01
LLM training
Pretraining data, scaling laws, synthetic data, RL from execution feedback, distillation and post-training systems built for production models.
Registration now open · Paris 2026
GenerationAI
The “NeurIPS” for AI engineers.
The conference for AI management from training to production—and the agent engineering discipline that turns frontier models into dependable systems.
Location
CNIF Forest, Paris
Date
December 1—2, 2026
2
days
80+
speakers
2000+
attendees
08
research tracks
The operating system for applied AI
Manage AI from the first training run to every decision in production.
GenerationAI connects model research to the systems discipline required to operate intelligent products in the real world.
01
Train the model
02
Engineer the agent
03
Operate the system
04
Measure the outcome
Conference theme
Open weight models vs Frontier Model
Defining the future of API
Will applications call closed frontier APIs, or run open-weight models behind their own endpoints? GenerationAI puts that defining architectural choice at the center of every track, talk and hallway conversation.
Agent engineering core
Agent engineering is the new production engineering.
The experience people supervise. The harness connecting tools, memory and context. The loops that plan, recover and finish the job.
Agent experience
Harness engineering
Loop engineering
Who is in the room
Built for people who ship models.
- +AI engineers
- +AI researchers
- +AI-native founders
- +AI tech decision makers
Every session is presented by the team that produced the result, with the methodology on the slide.
8 applied research tracks · in parallel
The complete AI lifecycle. From training clusters to production agents.
02
Agent experience
Designing the human surface of autonomous work: approvals, interruption, legibility, memory, trust calibration and collaborative control.
03
Harness engineering
Tool design, sandboxing, permissions, context assembly, observability and the execution layer that turns a model into a reliable system.
04
Loop engineering
Control flow for non-deterministic workers: planning, retries, reflection, context compaction, recovery and long-running agent loops.
05
Open weights vs frontier
What open-weight releases can now do, where closed frontier models still lead, and how that choice defines the future of the API.
06
Benchmarks & evals
Contamination, saturation, agentic task suites and internal evaluations that remain useful through model and harness changes.
07
Inference systems
Serving economics, KV cache strategy, speculative decoding and latency budgets for multi-step, high-concurrency agent workloads.
08
MCP & interoperability
Model Context Protocol servers in production: capability negotiation, authentication and the emerging interoperable tool ecosystem.
Agent systems
Design the loop, not just the prompt.
LLM training
Train the model for the real world.
80+ speakers
From frontier labs to the teams running models at massive production scale.
Meet researchers building the models, scale-up leaders turning them into fast-growing products, and engineering teams from large enterprises operating AI across complex, high-volume production systems.
OpenAI
Anthropic
Google Gemini
Microsoft Copilot
Mistral
Google DeepMind
HuggingFace
Nvidia
AMD
Intel
Kyutai
Gradio
ElevenLabs
Gladia
Pleias
INRIA
Fraunhofer Institute
Meta AI
Cohere
AI2
Nous Research
Allen Institute for AI
Stability AI
+ many more
December 1–2, 2026 · CNIF Forest, Paris