Inside the Next AI Stack: World Models, JEPA, and DPO
About Event TL;DR
A beginner-friendly paper clinic exploring three ideas shaping the future of AI:
JEPA
V-JEPA
DPO
You’ll leave with a clear mental model of a possible next AI architecture stack:
Perception → World Modeling → Alignment
This session is designed to help builders understand where modern AI research is heading and why these ideas matter.
What We’ll Cover (in Order)
1) JEPA — A New Paradigm for Self-Supervised Learning
Instead of reconstructing inputs, Joint-Embedding Predictive Architectures (JEPA) learn by predicting representations of missing information.
2) V-JEPA — Learning World Models from Video
Meta's V-JEPA extends JEPA to video to build world models. Instead of predicting pixels, it predicts future latent representations of scenes.
3) Direct Preference Optimization (DPO)
DPO simplifies the RLHF pipeline used to align language models.
Instead of: Human feedback → reward model → RL optimization
DPO does: Human preferences → direct optimization
The Big Picture
These three ideas map to three core capabilities future AI systems need:
Layer ➡️ Capability Perception ➡️ JEPA World Understanding ➡️ V-JEPA Human Alignment ➡️ DPO
Together they form a possible architecture stack for future AI systems.
We'll discuss:
why reconstruction-based learning may be fading • how world models enable agents and simulation • why alignment methods like DPO simplify LLM training • what these ideas might mean for AGI research
Learning Requirements
Comfortable skimming a technical paper • Curious about agents, robotics, or AI architecture trends • Beginner-friendly — no deep math background required
Who This Is For
AI engineers and researchers
ML students and curious builders
startup founders exploring AI infrastructure
anyone trying to understand the next wave of AI architectures
Suggested Pre-Reads
To get the most out of the session, we suggest skimming these resources beforehand (no need to read every detail):
V-JEPA overview https://ai.meta.com/research/vjepa/#world-models
Direct Preference Optimization paper https://arxiv.org/abs/2301.08243
JEPA paper https://openreview.net/pdf?id=BZ5a1r-kVsf
About AI Scholars
A peer-led learning journey for engineers, students, researchers, and builders. We read new papers, challenge the hype, and turn ideas into practical mental models you can use immediately.Inside the Next AI Stack: World Models, JEPA, and DPO Explained