Every few months a new model ships, a demo goes viral, and the same argument restarts. One camp says we’re two years out. Another says the whole thing is a parlor trick that will plateau. Both sides cite real evidence, which is part of why the question is so slippery.
So let’s try to answer it properly. Not with a vibe, and not with a single forecast from a lab founder with a financial stake in the answer. Instead: what can these systems do today, where do they still fall apart, and which measurable signals would actually tell us we’re getting close?
What “close” would even mean
AGI is usually defined as a system that can learn and perform any cognitive task a human can, at roughly human level or better. That definition is doing a lot of quiet work. “Any cognitive task” includes things nobody has thought to benchmark, in environments nobody has built. It also implies transfer: pick up a new field from a standing start, without a bespoke training run.
If you want the longer version of how these systems are actually built and where the definition gets fuzzy, there’s a solid primer on what AGI technology actually means and how it’s constructed. The short version: today’s frontier models are extraordinarily good pattern completers trained on a large slice of human text, and that architecture has real consequences for what they can and can’t do.
What the models can genuinely do right now
It’s worth being concrete, because the capability curve has been steep. On GPQA, a set of graduate-level science questions written to be Google-proof, frontier models went from barely above random guessing in early 2023 to outperforming domain PhDs on the multiple-choice version by late 2024. On competition mathematics, systems now score gold-medal range at the International Mathematical Olympiad under contest conditions.
Then there’s the metric that raised the most eyebrows: METR’s measurement of the length of tasks a model can complete reliably. Their estimate put the doubling time at roughly seven months across recent generations, meaning the length of task an AI can finish unsupervised has been growing fast. If that trend held, systems would be handling multi-day projects within a few years.
But the benchmarks keep breaking
Here’s the awkward part. Almost every benchmark built before 2024 has been saturated, contaminated, or quietly abandoned. When researchers tested models on ARC-AGI, a puzzle set designed to resist memorization, the best score sat around 5% for years. In late 2024 a system hit 87.5% on the semi-private set, at a compute cost of thousands of dollars per task. Impressive, and also a reminder that headline scores hide the economics.
FrontierMath tells a similar story. In early 2024, models solved under 2% of problems. A year later, reasoning models were clearing a quarter of them. Real progress, still a long way from the ceiling.
The long tail problem
Raw scores mask reliability. A model that gets 90% of coding tasks right still fails one in ten, and in production that failure is often silent: code that compiles, runs, and does the wrong thing. Agentic benchmarks show this bluntly. On multi-step tasks lasting hours, success rates decay sharply with each additional step, which is why a system that looks autonomous in a five-minute demo needs a human hovering over it after twenty.
Three gaps that haven’t really moved
- Continual learning. A deployed model doesn’t get better at your job from doing your job. Fine-tuning exists, but it’s a heavyweight process, not the effortless on-the-job learning a new hire manages in a week. Everything a model knows is baked in at training time.
- Long-horizon reliability. Error rates compound. Getting a model from 90% to 99.9% per step is a different kind of problem than getting it from 60% to 90%, and the field has far less experience with the second one.
- Grounded world models. Language models learn from descriptions of the world, not from the world. That shows up in physical reasoning, causal inference, and anything requiring trial and error against reality. Robotics progress is real but lags software by a wide margin.
Sample efficiency belongs on that list too. A teenager learns to drive in about twenty hours of supervised practice. No current system comes close to that kind of data efficiency on an open-ended task.
The people making bets, and why their dates keep sliding
Forecasts are everywhere and they disagree wildly. Shane Legg, DeepMind’s co-founder, has put roughly 50% probability on AGI by 2028. Demis Hassabis talks about five to ten years. Yann LeCun thinks the current approach is a dead end and the timeline is much longer, possibly decades. Sam Altman has suggested systems capable of real scientific discovery within a few years.
Notice that the dates have a habit of moving. A breakdown of AGI predictions and why the dates keep shifting is worth reading before you anchor on any single number, because the pattern is instructive: researchers tend to revise targets when a benchmark falls, not when a capability arrives.
It helps to look at specific scenarios rather than one fuzzy date. There’s a rigorous case for AGI by 2030 and what would have to be true for it — the assumptions are checkable, which makes the claim falsifiable in a way that “soon” never is. A longer-horizon view, the numbers behind the AGI by 2040 bet, lays out where the arithmetic starts to strain.
The compute and money reality
Training costs have climbed from tens of thousands of dollars for early GPT models to north of $100 million for frontier runs, with some estimates far higher. Hyperscaler capital expenditure crossed $200 billion in 2024 and kept climbing. That money buys scale, and scale has reliably bought capability so far.
The binding constraint is shifting from chips to power. Data centers are now queuing for grid connections measured in years, and a handful of operators have started signing nuclear agreements. If compute growth slows because you can’t plug in, the capability curve bends with it. That’s a physical constraint with a physical timeline, and it’s one of the few things in this story you can actually forecast.
Signals that would tell us we’re genuinely close
Watch for these rather than for demo videos:
- A model that improves measurably from a week of doing a real job, without a retraining run.
- Task-horizon doubling that holds for another two generations, verified by an outside evaluator.
- An AI that produces a novel, experimentally confirmed scientific result — not a plausible hypothesis, a confirmed one.
- Reliability above 99% on multi-hour agentic work, with the cost per task falling rather than rising.
- Evidence of transfer: a system that picks up an unfamiliar domain from a short briefing and performs at professional level within days.
Any one of those alone is a milestone. Two together would be a genuine shift, and none of them is currently demonstrated.
The honest answer, as of now
We are closer than most working scientists thought possible in 2019, and further than the loudest marketing suggests. The systems are spectacular at compressed, well-specified cognitive work: drafting, summarizing, coding against a clear spec, answering questions inside a knowledge domain. They are shaky at anything requiring sustained autonomy, genuine novelty, or learning on the fly.
The most defensible position sits somewhere between five and fifteen years for something a reasonable person would call AGI, with enormous uncertainty on both ends and a real chance the current paradigm stalls out entirely. What’s changed is that this is no longer a speculative question. There are measurable trends, published evaluations, and falsifiable forecasts. The next eighteen months will produce enough evidence to move the estimate meaningfully in one direction or the other — and that, more than any single prediction, is what makes the question worth tracking closely.