AGI — the goal

Artificial General Intelligence is the stated target of every serious frontier AI lab, the premise of hundreds of billions in capex, and a moving goalpost that gets redrawn every time a model clears the previous line. This page tries to be useful about it: what the word means (and doesn't), who's racing, what the benchmarks show, and where the doubts are.

1. What "AGI" actually means

"Artificial General Intelligence" was coined in 2002 by Shane Legg and Ben Goertzel as a way to distinguish the original dream of AI — a single system that can do any intellectual work a human can — from the narrow task-specific AI that had come to dominate the field. For two decades the phrase lived in fringe conferences. In 2023 it started showing up in 10-K filings.

The problem is that "general intelligence" is a contested concept even among humans. Psychologists debate whether there's a single $g$ factor; evolutionary biologists point to enormous cross-species variation in kinds of cognition; anthropologists note that most human intelligence is socially distributed, not individual. Importing all this into ML was always going to produce disagreement.

What you can safely say: AGI is not a binary. The question isn't "did we reach AGI?" but "on which axes of general-purpose cognitive work does the system match or exceed humans, and at what cost?" Every useful conversation about AGI starts by being specific about the operationalization.

2. The competing definitions

A short tour of the influential definitions you'll see cited:

• OpenAI (2018 charter). "Highly autonomous systems that outperform humans at most economically valuable work." Explicitly economic; vague on which humans and which work. Operationalized by Microsoft's contract as a revenue milestone: $100B in cumulative profit (per reporting in 2023).
• DeepMind (Morris et al. 2024). "An AI system that is at least as capable as a skilled adult human at most tasks." They propose a matrix of performance (5 levels) × generality (narrow vs general). By their own scoring, frontier LLMs are "Level 1 General — Emerging AGI." arXiv:2311.02462.
• Nilsson's employment test (2005). An AI is "general enough" if it could do the jobs that humans actually get hired for — including new ones. A pragmatic economic bar.
• Turing (1950). Pass an extended conversation with a human judge. Already passed in short-form controlled studies (Jones & Bergen 2024). Obsolete as a finish line; useful as a historical milestone.
• Coffee test (Wozniak). A robot that can enter an unfamiliar house and make a cup of coffee. A physical-embodiment benchmark that is still far from solved — which is why "AGI" discussions are mostly about disembodied cognition in 2026.
• Karnofsky's "PASTA" — Process for Automating Scientific and Technological Advancement. An AI capable enough to automate the hard parts of AI research. The self-improvement entry point; explicitly the thing the AI 2027 scenario organizes around.

3. DeepMind's 6 levels — a useful ladder

Morris et al. propose a two-axis framework. Don't treat it as a law; treat it as a shared vocabulary.

Two observations on this table:

• Narrow superhuman systems exist today in limited domains (chess, Go, protein folding, math-olympiad geometry). The gap is generality, not peak performance in any one area.
• Putting frontier LLMs at "Emerging General" is charitable or deflationary depending on your prior. If you care most about general-knowledge fluency, they're past Level 2. If you care most about reliability over long horizons, they're barely Level 1.

4. Who's building it — 2026 field guide

The "serious AGI lab" list is shorter than the "AI company" list. Entry requirements: (1) frontier-scale pretraining compute, (2) a public commitment to building AGI, (3) an in-house research agenda, not just a wrapper.

Two honorable mentions that don't fit the "frontier lab" bucket but matter:

• Mistral AI (Paris) — Europe's sovereign-AI play. Ships both open and commercial models. Strategic, not sheer frontier.
• Tesla FSD / Optimus team — the only serious attempt at embodied, real-world general intelligence at scale. If "AGI must be physical" matters to you, Tesla is the main shop.

5. Benchmark trajectories — the receipts

A few numbers worth memorizing because they show how fast things are moving:

What to take from this table: the benchmarks that used to be "hard" get saturated in 12–18 months. Each time a new one is published, people expect it to survive 5 years and it survives 1. That's the empirical fact anyone who argues for long AGI timelines has to contend with; that's the empirical fact anyone who argues for imminent AGI can't stop pointing at.

But there are also honest caveats:

• Benchmark gaming is real. Anything Internet-scrapeable ends up in pretraining data eventually. "Held-out" is harder than it sounds.
• Exam-style tasks reward a specific cognitive shape. They reward pattern completion on well-posed questions. Long-horizon, ambiguous, multi-tool work is not what MMLU measures.
• Test-time compute is a confound. o3's ARC-AGI performance used ~$2k of compute per task. That's impressive but also clarifies what "reasoning model" really means: throwing more serial compute at one problem.

6. Interactive: METR's time-horizon curve

Arguably the most interesting recent metric is METR's time-horizon measurement — the length of a software task (measured in how long it takes a competent human) that a frontier model can complete with 50% reliability. Between 2019 and 2025, METR reports this number has roughly doubled every 7 months. Drag the slider to see what that trend implies.

7. What might stop it

A sober engineer's list of things that might derail the "AGI by 2030" trajectory. None are fatal on their own; several together would be.

8. Further reading