A model can pass the bar exam, write a working Python script, and summarise a novel it has never seen. Ask that same class of system to run a two-week project without dropping a dependency, and it starts losing the thread somewhere around day three.
That gap is the road to AGI, and it is longer and stranger than most public arguments let on. The useful question isn’t whether we reach the end. It’s which stretches of the route are already paved, which are still swamp, and what evidence would honestly tell you that progress is real rather than performative.
What “AGI” is actually pointing at
Artificial general intelligence usually means a system that can pick up essentially any cognitive task a human can do, at roughly human level or better, including the ability to learn something new without being retrained from scratch. That last clause does an enormous amount of work. A model that aces a hundred benchmarks it was tuned on is broad. It is not general.
The distinction splits the road into two lanes. One is capability: can the system do the hard thing at all? The other is transfer: can it do a hard thing tomorrow, on a task nobody prepared it for, at the same level of reliability? Most of the noise collapses those two lanes into one, which is why the discourse swings so violently between euphoria and dismissal.
Four engines that have to keep running
Strip away the marketing and frontier progress has been powered by a small set of inputs. All of them need to keep moving for the road to continue.
- Compute. Training runs for frontier models grew by orders of magnitude through the 2010s and early 2020s. Inference demand has since become the tighter constraint.
- Data. High-quality human text is finite. Estimates for when the good material runs dry cluster in the late 2020s, which explains the scramble toward synthetic data, video, and proprietary corpora.
- Architecture and training methods. Transformer scale got us this far. Whether it carries us the rest of the way is genuinely unsettled.
- Evaluation. You cannot steer toward a target you cannot measure, and the current suite of benchmarks saturates within about eighteen months of being published.
Compute is expensive, not impossible
The bottleneck keeps migrating. It was chip supply, then it was cluster networking, and now it is increasingly electricity and land. Grid connection queues in Northern Virginia run to multiple years. That is a permitting problem dressed up as a physics problem, and permitting problems get solved more slowly than people expect.
Data is the harder wall
Training on model output risks a slow drift toward blandness, what researchers call model collapse. Careful filtering and human feedback blunt it, but they don’t eliminate it. The genuinely valuable data now is the stuff nobody has scraped yet: structured logs of real work, proprietary process data, physical-world recordings. Whoever holds that holds an advantage that a bigger cluster cannot simply buy.
Evaluation is the quiet bottleneck
A four-point gain on a multiple-choice benchmark tells you almost nothing about whether a system can debug a flaky test suite or notice that a spreadsheet formula is wrong. The field has been improvising harder tests for years, and each new one buys maybe two years of signal before it saturates too.
Where the route narrows
Context windows have grown from roughly 4,000 tokens to over a million. That is impressive and it is not memory. A long context window lets a model look at a lot at once; it doesn’t let it consolidate experience, form durable abstractions, or revise what it knows after deployment. Continual learning remains one of the least solved problems on the whole journey.
Then there is long-horizon work. One widely cited 2025 analysis of benchmark tasks found that the length of job a model could complete with 50% reliability was doubling roughly every seven months. Extrapolate that and the picture looks dramatic. But a 50% success rate is not something you ship. The curve that matters is the one for 99.9%, and nobody has a clean plot of that.
Money decides which road gets built
The big hyperscalers committed well over $300 billion in capital spending for 2025 alone, most of it aimed at AI infrastructure. That number shapes the route more than any research paper. It means the road bends toward whoever can sign gigawatt power contracts, which concentrates the work in a handful of organisations and makes the field’s direction a strategic decision as much as a scientific one.
Reading the signposts without getting fooled
Every few weeks a new roadmap lands, complete with milestone rows and plausible-looking dates. Most of them are marketing documents with a research paper stapled on. It helps to know how to read AGI roadmap updates without getting fooled: check whether the milestone is falsifiable, whether the date is tied to a specific capability, and whether anyone outside the company has reproduced the underlying result.
Public timelines also compress badly in retelling. The honest framing is that the answer to how long until AGI is a range, not a date, and the range is wide enough that planning around a single year is close to useless.
The map keeps getting redrawn
Ask working researchers and you get answers spanning three-quarters of a century. That spread isn’t laziness. It reflects real disagreement about whether scale alone closes the gap, and it’s worth understanding why researcher timelines run from 2027 all the way to 2100 before you treat any single prediction as authoritative.
Dates also drift in a consistent direction when the hard parts turn out to be harder than the demos suggested. That pattern is worth tracking on its own, and the reasons expert AGI dates keep drifting tend to cluster around the same three culprits: reliability, memory, and evaluation.
What would actually have to be true
It’s more productive to work backwards. A system that genuinely deserves the label would need continual learning without catastrophic forgetting, sample efficiency in the range of a human learning a new job in weeks rather than a model needing billions of examples, grounded understanding of the physical and social world, and reliability high enough that you’d let it act unsupervised.
Each of those is a research programme, not a scaling run. If you want the fuller version of that argument, it’s laid out in what would actually have to be true for AGI by 2050, and the list is longer than most roadmaps admit.
The unglamorous last stretch
Here’s the part of the road almost nobody maps. The distance between a system that works 90% of the time and one that works 99.99% of the time is not a 9.99% improvement. It’s the difference between a demo and a product, and it’s where most of the remaining work sits.
That last stretch is unglamorous. It’s error recovery, calibration, knowing when to stop and ask, and behaving predictably under distribution shift. It’s the reason a coding assistant can save you an hour and still cost you an afternoon when it confidently invents a function that doesn’t exist.
Whoever closes that gap first will have done something more consequential than winning a benchmark. They’ll have built a system you can leave alone with a task. Everything before that point is preparation, and the preparation is going to take longer, cost more, and look less cinematic than the highlight reels suggest.