A definition of AGI

Original Article ↗ Hacker News Discussion ↗

Shifting Definitions and Moving Goalposts

  • Many see AGI definitions as perpetually shifting: once a capability is achieved (chess, Turing test), it gets reclassified as “not real intelligence.”
  • Some argue that by older definitions (e.g., “one task = AI, many tasks = AGI”), current LLM systems already qualify as AGI or “baby AGI.”
  • Others counter that LLMs obviously fail stronger interpretations of the Turing test and show non‑human, brittle failure modes; they see today’s systems as impressive but narrow.

What This Paper Proposes

  • The paper defines AGI as “matching the cognitive versatility and proficiency of a well‑educated adult,” operationalized via the Cattell–Horn–Carroll (CHC) model of human cognition.
  • It assesses models across 10 cognitive domains, producing a “jagged” profile; GPT‑4 is ~27%, GPT‑5 ~57–58% of “AGI” on this scale.
  • Commenters find the framework interesting but note issues:
    • “Well‑educated adult” is vague and excludes much of humanity.
    • CHC is tightly tied to human psychometrics and vision/language, possibly unsuitable for non‑human or non‑embodied intelligence.
    • The additive scoring (10 axes × 10%) is seen as arbitrary; a model could be “90% AGI” yet unusable if one axis (e.g., speed, memory) is effectively zero.

Task, Job, and Economic Views of AGI

  • Alternative definitions focus on:
    • Tasks vs jobs (bundles of tasks) and what fraction must be automatable.
    • “AGI is when it can do any job a human can,” or when no non‑manual jobs remain.
    • Autonomy: a system that can be trained like a human, learn continually, pursue goals, and improve itself.
  • Some say most people actually mean ASI or recursive self‑improvement when they say “AGI.”

Limits of Current LLMs

  • Repeated concerns:
    • No continual learning or robust long‑term memory; weights can’t be updated on the fly from experience.
    • Poor data efficiency relative to humans; requires vast training data.
    • Hallucinations and fabricated citations, with no internal concept of truth vs fiction.
    • Dependence on scaffolding, prompts, and human oversight; cannot reliably own a role end‑to‑end (e.g., full job, long‑horizon project).
  • Supporters reply that within those constraints, LLMs already outperform many humans on a wide variety of text-based tasks and can plausibly pass “sloppy” Turing tests.

Intelligence, Consciousness, and Biology

  • Large sub‑thread on whether intelligence and consciousness are:
    • Purely computational (thus substrate‑independent) vs necessarily biological.
    • Tied to awareness, emotions, desire, embodiment, and long developmental learning.
  • Some argue awareness is central but unmeasurable, so psychometric AGI frameworks miss the core; others see that line as unfalsifiable special pleading.
  • Multiple‑intelligences style critiques: the paper covers linguistic/logical domains well but largely omits interpersonal, intrapersonal, bodily‑kinesthetic, and existential capacities.

Skepticism About the Framework and Hype

  • Several view the work as “intelligence SAT for models”: useful for tracking progress but not a true definition of AGI.
  • Others see human‑reference AGI definitions as inherently unstable and politically/financially loaded: useful mainly for marketing and investment narratives.
  • Forecasts like “50% chance of AGI by 2028, 80% by 2030” are compared to rapture predictions—attention‑grabbing but weakly grounded.

Meta: Boredom, Fear, and Fatigue

  • Some are tired of endless AGI definitional debates and compare them to popularized but long‑standing philosophy-of-mind discussions.
  • Others stress that a system able to automate research, self‑improve, or be cheaply replicated at scale would be transformative enough that trying to pin down “AGI” precisely is secondary to grappling with safety, economics, and social impact.