Have Large Language Models Hit a Ceiling? What We Know and Don't

When a curve slows, it is tempting to say: it is over. But a slowing curve does not mean a stopped one, so let us read what the evidence says. In 2020 Kaplan and colleagues described how the loss of a language model falls as its size, data and computation grow, following a power law that held over more than seven orders of magnitude. The meaning is that loss falls slowly, so we must multiply resources again and again to obtain the same amount of improvement.

Cost and electricity

Epoch AI estimates that the computation used to train frontier models has been growing about 5 times a year since 2020 (with a confidence interval of 4 to 6), and that the cost of training has been growing about 3.5 times a year (between 2.8 and 4.4). Behind the cost lies electricity: the International Energy Agency estimates that data centres consumed about 415 terawatt-hours in 2024, roughly 1.5% of the world's electricity, and that consumption could reach about 945 terawatt-hours in 2030 in its base case. The Agency states that uncertainty is large; its scenarios for 2035 range from about 700 to more than 1,700 terawatt-hours. These are real constraints, but not a fixed wall.

Data: how much human text is left?

The paper by Villalobos and colleagues points out that training datasets may reach the size of the public stock of human text between 2026 and 2032 if current trends continue. This is an estimate that depends on continuation and concerns public human text only, and the authors do not regard it as the end of the road; they suggest that generated data, transfer learning from data-rich domains and improved data efficiency could keep progress going. The quality of these alternatives remains an open question, for what matters is not the quantity of text but the meaning, diversity and truthfulness it carries.

Errors: what about hallucination?

OpenAI says, in research it published in September 2025, that hallucination, meaning plausible but false statements, arises from two things: training to predict the next word from text with no true-or-false labels, so that rare facts remain hard to predict; and evaluation systems that reward guessing over admitting ignorance. It holds that hallucination cannot vanish entirely, because some questions have no answer, but that it is not inevitable: it can be reduced by encouraging the model to abstain when uncertain and by adjusting evaluation. This is the verdict of the developer itself, not an independent assessment. On the difference between understanding and imitation, see the article "How Do Large Language Models (LLMs) Like ChatGPT Work?".

Another path to progress

Increase is not confined to inflating the model. Snell and colleagues showed in 2024 that giving a smaller model extra computation at inference time can in some cases, with total computation matched, outperform a model 14 times its size. These are results on specific tasks, and they do not mean the same holds for every task.

So the picture is not a "peak" followed by decline, but a shift of effort between axes: better data, longer thinking at inference time, and higher efficiency. No one knows which path will go furthest, and knowledge of the unseen belongs to God alone.

What does this mean for your decision?

• Do not build a plan on the assumption that every new version will be stronger by the same measure, nor on the assumption that improvement has stopped.
• Test the tool on your actual tasks and measure its results; performance on public tests does not settle the matter.
• Keep human review wherever harm would follow if the machine errs.