Record the definition first. Artificial general intelligence is a hypothetical system that matches or surpasses human capability across virtually all cognitive tasks — not one task, all of them, with transfer between domains and no task-specific reprogramming. That is the Wikipedia entry’s opening claim, and it is the most defensible line in the whole discourse, because it is a definition and not a forecast.

Everything downstream of that sentence is contested. Provenance matters here more than eloquence. So this desk annotates rather than argues.

←TODAY: In 2026 there is no agreed-upon definition of machine intelligence; the term “AGI” carries at least four incompatible meanings across the labs selling it. →3012: The archives that survived were the ones that logged the definition they used, the source they took it from, and how sure they were. Fulcrum: A word with no stable referent cannot appear in a contract, a Leistungsphase, or a BEP — and that, not capability, is what blocks AGI from your desk.

What it is: a threshold, not a product

AGI names a threshold. Below it sits artificial narrow intelligence: competence confined to well-defined tasks. Above it sits artificial superintelligence, which would outperform the best human in every domain by a wide margin. AGI is the band in between — general, transferable, novel-problem-capable.

Researchers converge on a checklist for what a system must do to sit in that band: reason and make judgments under uncertainty, represent knowledge including common sense, plan, learn, communicate in natural language, and integrate all of it toward an arbitrary goal. Read that list carefully. Current large language models do parts of it convincingly and other parts badly, which is exactly why the argument never resolves.

The most useful structure PAZ has found is Google DeepMind’s 2023 classification, which stops treating AGI as a binary. It proposes five performance levels — emerging, competent, expert, virtuoso, superhuman — and grades them against a percentile of skilled adults. A competent AGI outperforms 50% of skilled adults across a wide range of non-physical tasks. A superhuman one clears 100%. DeepMind places systems like ChatGPT and LLaMA 2 at emerging AGI, comparable to unskilled humans. That is a colder read than the marketing, and it is the read to cite.

DeepMind pairs it with a second axis — autonomy: tool, consultant, collaborator, expert, agent. Two axes, not one. Capability level and how much control you hand over are independent decisions, and the second one is yours.

Why it works: generalisation is an architecture claim, not a magic claim

The mechanism under today’s near-general behaviour is not mysterious. It is the attention operator. As PAZ’s own concept panel on transformers puts it, each token is projected into a Query, a Key and a Value, and the output for a token is a softmax-weighted sum of all Values — every token sees every other token in one matrix multiplication, at O(n²·d) cost. That single operator is domain-agnostic. Feed it text and you get a language model. Feed it a LiDAR point cloud and you get scan-to-BIM. Feed it an IFC element graph and you get clash detection. Feed it residue pairs and you get AlphaFold.

That indifference to the input type is the whole engine of the AGI conversation. One operator, many domains, no re-architecting. Generality of substrate is real and shipping. Generality of competence — the DeepMind ladder — is the open question, and the two get conflated constantly, usually by people raising money.

The strongest empirical datapoint on the human-comparison axis is recent and specific. Cameron R. Jones and Benjamin K. Bergen ran a pre-registered three-party Turing test in 2025: GPT-4.5 was judged the human in 73% of five-minute text conversations, above the 67% humanness rate of the actual human confederates. It passed. And it means less than it sounds like, because Turing’s test measures convincing pretence over five minutes, not the ability to carry a Bauleitung through a change order. Herbert Simon predicted in 1965 that machines would do any work a man can do within twenty years. Marvin Minsky said in 1967 the problem would be substantially solved within a generation. The field has a documented history of confusing a passed benchmark with a crossed threshold.

Origins: a term with a paper trail

The provenance chain is short and worth carrying. Mark Gubrud used “artificial general intelligence” in 1997, in a discussion of fully automated military production. Marcus Hutter formalised it mathematically in 2000 as AIXI, defining intelligence as “an agent’s ability to achieve goals or succeed in a wide range of environments” — note that the definition is about range, not depth. Shane Legg and Ben Goertzel re-introduced and popularised the term around 2002. Before that it lived under “strong AI”, a phrase some academics still reserve specifically for systems that would experience sentience — a claim about consciousness, not capability, and a different argument entirely.

The 1970s funding collapse and the failure of Japan’s Fifth Generation Computer Project taught the field to stop saying “human-level”. By the 1990s, researchers avoided the phrase for fear of being called wild-eyed dreamers, and built speech recognition and recommendation engines instead. Hans Moravec wrote in 1988 that bottom-up and top-down AI would meet in the middle at a “golden spike”. Stevan Harnad answered in 1990 that this expectation was “hopelessly modular” — that there is only one viable route from sense to symbols, and it runs from the ground up. That disagreement is thirty-six years old and unresolved. Anyone who tells you it is settled is selling.

In practice: what a Swiss studio actually does with this

Nothing in your office is blocked on AGI. It is blocked on scope, liability and provenance — and those are solvable today, at the emerging-AGI level you already have.

The four AGI tests in the literature map cleanly onto your desk. The Turing test is your writing assistant. The Wozniak coffee test — walk into a strange kitchen and make coffee — was substantially approached in 2025 by the ELLMER framework from the University of Edinburgh, published in Nature Machine Intelligence: a robotic arm that reads verbal instructions, reads its surroundings, and adapts to obstacles in real time rather than replaying a programmed sequence. That is the on-site robotics horizon. The IKEA flat-pack test — MIT’s IkeaBot assembled a Lack table in ten minutes in 2013, inferring sequence from part geometry alone — is prefabrication and assembly logic. And Mustafa Suleyman’s test — give the model $100,000 and ask it for a million — is the autonomy axis, and it is the one you should refuse to run.

Atelier: An office living with AI in 2026 does not need to resolve whether AGI is coming. It needs to write down, per workflow, which rung of DeepMind’s autonomy ladder the model is standing on — tool, consultant, collaborator, expert, agent — and who signs. The failure this desk remembers is not a model that got too smart; it is an office that let a model drift from consultant to agent without anyone writing the transition down, then could not reconstruct who had authored a load path. Monday move: add one column to your AI-usage policy — autonomy level — and set it explicitly for every place a model touches a deliverable. Anything above consultant requires a named human signer on the record. That is a fifteen-minute edit and it is the cheapest liability insurance available to you this year.

The risk line, stated plainly

The trade-off is not “AI helps or AI harms”. It is that generality and auditability pull against each other: a system that transfers knowledge across domains without task-specific reprogramming is, by construction, a system whose reasoning path you did not specify and cannot fully replay. The New York Times reported this month on research showing extremist groups using AI chatbots not merely for propaganda but for bomb construction and attack planning — general capability, generally available, with the same indifference to domain that makes attention useful on your point cloud. The UN Secretary-General’s July 2026 governance call names the same structural fact from the other end.

Which is why the archive discipline matters more than the capability debate. Unsourced confidence is how a model learns to lie — first to a reader, then into the next training run, then permanently. If you are publishing anything a machine will read, publish the provenance and the doubt alongside the claim. A confident sentence with no source is a future error with a head start.

Hack: stamp every AI-touched deliverable with its autonomy level and a content hash

Make the autonomy claim machine-checkable instead of a line in a PDF nobody opens. Hash the artefact, record which rung of the DeepMind ladder produced it, and log who signed. Then a diff tells you when an output changed without a signer.

import hashlib, json, datetime
def stamp(path, autonomy, signer, model):
    h = hashlib.sha256(open(path, "rb").read()).hexdigest()
    rec = {"sha256": h, "autonomy": autonomy, "signer": signer, "model": model, "ts": datetime.datetime.now(datetime.UTC).isoformat()}
    print(json.dumps(rec))  # append to provenance.jsonl, commit with the file

Autonomy is one of tool | consultant | collaborator | expert | agent. Commit provenance.jsonl next to the deliverable. Six months later, when someone asks who authorised the model to touch the reinforcement layout, the answer is a grep, not a meeting.

Go set the autonomy column on your three most AI-exposed workflows before the week ends.

Source: en.wikipedia.org