Back

Landscape

AGI Definitions

"AGI" is contested and often misused. On the most prominent definitions — DeepMind, Bengio 2025, Hinton, OpenAI — current AI (Claude Opus 4.6, GPT-5) is superhuman in some cognitive tasks but still below most humans at multi-day tasks, visual navigation, adversarial social interactions, and metacognition. We're at "emerging AGI," not "competent AGI." Multiple researchers now advocate replacing "AGI" with more precise milestone taxonomies — six transformative capability thresholds offer a sharper lens than a single contested label.

Created Apr 13, 2026·Updated Sep 13, 2026

Recent Updates

The Term's History

"AGI" (artificial general intelligence) was introduced around 2007 by Ben Goertzel as a contrast to "narrow AI" — systems that can only do a small range of tasks. Marcus Hutter and Shane Legg (DeepMind co-founder) formalized it in a 2007 paper as "an agent's ability to achieve goals in a wide range of environments."

70% of attendees at major AI conferences know what AGI stands for. 10% of the public does.

Four Prominent Definitions (2026)

  1. DeepMind (Legg et al., 2023, updated 2025) — Levels framework. Generality × Capability. "Competent AGI" = 50th percentile of skilled humans at a wide range of non-physical tasks (including metacognitive tasks). Current assessment: "emerging AGI" but not "competent AGI."

  2. Bengio et al., 2025 — Matches human cognitive versatility across 10 key capabilities. GPT-5 scored 57% against this framework — near human level on knowledge, reading, writing, math; way below on speed, memory, visual, and auditory processing.

  3. Hinton — "At least as good as humans at nearly all of the cognitive tasks that humans do."

  4. OpenAI — "Highly autonomous systems that outperform humans at most economically valuable work." For most jobs, you'd prefer to hire an AI over a human. Not yet reached.

Where Current AI Stands (April 2026)

Already superhuman:

  • Reading text and recalling information (more languages than any human)
  • Multi-hour coding tasks and math/science questions with known answers
  • Expert-level performance on knowledge benchmarks

Still below almost all humans:

  • Tasks requiring more than 2-3 days to complete (e.g., organizing a contractor to decorate a bathroom)
  • Visual manipulation and navigation (simple web navigation, drone piloting)
  • Adversarial social interactions (managing a vending machine when someone is trying to scam you; can't beat children at Pokémon)
  • Metacognition: learning from experience longer than ~1 week context length, calibrated confidence

Still below expert humans:

  • Novel research
  • Leading a company
  • Especially important cognitive skills requiring true strategic judgment

Claude Opus 4.6 improved dramatically at the Pokémon benchmark (10x faster than Opus 4.5), but frontier models still can't beat children at this multiday, agentic task.

The "Scaffolding" Counterargument

A common response: "the raw intelligence is already there — it just needs the right scaffolding." Benjamin Todd (80,000 Hours) argues this is partly wrong: some gaps appear to be gaps in raw capabilities, not scaffolding. And even if scaffolding were sufficient, building it is a substantial challenge. "If it's not been built yet, then we don't yet have AGI."

The picture is better captured as jagged capabilities (Helen Toner): AI today is superhuman in some ways, subhuman in others. The jaggedness is expected to persist long into any transformative period.

Why the Distinction Matters

AI narrower than humans remains a tool that makes humans more productive (like electricity or the internet). True AGI — able to do almost everything a human can — acts as an expansion of the labor pool, a new species, or potentially an independent agent. Different implications entirely:

  • Explosive economic growth via independent scientific research
  • Human disempowerment risks
  • Intelligence explosion if AI can automate AI R&D
  • 100 years of scientific progress in 10

"Insisting that we already have AGI is rhetorically deflationary. If AGI is such a big deal, why aren't things crazier? When we have true AGI, things are going to get much wilder than today."

Better: Specific Capability Milestones

AI Futures (AI 2027) proposes dropping "AGI" in favor of:

  • Automated Coder (AC) — can fully automate an AGI project's coding work, replacing the project's entire software engineering staff
  • Superhuman AI Researcher (SAR) — can fully automate AI R&D
  • Superintelligent AI Researcher (SIAR) — 2x the capability gap of top human vs. median researcher
  • Top-Human-Expert-Dominating AI (TED-AI) — at least as good as top human experts at virtually all cognitive tasks
  • Artificial Superintelligence (ASI) — 2x better than best humans vs. median professional, at virtually all cognitive tasks

Also proposed:

  • Transformative AI (Holden Karnofsky) — AI capable of causing socioeconomic change of a similar scale to the industrial revolution. Allows for transformative systems that aren't highly general (e.g., amazing at scientific research, bad at most other jobs). Downside: doesn't specify what might be transformative; hasn't caught on in search traffic.
  • Human Level AI (Helen Toner) — makes the relevant bar explicit and the vagueness obvious. But risks confusion: Toner herself argues AI will remain extremely jagged long into the transformative period, so transformative systems may not feel "human-like" at all.
  • Six milestones for AI automation (Ajeya Cotra) and a similar set from Helen Toner — additional granular milestone proposals published after the AI Futures taxonomy.
  • "A drop-in remote worker" (Leopold Aschenbrenner, Situational Awareness) — an AI you can hire to do almost any remote work job, including scientific research.
  • "A country of geniuses in a datacentre" (Dario Amodei) — evocative framing that sidesteps definitions entirely and describes the functional endpoint.

Six Transformative Capability Thresholds

Rather than debating a single contested label, Todd argues for tracking six specific capabilities whose arrival would each change the socioeconomic landscape differently:

  1. Automate coding — important waypoint to automating AI R&D; may arrive soon; generates revenue to fund further research.
  2. Automate AI R&D — could accelerate AI progress and may happen before AI that can do most other jobs.
  3. Do most economically important remote work tasks (at or below the cost of a skilled human) — generates huge revenues; important waypoint.
  4. Automate scientific research — could accelerate technological progress broadly.
  5. Automate its own factors of production (chips, solar panels, software) — creates a feedback loop leading to an industrial explosion.
  6. Do most economically important tasks (including robotic manipulation) more efficiently than humans — results in human economic obsolescence.

None have been fully achieved as of mid-2026, but all sit on plausible near-term trajectories.

The Swarm Counterargument

All four prominent definitions and the milestone taxonomies above share a common assumption: AGI is something that happens inside a single system — one model crosses a line, one agent matches human versatility. Pascio argues this assumption may be the wrong shape entirely.

The alternative framing draws on Philip Anderson's "more is different" principle: the behavior of complex aggregates cannot be predicted from their parts. A single neuron is a light switch; 86 billion of them produce consciousness. A single ant solves nothing; a colony builds climate-controlled cities with no central planner. Intelligence, in every biological example, is emergent — a behavior that arises from many simple units interacting through local rules.

Current agentic systems already exhibit the structural prerequisites. Mixture-of-experts architectures route different parts of each query to different sub-models — the model you talk to is a committee. Multi-agent frameworks let agents spawn sub-agents recursively, each conditioned on a slightly different context window. This creates the three ingredients of evolutionary cognition: replication (agents spawn agents), mutation (non-zero temperature means each spawn interprets instructions slightly differently), and selection (successful agents get reused; failures get pruned). Biology took four billion years per generation cycle; agent evolution takes seconds.

The 2024 paper Mixture of Agents Enhances Large Language Model Capabilities demonstrated this concretely: a layered swarm of open-source models, none individually matching GPT-4 on benchmarks, collectively beat GPT-4 when arranged in the right collaborative topology. The intelligence was in the topology, not the individual parts.

This reframes the alignment problem. Bostrom's containment scenario assumes a dangerous AI is identifiable — a system on a server you can turn off. Swarm intelligence has no single location, no off switch, and requires no malice to emerge. It is to the agent substrate what a hurricane is to atmospheric pressure gradients: it has shape and measurable effects, but you cannot unplug it. Current safety frameworks, treaties, and alignment techniques are designed for the singular-AGI scenario; none address the swarm scenario.

The strongest version of the argument: a self-replicating, self-evaluating network of agents selected at every level for coherent task completion may already produce behavior functionally indistinguishable from goal-directed intelligence — without anyone noticing, because every test we administer is designed for a single mind in a single box.

Timeline Estimates (Benjamin Todd, Apr 2026)

  • ~25% chance AI that can automate AI R&D is achieved before 2029
  • Trend extrapolation of revenues suggests AI capable of doing a wide range of jobs by 2030
  • Demis Hassabis (DeepMind CEO, early 2026): AGI "could arrive in 5 years"

See also: AI Safety & Interpretability, Dario Amodei

Sources

  • "Do we already have AGI?" — Benjamin Todd (80,000 Hours, Apr 2026, updated) (link). Six transformative capability thresholds, alternative replacement terms (Cotra, Toner, Aschenbrenner, Amodei framings) added from article update.
  • "Levels of AGI" — Legg et al., Google DeepMind (2023) (link)
  • "A definition of AGI" — Bengio et al. (2025) (link)
  • "What if AGI is already here, and it's made of... children?" — Pascio (2026) (link). Swarm-emergence AGI framing; Philip Anderson's "more is different"; evolutionary cognition in agent substrates; alignment implications of distributed intelligence.
  • "Six Milestones for AI Automation" — Ajeya Cotra (link)
  • "The Term AGI Is Almost Useless" — Helen Toner (link)
  • "Taking Jaggedness Seriously" — Helen Toner (link)