The gradual AGI thesis
"Image synthesis assisted by Qwen, an AI partner within the Global Future Nexus ecosystem."
For Yann LeCun, AGI won't arrive as a thunderclap or a single "eureka" moment. It will emerge through a slow, incremental evolution—a path from mouse-level intelligence to superhuman capability, built not on larger language models but on systems that truly understand the physical world.
The Thunderclap That Isn't Coming
The prevailing narrative of AGI arrival is cinematic: a sudden breakthrough, a formal announcement, a moment when the world collectively recognises that machines have surpassed us. For Yann LeCun, one of the three "godfathers" of modern AI, this framing is not only wrong—it is dangerously misleading.
At the India AI Impact Summit in February 2026, LeCun delivered a characteristically blunt assessment: "There's no such thing as AGI. General intelligence is complete nonsense. There is only human-level AI. It's not going to happen next year. It will take a long time. It's not like we will discover one secret that will unlock intelligence". Progress will be slow, more difficult, and take much longer than the industry hype suggests.
The Seven Stages of Machine Intelligence
LeCun has articulated a concrete roadmap for this gradual evolution—a seven-stage progression that moves from simple learning systems to superhuman intelligence.
Stage 1: Learning how the world works —systems that observe and learn from the environment like a baby animal, developing foundational understanding of physics, causality, and space.
Stage 2: Objective-driven systems —machines that pursue goals while operating within explicit guardrails and safety constraints.
Stage 3: Planning and reasoning —systems that can predict consequences and plan actions to satisfy objectives, rather than merely generating the next token.
Stage 4: Hierarchical planning —the ability to break complex tasks into sub-goals, dramatically expanding problem-solving capability.
Stage 5: Scaling up —systems that progress from mouse-level intelligence to the level of a dog or crow—still far below human capability, but already qualitatively different from today's LLMs.
Stage 6: Training across environments —exposure to diverse tasks and settings to build flexibility.
Stage 7: Superhuman AI —systems that surpass humans in almost all domains, not through a single breakthrough but through the accumulated progress of the previous six stages.
Throughout this progression, LeCun emphasises that safety and control must evolve in parallel. "We'll adjust the guardrails to make those systems controllable and safe as we scale them up". The gradual nature of the transition is precisely what enables this—each stage provides the institutional learning needed for the next.
Why Gradual Matters
The gradualism thesis has profound implications for how we approach AGI governance. If intelligence emerges through steady, incremental progress, then the window for establishing governance frameworks is not a narrow crisis window—it is a multi-decade period of institutional co-evolution. If AGI is not a single point of failure but a continuous process of capability expansion, then governance cannot be a single moment of action—it must be a continuous practice of adaptation.
The LeCun vision suggests that the most urgent task is not preparing for a sudden arrival but building the institutional capacity to adapt alongside incremental progress. This is precisely the kind of anticipatory governance that GFN's framework is designed to enable. The gradual thesis does not reduce urgency—it redefines it as sustained attention rather than crisis response.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
