# What is AGI, and are we close to it? By Amanda, cofounder of [Wistkey](https://wistkey.com) · 13 Sep 2026 · AI Canonical: https://styamanda.ai/blog/what-is-agi/ Players keep asking whether AGI is about to spawn. The awkward answer is that nobody agrees on the boss's exact stats, several guilds sell maps to its cave, and the release date remains an estimate. AGI is a useful idea—but only after we say what we mean by it. ## AGI means broad, human-level capability **Artificial general intelligence, or AGI, usually means an AI system that can perform a wide range of cognitive tasks at roughly human level or better—not merely excel at one benchmark or profession.** We are closer in some abilities than we were a few years ago, but there is no accepted test that says "AGI achieved", and no reliable date for when every important gap will close. The word *general* does most of the work. A chess engine can beat every human at chess and still cannot plan a holiday. A language model can write code, explain a contract, inspect an image, and use tools, which feels much broader. Yet breadth in a demo is not the same as dependable competence across unfamiliar, messy, long-running work. Even leading labs use different thresholds. The [OpenAI Charter](https://openai.com/charter/) describes AGI as highly autonomous systems that outperform humans at most economically valuable work. A [Google DeepMind research framework](https://deepmind.google/research/publications/66938/) instead separates generality from performance and treats autonomy as another deployment dimension. Those are related ideas, not one universal measuring stick. > AGI is not one magic score. It is a claim about breadth, competence, reliability, and independence all at once. ## Why today's AI can look like AGI Modern models are unusually broad. One system can move between writing, coding, visual interpretation, research, translation, and planning without being rebuilt for each task. Add tools, memory, and a loop that lets it act, check, and try again, and the result starts to resemble a capable digital colleague. That is a real change. As I noted in [what AI agents can actually do now](https://styamanda.ai/blog/what-ai-agents-can-do-now/), an agent can already complete useful multi-step work when the goal, tools, permissions, and stopping rules are clear. At [Wistkey](https://wistkey.com), this practical layer matters more than winning a naming argument: can the system complete the workflow safely, and can a person verify the result? But a system may be superhuman on one task, competent on twenty, and strangely brittle on the twenty-first. Fluency hides that uneven map. When the answer arrives in polished prose, it is easy to confuse "sounds like a person" with "can reliably do what a person can do". ## What is still missing? There is no single missing component labelled AGI. The remaining problems overlap: - **Reliability:** success once is not the same as success repeatedly, especially when small mistakes compound over a long task. - **Transfer:** a general system should apply what it learned to genuinely new situations, not only variations that resemble its training and tests. - **Long-horizon work:** real projects involve changing requirements, partial information, interruptions, and decisions whose consequences appear hours or days later. - **Grounding:** confident language must stay connected to evidence, tools, and the physical or organisational world. - **Judgment:** knowing what to optimise, when to ask, when to stop, and which trade-off a person actually wants is not captured by getting a tidy final answer. - **Safe autonomy:** capability and permission are different stats. A system able to act widely also needs limits, monitoring, and recovery when it goes wrong. These gaps are hard to compress into a leaderboard. [ARC-AGI-2](https://arcprize.org/arc-agi/2), for example, tests unfamiliar visual reasoning tasks that people can solve but that expose failures in symbolic interpretation, combining rules, and applying rules differently by context. It is useful evidence about one kind of generalisation. Passing it would still not prove that a system can run a company, care for a child, repair a boiler, or make wise decisions under uncertainty. ## Why AGI timelines disagree Predictions differ partly because forecasters are predicting different things. One person means "better than most humans at most screen-based knowledge work". Another means "able to learn almost any intellectual task a human can". A third quietly adds robotics, consciousness, or complete economic independence. Those finish lines can be years apart even if everyone agrees about the underlying technology. Progress is also lumpy. Models can gain a striking capability between releases, while stubborn failures remain. Better tools and agent scaffolding may improve real work without changing the base model very much. Conversely, a benchmark jump may not survive contact with a cluttered inbox, an undocumented codebase, or a colleague who changes the brief halfway through. Researchers are building better rulers. [METR's task-completion time horizon](https://metr.org/time-horizons/) estimates the difficulty of software tasks an agent can complete at a stated reliability level, using the time a human expert would need as a proxy. METR explicitly warns that this is not simply how long an AI can run, and that results depend on the task suite. That caveat is the point: a measurement can illuminate one slice of progress without becoming a universal AGI clock. ## How to judge an AGI claim When the next headline says AGI has arrived—or will arrive next Tuesday—ask five questions: - **What definition is being used?** Look for an operational threshold, not a circular phrase such as "AI that is generally intelligent". - **How broad is the evidence?** One exam, game, or coding benchmark shows a capability, not generality. - **How reliable is it?** Check repeated performance, failure rates, human intervention, and whether success depends on cherry-picked examples. - **What sits around the model?** Tools, retrieval, memory, prompts, compute budgets, and human reviewers may be doing essential work. That does not make the result fake, but it changes what has been demonstrated. - **Can it handle novelty and mess?** The strongest evidence comes from unfamiliar tasks and real environments where the route is not pre-labelled. This is the same discipline needed for [AI evaluation more broadly](https://styamanda.ai/blog/ai-evaluation-explained/): define the task, test the actual system, inspect the failures, and resist turning a convenient score into a claim it cannot support. ## So, are we close? We are clearly closer to broadly capable AI systems, and useful autonomy is improving quickly. We are not close to agreement on what observation would settle the AGI question. Anyone offering a precise arrival date is combining evidence with assumptions about definitions, future research, engineering, cost, deployment, and human institutions. The practical response is neither to dismiss AGI nor to wait for a ceremonial unlock screen. Track capabilities that affect your work now. Build evaluation and oversight before systems become more autonomous. Keep plans that still function if progress is faster—or slower—than the loudest forecast. My quest log therefore has no "defeat AGI" checkbox. It has smaller entries: define the job, grant the minimum permissions, test the ugly cases, keep a human able to interrupt, and update the map when the terrain moves. Less cinematic, perhaps. Considerably better survival odds.