Urban data science
Cities, heterogeneous data, change.
Mindstone AI Meetup · · EPITECH, Lyon
When the old constraints disappear
Work and creativity
in the age of AI
What becomes possible?
What still matters?
John Samuel
CPE Lyon · LIRIS (CNRS UMR 5205)
Researcher · Educator · Creator
johnsamuel.info
My research lens
Data. Space. Time.
DATASPACETIMEKnowledge
we can explore
we can explore
Cultural heritage
Sources, memory, storytelling.
3D visualization
Making complex relationships visible.
How can heterogeneous data become knowledge that humans can explore?
Research in practice
Three current projects
Images → Context
ANR AGAPE1
Discovering spatial iconographic heritage
Iconographic sources + 3D positioning
Link heterogeneous sources.
Visualize them in 3D.
Patterns → Understanding
IA.rbre2
Data for Lyon’s climate resilience
Aerial imagery, LiDAR, urban data
Map plantability.
Assess heat vulnerability.
Terms → Knowledge
FSP TAC3
Thesaurus Automation Curation
LLMs + expert curation
Propose concepts and relations.
Align to existing vocabularies.
From abundant material to structured, explorable knowledge.
1 ANR : Agence nationale de la recherche · agape-anr.github.io
2 IA.rbre : TelesCoop, Métropole de Lyon, LIRIS · France 2030 · iarbre.fr
3 FSP : Fondation des Sciences du Patrimoine · github.com/VCityTeam/TAC
An intellectual journey
From symbolic to neuro-symbolic AI
My starting point
Symbolic
Rules
Datalog · declarative queries
Explicit semantics
A different capability
Neural
Patterns
Learned representations
Statistical associations
An emerging direction
Neuro-symbolic
Together
Discover possibilities
Organize, verify, reuse
FSP TAC, in miniatureCandidate termsExplicit relationsExpert validation
Neural models discover possibilities. Symbolic models make them reusable.
Frugal AI starts before training
Do we need to train another model?
Think of your next AI project. Where would you start?
- 01
Question the need
Could retrieval or explicit rules suffice?
- 02
Reuse what exists
Can an existing, smaller model do the job?
- 03
Adapt with evidence
Fine-tune or train only when the task justifies it.
Count the whole cost: data · energy · expert time
Choosing what to compute is itself an act of intelligence.
Programming in human languages
Code in the language you think in
ThenExploring a multilingual
command line
Nowmultilingual
A programming language
Different languages → shared program semantics
Actual code · Hello World
English
print("Hello world")Français
afficher("Bonjour le monde")↓
Localized builtins · shared execution model
Examples from the project README.
More people should be able to express computational ideas in their own languages.
Who gets to shape what we know?
Intelligence as a commons
Wikipedia
Wikidata
Wikifunctions
Abstract Wikipedia
Shared facts
C · developer · Dennis Ritchie
C · influenced · C++
Language-aware expression · illustrative example
C was developed by Dennis Ritchie and influenced C++.
C a été développé par Dennis Ritchie et a influencé C++.
Open source · Open data · Reproducibility · Multilinguality
Communities provide what data alone cannot: context, correction and meaning.
Teaching C at CPE Lyon
Learning changes. So must assessment.
AI makes practical work easier. Fundamentals still matter.
Questions & quizzes
Adapt questions to learning in the age of AI.
Educational games
Exploring learning through play.
New ways to evaluate
Assess understanding as well as the result.
Two complementary settings
Without internet
Written exams
Reason about C fundamentals.
Trace code. Explain memory.
With AI
Practicals
Build, test and debug.
Explain the code and the choices.
When AI can produce the code, how do we evaluate what students understand?
From content to creative systems
Don’t just generate an image. Build a world.
github.com/multilingualprogramming
A creative loop
Define rules.
Run variations.
Recognize surprise.
Repeat → branch → vary → observe
Fractals · cellular automata
Probabilistic rule-based worlds.
If variations are endless, what makes one creation worth keeping?
Abundance must not become homogenization
Yes, AI can generate this.
But did my eyes
see it?
Who chose the data? The language? The frame?
Abundance should widen creative diversity, not narrow it.Attention. Experience. Responsibility. Meaning.
Golden hour, Lyon · Photograph by John Samuel