Artificial general-purpose intelligence
The pursuit of artificial general intelligence risks distracting us from the most important breakthrough of this era: the first general-purpose artificial intelligence.
TL;DR: Generative AI is the first general-purpose artificial intelligence we have created. That is a different claim, and a more defensible one, than saying we are approaching artificial general intelligence (AGI). The two get conflated because both compress to the word 'general'. One is judged by breadth of application, in the tradition of general-purpose technologies such as electricity and the printing press. The other is judged by what the machine is: whether it matches human cognition. The conflation is becoming harder to see because modern AI systems wrap models in harnesses - tool selection, guardrails, oversight - that make the system more general than the model, and the model appear more intelligent than it is. Separating the two matters, because the first is already here: a general-purpose technology for intellectual work, extending the reach of human thinking whether or not AGI ever arrives.
Death of a meme
For two years, one of the internet’s favourite proofs that modern AI is dumb was a single question: how many ‘r’s are in the word ‘strawberry’? Large language models kept getting it wrong. The explanation was straightforward. A model that predicts text from statistical patterns in tokens is the wrong tool for counting letters. But that is not how the failure was read. Each screenshot was offered as a verdict: whatever these systems were, they fell short of real intelligence.
It was the wrong verdict, because it was the wrong test. The word ‘general’ means two different things in the AI debate, and the meme judged one against the other.
Two meanings of ‘general’
The first meaning is the economist’s. In 1995, Timothy Bresnahan and Manuel Trajtenberg coined the term ‘general purpose technologies’ to describe innovations like the steam engine and electricity: technologies that spread across many sectors and spawn waves of complementary innovation.1 A general-purpose technology is judged by breadth of application, not by any property of the technology itself. Nobody asked whether electricity was intelligent. They asked what it could power.
By that test, transformer-based large language models qualify as a general-purpose technology. The same underlying architecture writes, translates, classifies, codes, summarises, tutors, and drafts across virtually every domain of knowledge work.
The second meaning belongs to cognitive science and the ambitions of the AI field. Artificial general intelligence (AGI) is judged not by what the technology can be applied to but by what it is: a machine that matches human cognition in its ability to learn, generalise, and understand across domains. Whether current approaches are on a path to that destination is contested, with researchers on both sides of the question. I set out my position in a previous post (No LLM will achieve general intelligence).
The trouble is that both claims compress to the same word. ‘AI is general’ can mean ‘this technology applies almost everywhere’ or ‘this machine thinks like us’.
Every previous general-purpose technology expanded what people could do. Steam and electricity transformed physical work. Printing transformed the reproduction and spread of information. The computer performed calculations at a speed no human could match, but the thinking stayed with us: it did what it was instructed to do. Generative AI is the first technology to determine for itself how to complete intellectual work - the draft, the analysis, the code. It extends the reach of human thinking the way earlier technologies extended the reach of physical effort.
That is the breakthrough of the current era of AI. It does not require the machine to understand anything for the breakthrough to be real.
The harness illusion
There is a further complication. Almost nobody interacts with a raw model any more.
The chatbot is now a composed system. Behind the conversation sit instructions that shape behaviour, retrieval that fetches current information, code execution for tasks that need precision, safety filters, and increasingly a routing layer that decides which model, or which tool, should handle each part of a request. This collection of logic and controls is increasingly being referred to as the harness. The model generates. The harness decides what gets generated, checks it, supplements it, and sometimes overrules it.
The harness has two effects, and they pull in opposite directions.
The first is that the harness broadens the system’s practical competence beyond that of the model itself. The jagged intelligence exhibited by language models - excellent on one task, confidently wrong on a neighbouring one (Why current AI is both brilliant and dumb) - gets smoothed at the system level. Counting goes to code. Arithmetic goes to a calculator. Current events go to search. Each weakness in the model becomes a routing decision in the harness. It is why the strawberry meme quietly died. The models did not learn to count. The systems were designed to stop asking them.
The second effect is that the harness makes the model appear more intelligent than it is. The demonstration everyone sees is the composed system. Every smoothed edge feeds the narrative that the intelligence inside is approaching something human, when part of what improved was scaffolding: deterministic tools, retrieval, and rules doing what statistical generation cannot.
The harness does not eliminate the jagged edge of large language models. It merely relocates it. Instead of asking whether the model can count, we ask whether the system recognised that counting was required and invoked the correct tool. The risk of error moves from the model’s capability to the harness’s judgement. That makes failures rarer, but also subtler, harder to predict, and often harder to diagnose.
The law already makes the distinction
The distinction between general-purpose and generally intelligent is not only a philosophical one. It already has legal form.
The EU AI Act, the first comprehensive law regulating AI, defines a ‘general-purpose AI model’ as one that displays significant generality and is capable of competently performing a wide range of distinct tasks.2 That is the economist’s definition: breadth of application, not human-like cognition. The Act defines general-purpose AI without ever needing the concept of AGI.
The Act also separates the two layers this post has been describing. The general-purpose model is regulated as one object; the AI system built on top of it is another. The same model can underpin many different systems, each with its own obligations.
That distinction has proved significant in practice. The obligations applying to general-purpose models were implemented broadly as planned.3 By contrast, the use-based obligations applying to AI systems in specific contexts, such as recruitment or credit scoring, have been delayed after policymakers concluded that the standards and assessment infrastructure needed to operationalise them were not yet ready.4
Regulators found the model layer comparatively tractable and the composed-system layer much harder. The law has independently located the difficulty in the same place this post does: not in the model itself, but in the systems we compose around it.
Measuring the wrong thing
The strawberry test judged a general-purpose technology against a general-intelligence claim. It is an old mistake. In the 1980s, Hans Moravec observed: what is hard for us - massive and complex calculations - is easy for machines, and what is easy for us - recognising a cat in a photograph - was, back then, impossible for machines.5 So each generation finds the simplest thing the machine cannot do and treats it as the measure of the whole. A technology does not have to do everything a human can do to change what humans can do.
What matters now is knowing where the model ends and the harness begins, what each can be relied upon to do, and where human judgement still has to sit.
AGI may arrive or it may not. The first general-purpose artificial intelligence already has. The question is not whether the machine is approaching our intelligence. It is what we now do with the reach it gives ours.
Related posts
No LLM will achieve general intelligence (August 2025)
Why current AI is both brilliant and dumb (March 2026)
Scaling our intelligence (June 2026)
References
Bresnahan, T.F. & Trajtenberg, M. (1995). “General Purpose Technologies: ‘Engines of Growth’?” Journal of Econometrics, 65(1), 83-108.
Regulation (EU) 2024/1689 (the EU AI Act), Article 3(63), definition of ‘general-purpose AI model’; see also Recital 97 on the separation of models from AI systems.
European Commission, AI Act implementation timeline: obligations for providers of general-purpose AI models applicable from 2 August 2025.
Digital Omnibus on AI: provisional political agreement reached 7 May 2026; endorsed by the European Parliament on 16 June 2026 and by the Council on 29 June 2026; publication in the Official Journal pending (as of 7 July 2026). Annex III high-risk obligations deferred to 2 December 2027; Annex I to 2 August 2028.
Moravec, H. (1988). Mind Children: The Future of Robot and Human Intelligence. Harvard University Press.