AGI — Artificial General Intelligence
AGI is not a machine that wins at chess; it is a machine that can be handed a strange job on Monday and make useful progress by Friday. The hard part is not one benchmark score. It is transfer: taking skills from code, language, science, planning, tools, and messy human instructions, then using them outside the training script.
The term became popular in the 2000s through Ben Goertzel, Shane Legg, and others who wanted a name for AI that was not trapped inside one task. Legg and Marcus Hutter’s 2007 definition puts the core idea plainly: intelligence is an agent’s ability to achieve goals across many environments.
What the label is trying to measure
The cleanest disagreement is over the test. OpenAI’s 2018 Charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. Legg and Hutter frame it as goal achievement across a wide range of environments. Morris et al. 2024 split the question into levels: performance, generality, and autonomy.
Those are not the same thing. A system can beat most humans at coding benchmarks and still fail at a household repair. It can write a good legal memo and still hallucinate a filing date. It can plan a 40-step workflow, then break because one website changed its button text.
The arrow that matters is not from "more tokens" to "more intelligence." It is from competence in a known setting to competence in a new one.
The timeline fight
The 2023 Expert Survey on Progress in AI collected responses from 2,778 researchers who had published at top AI venues. Its aggregate forecast put a 10% chance of unaided machines outperforming humans in every task by 2027, and a 50% chance by 2047. The range matters more than the median: the field does not agree on what evidence should move the date.
There are two live camps. One reads the progress from GPT-4 onward as evidence that scale, better data, tool use, and inference-time search can keep pushing systems toward generality. The other, associated with critics like Gary Marcus and François Chollet, argues that current systems still lack grounded reasoning, causal models, and durable abstraction.
What's contested
The empirical question is whether current systems fail because they are too small, too weakly trained, or built on the wrong kind of architecture. Scaling-law believers expect new abilities to keep appearing as compute, data quality, and training methods improve. Skeptics expect the curve to flatten on tasks that require causal understanding, long-horizon agency, or learning from sparse experience.
The interpretive question is sharper: if a system solves 80% of economically valuable cognitive tasks but remains brittle on the last 20%, do we call it AGI, or do we reserve the label for something closer to a human apprentice with memory, agency, and embodied common sense?
Why this has to do with other realms
AGI is an AI problem, but it keeps walking into philosophy because "general intelligence" smuggles in a theory of mind. If a model predicts the next token well enough to plan, debug, persuade, and invent experiments, the old boundary between imitation and understanding gets harder to police. That is why this page sits next to concept world as simulation and concept technological singularity, not just realms/ai computing.
It also belongs near economics. If OpenAI’s "economically valuable work" threshold is the chosen test, AGI becomes less like a lab milestone and more like a labor-market instrument. The useful question shifts from "is it conscious?" to "which decisions can it make cheaper, faster, and with fewer hidden errors than a trained person?"
An open question
If the first AGI arrives as a bundle of uneven agents rather than one clean mind, will anyone agree on the date it happened?
Key Sources
- Shane Legg and Marcus Hutter, "Universal Intelligence: A Definition of Machine Intelligence" (Minds and Machines, 2007) — the canonical formal definition.
- Meredith Ringel Morris et al., "Levels of AGI for Operationalizing Progress on the Path to AGI" (ICML, 2024) — a useful ladder for performance, generality, and autonomy.
- OpenAI Charter (2018) — the "most economically valuable work" definition that shaped public AGI language.
- Katja Grace et al., "Thousands of AI Authors on the Future of AI" (arXiv:2401.02843, 2024) — 2,778 researcher forecasts on high-level machine intelligence.
- François Chollet, "On the Measure of Intelligence" (arXiv:1911.01547, 2019) — argues for measuring skill acquisition and abstraction, not memorized task performance.
Further Reading
- concept asi — the next threshold after general intelligence becomes recursive capability.
- concept scaling laws — why some researchers expect smooth capability gains from bigger training runs.
- concept alignment problem — the control problem hiding behind the capability question.
- Artificial General Intelligence edited by Ben Goertzel and Cassio Pennachin (2007) — the early research community naming its target.
- Bubeck et al., "Sparks of Artificial General Intelligence" (arXiv:2303.12712, 2023) — a dated but useful snapshot of the GPT-4 shock.
See Also
- concept asi
- concept technological singularity
- concept scaling laws
- concept alignment problem
- concept world as simulation
- realms/philosophy
Abhishek's take
I do not find the label useful as a trophy. I find it useful as a procurement test: can the system notice a broken assumption, ask for the missing constraint, and repair the decision before a human review catches it? The AGI question that matters to me is not "does it think like us?" It is "which parts of judgment stop being scarce first?"
Tags: #agi #agi-timelines #transformers #large-language-models #alignment #capability #generality