrete business uses. For a CTO, this implies a reclassification of the projects around verifiable and measurable use cases, far from general promises.
2. Strengthen digital sovereignty by going back down the OSI battery
In a remarkable conference, Ophélie Coelho, specialist in digital policies, revisited the Seven layers of the OSI model Through geopolitical reading: from control of submarine cables to data, including data and bones. It recalls that the territorialization of infrastructure is an issue of sustainable power: “The one who controls traffic nodes controls digital sovereignty. »»
This message implies, for CTOS, to Relocate strategic attention to low layers : choice of host, cloud dependence, bone control, network resilience. At a time when sovereignty is reconfigured around international tensions, the physical dimension of infrastructure can no longer be relegated to purchases or IT.
3. Anticipate the quantum transition, today
Fanny button, active member of the French quantum ecosystem (Quantum Lead & Product Manager at OVHCLOUD, and CO founder of France Quantum), has delivered an unambiguous inventory: France is In the world top 3with six projects among the ten more advanced. The quantum revolution is no longer speculation: “The first quantum engines will be operational within three years. You have to learn to pilot them now. »»
His message: Create immediate acculturation conditionsthrough simulators (type notebooks IBM or ATOS), and mobilize technical teams-including non-specialists-around hybrid R&D projects. For a CTO, this means inscribing quantum in watch, training, and prototyping priorities, without waiting for “market maturity”.
4. Master the war of stories: AI and language as weapons of influence
In a dense intervention, Élodie Laye Mielczareck, Sémiolinguiste and Analyst of the speeches, exposed the four levels of distortion between language and reality. Level 4 – the most critical – corresponds to the use of language for deny the real. In the era of generative AI, this mechanics are multiplied: Fake News, narratives of influence, automated belief bubbles.
For CTOS, this analysis calls for a vigilance on automated linguistic uses: chatbots, automatic summaries, voice. The question is no longer just technical but ethical: what stories do our systems produce? And for what ends? Algorithmic governance and human validation become critical issues to avoid drifts.
5. Understand the internal representation models of LLMS
Thibaut Géraud, alias Mr Phi, has shown that LLMS develop internal representations of the world. Through experiences on chess, he has shown that some models reach a Elo score of 1800equivalent to a professional level, thanks to a unique prompt: “Play as a Professional Chess Player. »»
This observation shakes up the traditional vision of LLMS as simple predictors. They build “World Models”, a sort of dynamic memory. For a CTO, this implies reconsider the abstraction capacities of these modelsbut also of wondering about their traceability, their biases, and their potential opacity. The development of internal visualization functions or audits of models becomes crucial.
