Ask five AI researchers to define AGI and you’ll get five different answers, usually followed by a nervous laugh. That vagueness matters more than it sounds, because you can’t audit progress toward a target nobody has pinned down. It also makes every shortfall look temporary. If the goalposts shift every couple of years, nothing ever truly fails.
The honest picture is more interesting than either the doom posters or the launch-day press releases suggest. There are concrete, measurable obstacles sitting between today’s systems and anything you’d reasonably call general intelligence. Some are physical. Some are mathematical. A few are conceptual in ways that a bigger GPU cluster won’t solve.
First, the definition problem eats everything else
AGI has meant, at various points, a machine that passes the Turing test, a machine that beats humans on every benchmark, a machine that can do any job a person can do, and a machine that generates $100 billion in profit. Those are wildly different bars. Two of them are already arguably cleared. One may be decades away.
This is why the goalposts for AGI keep sliding every time a milestone falls. Chess was once a proxy for general thinking. Then Deep Blue won, and chess became a narrow task overnight. The same thing happened to Go, to protein folding, and to bar exams. Any critique of AGI limitations has to start by admitting the target itself is a moving object.
Compute and power put a ceiling on scaling
Frontier training runs have gone from thousands of dollars to something in the region of $100 million per run in under a decade. A modern frontier cluster can pull tens of megawatts, and that’s before you count cooling. An H100 draws around 700 watts under load; 30,000 of them is a small power station that happens to compute.
The human brain does everything it does on roughly 20 watts. That gap isn’t a fun trivia point. It’s a signal that current architectures are extraordinarily inefficient at whatever it is brains actually do, and that simply buying more electricity has limits imposed by grids, capital, and chip fabrication capacity.
The high-quality data well is emptying
Researchers at Epoch AI have estimated that usable public text data could be substantially exhausted somewhere between 2026 and 2032. Common Crawl holds hundreds of billions of pages, but a large share is spam, duplication, or boilerplate. Scraping harder doesn’t create new knowledge.
Synthetic data is the popular answer, and it works in narrow domains like mathematics and code where correctness can be checked automatically. In open-ended domains, training on model output tends to amplify the model’s own errors. Anyone who has watched successive generations of AI-generated text flatten into the same rhythm has seen a mild version of this drift.
Sample inefficiency is the awkward one
A four-year-old learns what a dog is from a handful of encounters and can then recognise a chihuahua, a Great Dane, and a cartoon. Frontier language models train on more than ten trillion tokens to reach comparable (and in some ways narrower) flexibility.
That ratio, roughly a billion times worse sample efficiency on some tasks, is not a footnote. It suggests current systems are doing something closer to sophisticated pattern retrieval than concept formation. When a model fails at a task with a trivially different surface form, you can see the difference between the two.
No body means no grounded world model
A language model has never dropped a glass, missed a step, or felt a cup go warm. Every physical claim it makes derives from statistical co-occurrence in text written by beings that do have bodies. That works surprisingly well for describing physics and noticeably worse for predicting it.
Robotics data is brutally expensive. Teleoperated manipulation datasets are measured in hours, not petabytes, and hardware breaks. Multimodal video models have improved fast, yet they still generate a person with six fingers handing a glass to a person with seven. Fluency in pixels is not a model of consequence.
Benchmarks reward the wrong behaviour
- Contamination: test questions leak into training corpora, so scores partly measure memorisation.
- Saturation: MMLU climbed from roughly 40% to over 90% in a few years and now separates models poorly.
- Goodhart’s law: optimise hard for a benchmark and you get a benchmark specialist, not a general reasoner.
- Coverage: no standard test measures whether a system can notice it’s been handed a problem it doesn’t understand.
Reliability compounds brutally on long tasks
Here’s an arithmetic problem nobody enjoys. A system that completes each step of a task with 99% reliability succeeds on a 100-step task about 37% of the time. Drop to 95% per step and a 20-step workflow lands around 36%. Autonomy is a compounding bet, and the compounding is unforgiving.
This is exactly why the tools that do earn their keep are scoped narrowly. Generative video is a genuine production asset inside limits like the 125 one-time credits on Runway’s free tier. Writing assistants shine on drafts and rewrites rather than end-to-end publishing, as guides to HyperWrite’s free plan make clear. Design tools such as Canva Magic Studio remove specific friction instead of replacing the designer. Narrow scope is a feature, not an apology.
Memory and continual learning remain unsolved
A million-token context window is not memory. It’s a bigger desk, and it gets cleared every session. Models don’t accumulate experience the way a professional accumulates expertise over fifteen years. Retraining from scratch is the standard update path, which is why catastrophic forgetting is such a persistent headache in research.
Frozen weights also mean a deployed system cannot learn from the mistakes it makes in front of you. It will make the same error on Monday that it made on Friday, cheerfully, forever.
Running models yourself shows both progress and the ceiling
The open-weight ecosystem is the best reality check available. Tools like Ollama for running LLMs locally put a capable assistant on a laptop with two commands, which would have been science fiction in 2019. Quantisation work such as Transformers support for llama.cpp quants keeps shrinking the hardware required for decent output.
Spend an afternoon with a 7B model and the limitation becomes tactile. It handles a summarisation task beautifully and then confidently invents a citation. Capability arrived in patches, not as a general faculty.
What would actually have to change
Several things would need to move at once, and none of them are on a predictable schedule. Continual learning that doesn’t destroy prior knowledge. Sample efficiency within a few orders of magnitude of biological systems. World models grounded in interaction rather than text. Evaluation methods that measure judgment instead of recall. Energy budgets that don’t require a substation per training run.
The likeliest near-term reality is a patchwork: systems that are superhuman at specific reasoning, retrieval, and generation tasks, bolted together with retrieval, tooling, and human review, still prone to confident nonsense at the edges. That patchwork is already enormously useful, and it is not AGI. Holding both ideas at once is the most accurate position available, and it’s the one that keeps you from being surprised in either direction.