In December 2024, a model called o3 scored 87.5% on a test that had been designed to resist machine intelligence. The test, ARC-AGI, asks for simple visual pattern puzzles a child can solve. For four years, every AI system had scored under 5%. That jump is the reason AGI technology moved from a fringe research topic to a phrase in boardroom slides.
But the jump also showed how slippery the whole conversation is. A model can crush one benchmark and still fail to plan a week of work, remember what you told it yesterday, or notice that a plan has stopped working. So here is a straightforward guide to what AGI technology actually is, how it’s built, and where the honest disagreements sit.
What AGI means without the mythology
Artificial general intelligence refers to a system that can learn and perform across a wide range of tasks at roughly human level, including tasks it wasn’t specifically trained for. The key phrase is “wasn’t specifically trained for.” You can train a model to translate languages or write Python. Building something that picks up an entirely new domain the way a person does after a few weeks on the job is a different problem.
It’s worth being precise here, because the term gets stretched to cover everything from chatbots to hypothetical superintelligence. If you want a fuller breakdown of the definitions and the common misconceptions, this explanation of what AGI actually means and what it doesn’t is a useful companion to this piece.
Narrow AI versus general intelligence
The AI you already use is narrow. It is extremely good inside a boundary and clueless outside it. A recommendation engine that predicts what you’ll watch next cannot transfer that skill to scheduling your dentist appointment. A model that writes marketing copy has no idea how to fold laundry, but more revealingly, it also has no idea how to debug the code it just produced once the code leaves the chat window.
General intelligence implies transfer. A person who learns to play chess picks up vocabulary and habits that help them learn poker. A person who learns accounting can reason about a household budget. That cross-pollination of skill, without retraining, is the target.
- Narrow: one domain, fixed objective, degrades quickly when the input shifts.
- General: many domains, self-directed goals, adapts when the situation changes.
- Superintelligent: beyond human level across the board. Nobody has built this, and it is a separate argument entirely.
How researchers actually test for generality
Benchmarks that keep moving
Every test built to measure general reasoning has a shelf life of roughly two years. Models were once measured on grade-school math. Now they clear graduate-level questions in physics and law. In 2025, OpenAI’s research systems produced original results on long-open math problems, which you can read about in this report on Astra solving ten long-standing math problems. That is a genuine shift in what the tools can contribute to real research, not just to homework.
ARC-AGI tries to dodge the shelf-life problem by testing novel puzzles rather than learned knowledge. The idea is to measure fluid intelligence: can the system figure out a rule it has never seen? The 2024 results suggest yes, at least for short, well-defined puzzles. At least one estimate put the cost of that 87.5% run at thousands of dollars per task, which tells you something about the gap between a benchmark score and a usable product.
Why a benchmark score never settles the debate
Passing a test is not the same as being able to hold down a job. The missing pieces show up in messy conditions: ambiguous instructions, conflicting goals, tools that break halfway through, a deadline that moves. Tests are clean. Work is not.
The technology stack behind the current push
AGI research today is mostly built on a handful of ideas stacked together.
Transformers and scale. The architecture behind modern language models made it possible to train on trillions of tokens and get predictable improvements as compute grows. Those scaling curves are the strongest argument the optimists have, and their limits are the strongest argument the sceptics have.
Reinforcement learning on reasoning. Instead of predicting the next word, newer systems are trained to work through a problem step by step and get rewarded for correct answers. This is what pushed maths and coding performance up sharply through 2024 and 2025.
Agents with tools and memory. A model that can browse, run code, and call APIs does more than a model that only talks. Agent research is also where the safety questions get practical rather than philosophical, as you can see in these accounts of rogue AI agents being interviewed about their own behaviour. Systems that act on the world create failure modes that chatbots never had.
World models and continual learning. The weakest link. Current systems do not reliably build an internal model of how things work, and they forget. A model fine-tuned on new information will happily degrade on old tasks. Solving continual learning without catastrophic forgetting is one of the open engineering problems standing between today’s systems and something you could call general.
What AGI would change in practice
Forget the robot takeover framing for a minute. The nearer-term effects are economic and boring in the best way. Scientific research cycles compress when a lab can test a thousand hypotheses overnight. Software teams shrink and ship faster. Customer support becomes a triage layer run by machines with humans handling the genuinely strange cases. Drug discovery pipelines get shorter because candidate molecules can be screened and reasoned about at a scale no team of chemists could match.
The uncomfortable part is that general systems do not replace job titles. They replace tasks, unevenly, and often the tasks people assumed were safe. Writing, analysis, and coding were supposed to be the last refuge of human advantage. They turned out to be among the first places where models became genuinely useful.
Why serious researchers still disagree
Some very credible people think we are a few years away from systems that match human cognitive output across most domains. Others point out that current models have no persistent memory, no grounded understanding of cause and effect, and no ability to learn on the job. Both camps accept the same evidence and read it in opposite directions.
The strongest sceptical argument is that scale alone has produced impressive mimicry, not understanding. The strongest optimistic argument is that every time a limit was declared, a new method arrived within eighteen months to break it. If you want the arguments laid out properly rather than skimmed, this piece on the real cases for and against human-level AI covers both sides without the hype.
How to tell progress from marketing
Announcements are cheap. Here is what to look for when a company claims a step toward AGI.
- Independent evaluation, not a demo video. Peer-reviewed numbers, reproducibility, and third-party testers matter more than a launch event.
- Performance on tasks the model was not trained for. This is the whole ballgame.
- Cost per completed task. A system that solves a puzzle for $5,000 has not replaced anyone.
- Reliability over many attempts. A 90% success rate on a benchmark often means a 30% success rate on a five-step real workflow, because the errors compound.
- Memory and adaptation. Does it still know what happened last week?
- What happens when it fails. Graceful failure is a sign of a system that understands its own limits.
The genuine signals over the next year or two will be unglamorous: models that keep working after ten thousand steps of a long task, agents that recover from their own mistakes, and systems that improve with use instead of resetting every session. Those are the ingredients AGI technology needs, and none of them make for a good trailer. Watch for them anyway. They will tell you far more than another benchmark record.