In March 2023, a Microsoft research team published a 155-page paper called “Sparks of Artificial General Intelligence,” arguing that GPT-4 showed glimmers of general reasoning. Within days, other researchers were publishing rebuttals. That argument is the whole story of AGI in miniature: the term describes something enormous, nobody agrees where the line sits, and the disagreement is as much about definitions as about technology.
Here’s the useful version. Artificial general intelligence is the idea of a machine that can learn and carry out any intellectual task a person can, at roughly human level or better, without being rebuilt for each new job. Not a system that plays chess brilliantly and nothing else. A system that could pick up chess, tax law, or conversational Portuguese from the same starting point.
The three ingredients that separate AGI from everything built so far
Most working definitions circle the same three properties. Strip away the philosophical padding and you get:
- Generality. Skill transfers across domains. A system that learns to plan a wedding can apply the same planning machinery to a software migration.
- Autonomy. It acquires new skills without a team of engineers retraining it on a curated dataset. You describe the task and it works out the rest.
- Human parity. Performance lands around or above what a competent adult manages, across the full spread of intellectual work rather than a chosen slice of it.
Miss any one of these and you have something impressive that isn’t AGI. A system that matches humans on 90% of tasks but needs a fresh training run for each one is a very good tool. A system that learns anything but does it at the level of a distracted eight-year-old is a demo.
Where the definition gets slippery
“Human level” sounds precise until you try to pin it down. Average of which humans? A median adult, a PhD in the relevant field, or the best expert alive? Then there’s the scope question: does general intelligence include reading a room, or parallel parking? Most benchmarks quietly exclude anything needing a body or social intuition, which makes the target easier to hit than the phrase suggests. If you want a plain-English answer behind the hype, this is the honest starting point.
Narrow AI versus AGI, in concrete terms
Nearly every AI system in commercial use is narrow: excellent at one job, helpless one step outside it.
- AlphaFold predicted structures for roughly 200 million proteins and reshaped biology. It cannot book a flight.
- Chess engines have outclassed world champions since 1997. Ask one to explain a weak opening in plain language and it has nothing.
- Recommendation systems shape a huge share of what you watch and buy. They have no idea what a recommendation is.
Large language models blur this line because they are genuinely general-purpose. One model writes SQL, drafts a lease clause, and debugs Python. That range is new and real. It still isn’t general intelligence, because the range comes from absorbing a large slice of the internet, not from an ability to acquire unfamiliar skills on the fly. The definitional fight over what we’re actually building is worth understanding before you take any timeline seriously.
Why the goalposts keep sprinting away
There’s a running joke in AI research: any task a machine can do stops counting as intelligence. It stings because it’s half true.
MMLU, a 14,000-question exam spanning 57 subjects, launched in 2020 as a hard test of broad knowledge. By 2024, frontier models were scoring in the high 80s and the benchmark was effectively retired. Researchers swapped in GPQA, written by PhDs to be Google-proof, then FrontierMath, then Humanity’s Last Exam. Each new test falls faster than the last.
ARC-AGI is the most instructive case. Designed in 2019 to measure raw pattern-finding on simple visual puzzles that resist memorisation, it kept models near zero for years while humans scored around 85%. In December 2024, OpenAI’s o3 hit roughly 76% on the semi-private evaluation set, at a reported cost of about $20 of compute per task. Both things are true: that’s a real jump, and it’s a narrow, expensive slice of what a person does before lunch. The moving-goalpost problem explains a lot of why expert forecasts disagree so wildly.
Four roadmaps, and why researchers split on them
Scale the existing recipe
The bet behind most frontier labs: more compute, more data, more parameters, plus reinforcement learning on reasoning traces. GPT-3 had 175 billion parameters in 2020; successors are larger and trained on far more text. Believers point to abilities that appeared only past a certain scale. Sceptics point to brutal cost curves and a finite supply of good training data.
Build something structurally different
Others argue transformers lack a key ingredient, whether that’s persistent memory, causal world models, or a modular design that separates reasoning from recall. Yann LeCun has pressed this case for years, and the fact that nobody has shipped a clearly better architecture keeps the argument alive.
Embodiment and world models
Intelligence evolved to control bodies in physical space, and roboticists take that seriously. A system that can’t predict what happens when you tilt a cup may be pattern-matching rather than understanding.
Agents and tool use
The pragmatic route: keep the model as a reasoning core and let it call search engines, code interpreters, and other models. It patches over real weaknesses, though it doesn’t produce a system that teaches itself.
What today’s best systems still cannot do
Four gaps show up again and again in evaluations:
- Continual learning. Models are frozen after training. They don’t get better at your job by doing it for a year the way a junior analyst does.
- Sample efficiency. A teenager learns to drive safely in about 20 hours behind the wheel. Training runs burn trillions of tokens to reach comparable competence.
- Long-horizon reliability. Errors compound. A 95% success rate per step leaves roughly a 13% chance of clearing 40 steps cleanly.
- Calibration. Systems state wrong answers with the same confidence as right ones, which is a problem when the work matters.
Working memory is one area moving fast. Models that hold a million tokens of context, such as Kimi’s long-context assistant, can keep an entire codebase or contract set in view while they reason. That genuinely helps. It doesn’t change the fact that the model stops learning the moment the conversation ends.
How to read a timeline claim
A 2023 survey of 2,778 AI researchers put the median forecast for human-level machine intelligence around 2047, with enormous spread. Lab CEOs say sooner, in the range of a few years. Prediction markets land somewhere in between.
The gap isn’t stupidity on anyone’s part. It’s disagreement about definitions (does AGI need to act in the world? does it need to be cheap?) plus a difference in incentives, since timelines attract attention and investment. When you see a date, ask what specific capability is being claimed and what result would prove the claim wrong.
A short checklist for judging AGI claims
Next time a headline announces that AGI has arrived, run it through this:
- Does the claim rest on a benchmark, and was that benchmark designed before or after the result came in?
- Can the system complete the task without a human breaking it into steps?
- How much compute did the demonstration cost, and does that survive contact with real usage?
- What happens when the task shifts slightly off the tested distribution?
- Did the system learn during the trial, or was it retrained afterwards and quietly re-run?
Those five questions filter out most of the noise. What’s left is a fair picture: today’s models are the most capable general-purpose tools anyone has built, doing real work in writing, analysis, and code, and they are not yet agents that can teach themselves a new field over a weekend. Teams that plan around the tools already in front of them tend to do better than teams waiting for the version in the pitch deck.