AGIartificial general intelligence

The word turns up in vendor announcements, policy consultations, funding rounds and newspaper headlines, and it is doing different work in each. Being able to read it accurately, as an aspiration with no agreed definition rather than a described product, is what separates a sober assessment of a tool from an argument about the future. For a decision you are making this quarter it is a distraction: what matters is what a system does on your actual work, which is measurable today. Understanding the debate is still worth an hour, because it explains why serious people give you opposite answers about the same technology.

There is no agreed definition, and that is not a detail. Some proposals set the bar at breadth, doing most tasks a person can do at a desk. Some set it at economic substitution, doing enough work well enough to replace a role. Some set it at learning, picking up an unfamiliar job from the same instructions a new starter would get. These are different targets, and a system could meet one while failing another.

There is no agreed test either, and proposed tests have a history of being cleared without settling anything. When a benchmark falls, one camp reads it as evidence of progress towards the goal and the other concludes the benchmark was measuring something narrower than it claimed. Both readings can be defensible at once, which is why the argument survives each new result.

The optimistic position, held by researchers at several of the leading laboratories, is that current methods plus more computing, better training and access to tools will get there, and that recent capability gains are the early part of that curve. On this view the remaining problems are engineering problems rather than conceptual ones, and the timescale is short.

The sceptical position, held by other researchers of equal standing, is that current methods have a ceiling. The arguments vary: that these systems learn correlations in material rather than causal structure, that they cannot learn continuously from experience the way people do, that they have no grounding in a physical world. On this view something conceptually new is needed and nobody has it.

A third position holds that the term is too vague to be a research target. Intelligence in people is not one quantity, so a threshold defined against it inherits the vagueness, and a goal nobody can specify cannot be shown to have been reached. Researchers taking this line tend to argue for measuring specific capabilities instead and letting the label look after itself.

What the camps do agree on is worth stating, because it is more than the disagreement suggests. Measured performance on many tasks has risen quickly. The measurements are contested and often flattered by material that leaked into training. And current systems fail in ways people do not, being confidently wrong about things they have no basis for, which is a different shape of error from human mistakes.

Unevenness is the specific problem with the word general. Present systems can handle a question a specialist would find demanding and then come apart on something a child manages, and their confidence does not change between the two. Part of what people mean by general is knowing roughly where your competence ends, and that is not something the current approach supplies.

Nearby words are used loosely and are worth separating. Superintelligence means well beyond a capable person rather than level with one, and is a further step in the same speculative territory. Human-level hides the question of which person doing what: level with a competent professional in one field is a very different claim from level with a person in general.

None of this changes how you should evaluate a tool. The honest test is unchanged whatever anyone believes about the future: give the system work of the kind you actually have, check the output against something you trust, and decide from that. A vendor's position on the long-term question tells you about its ambitions and its investors, not about whether the product will help on Tuesday.

• Hand a general assistant a task well outside anything it was specifically built for and it will usually make a serious attempt, which is the breadth that started this conversation in the first place.

  Claude

• Ask a research assistant a question where two of its sources contradict each other, and judge for yourself whether what comes back is understanding or fluent assembly.

  Perplexity

• In one conversation, give an assistant a task it handles well and one it cannot, and notice that its confidence sounds the same in both, which is the unevenness the argument is about.

  ChatGPT

Is AGI close?

There is no agreed answer and no consensus among researchers. Some at leading laboratories expect it within years and argue current methods will scale to it. Others of equal standing argue those methods have a ceiling and something conceptually new is needed. Anyone giving you a confident date is stating a position, not reporting a finding.

What would actually count as AGI?

Nobody has settled this, which is why the argument persists. Candidate bars include breadth across desk work, replacing an economic role, and learning an unfamiliar job from ordinary instructions. A system could plausibly meet one and fail another, so claims that a threshold has been crossed depend heavily on which definition the claimant chose.

Should this change decisions we are making now?

For nearly all commercial decisions, no. The useful evaluation is unchanged either way: give a system work of the kind you actually have, check the output against something you trust, and buy on that. Long-horizon planning is where the question starts to bear, and even there the honest input is a range rather than a date.

Is AGI the same as superintelligence?

No. The first describes a system level with a capable person across most intellectual work. The second describes one well beyond that, and is a further step into the same speculative territory. The words are frequently used interchangeably in press coverage, which muddles two claims of quite different strength.

Why do AI companies talk about it so much?

Several reasons at once, and they are not all cynical. It states an ambition that attracts researchers and capital, it frames the safety commitments a company wants credit for, and in some cases the people saying it believe it. It is a strategy statement rather than a product description, and reading it as the latter is a category error.

Are today's systems a step towards it?

That is the disputed question rather than the background to it. The optimistic reading is that recent capability gains are the early part of a curve leading there. The sceptical reading is that they are rapid progress along a road that ends somewhere else. Both camps are looking at the same results.

• AGI is a hypothetical system matching a capable person across most intellectual work, not a product on sale.
• There is no agreed definition and no agreed test, which is why proposed benchmarks keep falling without settling anything.
• Researchers of equal standing hold opposite positions on whether current methods can reach it.
• Present systems are uneven and their confidence does not track their competence, which is the specific gap the word general points at.
• For decisions this quarter it changes nothing: evaluate a system on your own work and buy on that.

• Claude

  A general assistant, where the breadth that fuels the argument is easy to observe.

• Perplexity

  Research answers you can check, which is the honest alternative to arguing about capability.

• ChatGPT

  The same unevenness within one conversation, ability and confidence coming apart.