What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
There is no reliable date for when AI models will stop improving, and current evidence does not show that a permanent plateau is imminent. Some gains from scaling familiar training methods may slow or run into limits, but progress can also come from better algorithms, training data, post-training and inference methods. A slowdown in one model family or benchmark is not proof that AI as a whole has reached a ceiling.
What would it mean for AI models to be stuck?
“AI has plateaued” can describe several different things, and they are not interchangeable. Training loss might improve more slowly; a model family might gain less from additional pretraining compute; performance on a particular test might stop rising; or users might see little improvement in practical usefulness. Evidence for one of these does not automatically establish the others.
As an Amazon Associate I earn from qualifying purchases.
A limit to one method is not a limit to every method
The key question is whether a claim concerns pretraining scale, post-training, inference-time techniques, or a particular task. If a familiar way of making a model larger produces smaller gains, that is evidence about that scaling approach under those conditions—not proof that no other technique can improve capability.
Capability is harder to forecast than broad trends
The UK-led International Scientific Report on the Safety of Advanced AI says aggregate performance across many tasks can be partly predicted from model scale, while specific capabilities cannot currently be forecast reliably far in advance. Broad trends therefore support more confidence than predictions about when a particular skill will appear.
#1 Best Overall
What has driven improvement, and what could slow it?
Progress has reflected interacting changes in training compute, training data, and AI techniques—not parameter count alone. The historical rates below describe observed change, not a rule that must continue.
| Measure | Historical rate reported | How to interpret it |
|---|---|---|
| Frontier-model parameters | 2.4× per year since 2010 | OECD’s 2026 report describes historical growth; it does not guarantee the same rate in the future. |
| Training data | 2.6× per year since 2010 | OECD’s 2026 report describes historical growth in the amount used for frontier models, not an unlimited supply of suitable data. |
| Training compute | More than 4× per year since 2010 | OECD’s 2026 report describes historical growth in compute used to train frontier models, not a direct measure of capability. |
These rates come from OECD’s 2026 report on possible AI trajectories through 2030. The OECD cautions that scaling relationships are patterns in past data, not immutable laws. Results depend on the model, task, training regime, resources, and how improvement is measured.
Where might scaling run into limits?
High-quality data is not an infinite resource
A 2024 ICML position paper examines whether the supply of public, human-generated text could constrain language-model scaling. Its scope is that kind of text for LLM training. It does not establish that all useful data is exhausted: conclusions depend on assumptions about access, reuse, data quality, and alternatives. See “Position: Will we run out of data? Limits of LLM scaling based on human-generated data”.
Compute depends on more than chips
Large training runs require computing hardware, but also electricity, capital, data, and time. Samaritan Research’s August 20, 2024 analysis identifies power, chip manufacturing, data, and latency as constraints on scaling. It estimates that a training run of 2×1029 FLOP would likely be feasible by 2030 under its assumptions; that is an infrastructure scenario, not a claim that such a run will happen or produce a particular capability. Read the analysis and its assumptions.
Some scaling choices have diminishing returns
OpenAI’s discussion of training scaling identifies economic willingness to spend and how well work can be parallelized as factors affecting compute growth. It also describes rapidly diminishing algorithmic returns from batch sizes that are too large. The limits vary by task and are not fully understood. This is an example of one way of scaling encountering weaker returns, not evidence that all routes to improvement have ended.
What do forecasts say about the next few years?
Forecasts in this area are conditional scenarios, not promises about future capability. Their baselines and assumptions differ, so their figures should not be treated as directly comparable.
| Source and date | Scenario or estimate | What it does—and does not—show |
|---|---|---|
| International Scientific Report, interim report | If recent trends continue, by the end of 2026 some general-purpose models could use 40–100× the compute of the most compute-intensive models published in 2023, alongside methods that use compute 3–20× more efficiently. | A conditional projection about compute and efficiency, not an observed outcome or direct prediction of capability. |
| Samaritan Research, August 2024 | A 2×1029-FLOP training run would likely be feasible by 2030 under the analysis’s assumptions. | An estimate of infrastructure feasibility, not proof the run will occur or what it would achieve. |
| Epoch AI | Model-count projections include conservative and aggressive scenarios and discuss a possible capability plateau at some level of effective training compute. | A conditional possibility; the cited discussion does not claim that such a plateau has been measured or dated. |
The 2026 and 2030 scenarios come from the International Scientific Report and Samaritan Research, respectively; the model-count scenarios are discussed by Epoch AI. Compute, parameters, and capability are different measures, so none of these figures gives a date when AI will stop improving.
Does a benchmark plateau mean AI has stopped getting better?
No. A benchmark measures performance on a defined set of tasks under a defined evaluation method. Scores may stop rising because models are nearing the test’s ceiling, the test has been exposed during training, or its design and format fail to distinguish newer capabilities. A 2026 systematic study treats benchmark saturation as a measurement issue involving benchmark design, data construction, and evaluation format; saturation on a test is not the same as a demonstrated ceiling on AI capability. See “When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation”.
When you encounter a claim that models have plateaued, check what is actually being measured: training loss, a benchmark score, performance in one domain, cost-adjusted results, or real-world usefulness. Also ask whether comparisons keep the evaluation method and budget consistent, and whether the test is close to its ceiling. Without those details, a flat score is difficult to interpret as evidence about broader progress.
Could AI improve without simply scaling pretraining?
Yes. More efficient algorithms can make a given amount of compute go further; post-training can change how a model follows instructions or handles tasks; and inference-time methods can spend additional compute while answering. Better data and better evaluation can also change what models learn and what researchers can detect. These routes do not guarantee continued rapid progress, but they mean that a limit in one pretraining recipe would not, by itself, settle the broader question.
The central uncertainty is whether scaling up and refining existing techniques can sustain rapid gains, or whether substantial advances will require unpredictable breakthroughs. The International Scientific Report frames that as an open debate, including whether current approaches can address harder challenges such as common-sense reasoning and flexible world models. The evidence supports neither a guaranteed breakthrough nor a known stopping point.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
