What do we really mean by AGI, and how close are we in practical terms? If GPT-style models feel intelligent, why do they collapse on ARC-AGI?

Inspired from my research dissertation on AGI and experiments with agentic systems, this session offers a grounded, developer-oriented walkthrough of ARC-AGI and its role in evaluating general intelligence in AI systems. We start by clarifying the levels of AI capability, from narrow systems to AGI and ASI and then present benchmarks and most promising approaches.

We will dive into:
- What AGI/ASI mean
- What ARC-AGI is testing (and what it is not)
- Why generalisation, abstraction, and efficiency matter
- How ARC-AGI compares to other benchmarks and evaluation paradigms

The second half zooms out to survey the most promising technical approaches toward AGI, including:
- Agentic and curriculum-based training
- Neuro-symbolic and abstraction-driven systems
- World-model-centric learning
- Hybrid approaches combining learning, reasoning, and memory

The goal is helping developers understand what progress toward AGI actually looks like, what signals are meaningful, and where current systems still fundamentally fall short.