AGI is a description of behaviour, not a piece of machinery
Ask five AI researchers how AGI works and you’ll get five different diagrams, most of them sketched on the same napkin. That disagreement isn’t evasion. Artificial general intelligence is defined by what a system can do — handle unfamiliar problems across unrelated domains without being retrained for each one — rather than by a specific blueprint. There is no spec sheet to read off.
What we can do is trace the mechanisms any credible answer needs. Strip away the branding and a general system has to learn, remember, build a model of the world, plan, and check its own output. Everything else is implementation detail.
What a general system has to do that a narrow one can’t
Hand a language model a task it has never seen: “Find me a train to Leeds on Thursday, under £40, aisle seat, and rebook it if the service is cancelled.” Nothing in its training data was labelled “train booking.” It has to infer the goal, work out which tools exist, fill in a payment form, and notice when the result doesn’t match the brief.
That’s a chain of distinct capabilities, and each one can fail on its own.
- Comprehension. Parsing an instruction full of gaps and filling them in sensibly.
- Transfer. Borrowing knowledge from an unrelated domain. Booking a flight teaches you something about booking a train.
- Planning. Splitting the goal into steps before acting, then revising when step three collapses.
- Memory. Keeping what happened last Tuesday available today. Most models start every session cold.
- Self-checking. Spotting that a £62 fare does not satisfy “under £40,” without a human pointing it out.
The relative weight of those traits is contested, and the five traits that genuinely make an AI system general are a useful yardstick for judging any claim that a new model has crossed the line.
The learning loop: pretraining and everything after it
Modern systems learn in stages. Pretraining is the expensive part. A model reads trillions of tokens of text and code and is trained on one deceptively simple objective: predict the next token.
That objective turns out to be a strong lever. To predict text well you have to absorb grammar, factual associations, argument structure, code syntax, idioms, and a rough sense of how ideas follow one another. Nobody designed a curriculum. The signal did the work.
Fine-tuning and reinforcement learning
Raw pretraining produces something closer to an autocomplete engine than an assistant. Post-training reshapes it: examples of good instructions, human preference comparisons, and reinforcement learning against tasks with checkable answers. This is where a model learns to follow a brief, refuse the obviously dangerous request, and keep answers a sensible length.
In-context learning, the closest thing to on-the-job learning
Show a model three worked examples inside the prompt and its behaviour changes immediately, no weight updates required. Context windows now run to hundreds of thousands of tokens, which lets a system hold an entire codebase or a long case file in view while it works. The catch: close the window and everything learned inside it evaporates. That is not learning in the human sense.
It’s why machine learning is a toolbox and AGI is a destination — the techniques are real and powerful, but none of them independently delivers the general behaviour.
World models: the internal picture that makes transfer possible
Here’s the most interesting finding in the field. Researchers trained a small transformer only to predict legal moves in Othello, never showing it a board. Probe the model’s internal states and you can read off the current position of the pieces. Chess models behave the same way. Prediction on the right data forces a compressed model of whatever process generated it.
That is roughly how a general system gets away with answering questions about situations it has never directly encountered. It has an internal simulation, of sorts.
The limit is grounding. A model that has only read about water doesn’t know that a full bucket is heavy. It knows the token “heavy” follows the token “full” in enough sentences. That gap between describing and knowing is one of the hardest problems in the field.
Planning, tools, and agents that audit their own work
The current working pattern for a general agent is a loop: reason about the state of the task, take an action (call a search API, run Python, edit a file), observe the result, and repeat. Give the model a code interpreter and it can test its own arithmetic rather than guessing. Give it a browser and it can check a fact instead of confabulating one.
Verification is where generality gets won or lost
A system that can validate its output outperforms a smarter one that can’t. This is why AI progress looks so uneven. Code has a compiler and unit tests. Maths has proofs. Strategy has a scoreboard. Prose, legal drafting and product judgement have no compiler, so errors survive to the user. Benchmarks built to resist memorisation, like ARC-AGI, show the same thing: models that look fluent on exams still trail humans sharply on genuinely novel puzzles.
Multi-step reliability compounds the problem. A step that works 95% of the time sounds excellent until you chain thirty of them: 0.9530 is roughly 21%. Long-horizon tasks therefore fail at rates that surprise people who only ever tested single prompts.
Why today’s models land somewhere in the middle
Contemporary systems are broad but shallow. They’ll draft a sonnet, debug Python and plan a fortnight in Portugal, then drift on the twentieth repetition of a task they nailed the first time. Sample efficiency is the sore point. A child learns “hot” from one painful touch. A model needs thousands of labelled examples for the same concept, or a very expensive pretraining run that happened to include the right data.
Whether scaling alone closes that gap, or something architecturally new is required, is genuinely unresolved. Serious researchers argue both ways, and the real arguments for and against human-level AI are worth reading in full rather than in summary. What counts as progress is also murkier than headlines suggest — sorting real AGI examples from near-misses takes more care than most coverage admits.
The parts nobody has cracked yet
Continual learning without catastrophic forgetting tops most researchers’ lists. Fine-tune a network on new material and it tends to lose old skills; the weights overwrite. Humans manage this effortlessly and no one has a clean solution at scale.
Grounded causal understanding is the second. Models correlate well and intervene badly. Ask one what happens if you remove the battery from a torch held upside down and you’ll usually get the right answer, but the reasoning is thinner than it appears.
Evaluation is the third. We still measure generality with benchmarks that leak into training data, then argue about the scores for six months.
Beyond all of this sits a separate question about what happens when capability keeps climbing past the human range. That’s a different threshold from the one AGI describes, and where the line between AGI and superintelligence actually falls is a debate worth keeping distinct from the engineering one.
The honest state of play: the ingredients are visible and partly built. A system that maintains a decent internal model of the world, plans over it, acts through tools, and audits its own results will read as general long before it resembles a person. Whether that assembly adds up to intelligence in the way the term is usually meant is the open question nobody has answered yet.