Nobody agrees on what general intelligence is, which is exactly why nobody agrees on when it arrives. Learn the competing definitions, set your own assumptions, and this console will compute your personal countdown to the two most contested dates of the century.
Begin BriefingThe Turing Test is retired. In its place, the field has fractured into rival paradigms. Your definition of intelligence is the single biggest lever on your timeline, worth more than a decade of variance on its own.
Performance x Generality. DeepMind grades systems by what they can do, not how they do it: capabilities over processes, potential over deployment, and AGI as a continuous path rather than a finish line. Today's frontier LLMs sit at Level 1, "Emerging AGI," with pockets of higher performance in narrow domains. Notably, this framework says AGI does not require consciousness, sentience, or a body.
Narrow, non-AI computation. Your calculator lives here.
Consistently better than an unskilled human across broad cognitive tasks. Current frontier LLMs.
Matches the 50th percentile of skilled adults across domains.
Reaches the 90th percentile of skilled adults.
99th percentile. Better than nearly every human alive.
Outperforms 100% of humanity. The doorway to ASI.
Intelligence as labor substitution. OpenAI's five stages are openly economic: each level describes how much human work the system can absorb. The quiet implication is big. If AGI is defined by its capacity to replace cognitive labor and generate capital, its arrival is not just a science milestone but a macroeconomic shock.
Fluent conversational language, but needs continuous human prompting.
Human-level multi-step problem solving without hand-holding.
Takes initiative and executes long-horizon tasks in the background.
Invents net-new ideas, adding discoveries to human knowledge rather than remixing it.
Runs the strategic, managerial, and operational work of an entire company alone.
Intelligence is skill-acquisition efficiency. François Chollet argues that equating AGI with economic task automation confuses fluid intelligence with sophisticated memorization. Frontier models interpolate brilliantly inside their training distribution and break when forced to extrapolate. His ARC-AGI benchmark resists memorization with visual puzzles built on "core knowledge priors" like object permanence, topology, and counting. The test: if a child learns a rule from three examples and a model needs billions of tokens, the model is doing scale-dependent statistics, not intelligence.
Verdict: scaling alone will not get there. A breakthrough in program synthesis and real-time adaptation is required.Language is not enough. Yann LeCun holds that autoregressive LLMs are structurally doomed for AGI: text is a low-bandwidth compression of reality, and next-token errors compound over long horizons until the system drifts into hallucination. A four-year-old has absorbed roughly 1014 bytes of continuous sensory data through the optic nerve, more than the filtered text corpus of any frontier model. His answer is JEPA-style world models that predict in abstract representation space, enabling cause-and-effect reasoning, hierarchical planning, and physical common sense. Under this view AGI is a spatial milestone, not a linguistic one.
Verdict: AGI needs grounding in the physical world. Timelines stretch accordingly.Beyond parity. ASI is intelligence that vastly exceeds the best human minds across virtually every domain. Nick Bostrom maps three routes there, and adds two warnings: the orthogonality thesis (any level of intelligence can serve any goal, so brilliance does not imply benevolence) and instrumental convergence (nearly any goal creates pressure toward self-preservation and resource acquisition).
Human-grade cognition running orders of magnitude faster. Experiences our world in slow motion; finishes decades of research in minutes.
A coordinated swarm of smaller minds whose aggregate output eclipses any single intellect. Civilization-scale progress, parallelized.
Qualitatively smarter, with cognitive moves we cannot conceive of, the way primates cannot conceive of calculus.
The bridge is recursive self-improvement: an AGI that rewrites its own architecture, each version better at improving the next. Growth like dy/dt = my goes exponential unless it hits diminishing returns, physical experiment bottlenecks, or "model collapse," where training on your own synthetic output erodes real-world competence.
This engine implements a parametric model synthesized from the expert literature: pick a philosophical baseline, then apply real-world modifiers. YAGI = baseline + compute + algorithms + data + infrastructure. YASI = YAGI + takeoff gap. Every choice moves the dial below in real time.
Timeline variance is driven less by chip fabrication forecasts than by ontological disagreement about what intelligence is. Switching from "scaling laws" to "embodiment" moves your date 17 years before you touch a single constraint.
Even a mind running a million times faster than yours must wait for gigawatt data centers to be built, permits to clear, and experiments to run. Physical bottlenecks and the parallelization penalty are humanity's built-in brake on an instant intelligence explosion.
The most rigorous takeoff models land between the extremes: an AGI-to-ASI transition of roughly 1 to 4 years. Fast enough to upend global power structures, slow enough to be measured in months rather than minutes.