The New Law of Diminishing Returns

For years, the playbook in artificial intelligence was simple: make the model bigger, feed it more data, and buy more GPUs. Every major advancement in the field seemed to validate this approach. The tech sector poured billions into scaling large language models, assuming that model performance would continue to rise if we just kept feeding them compute and internet-scale text. That assumption is now breaking, and the data proves it. To fine tune LLM systems effectively, quality data has to be the foundation.

In 2025, researchers studying advanced reasoning systems found that adding more computational steps no longer delivered proportionate improvements. [1] This means the returns from scaling AI models are slowing, while the costs to run them are accelerating across the board. The same pattern that once propelled the field forward is now working against it. The industry’s obsession with scale has created models that are large enough to understand everything but too inefficient to sustain anything.

The integrity of datasets for machine learning dictates every downstream outcome. This is the moment to stop asking “How big is the model?” and start asking “How clean is the data?” and whether the data quality supports sustainable model performance.

The Scaling Illusion

Researchers analyzed the performance of dynamic reasoning systems, large models that extend their reasoning by looping over their own outputs. In the tests, each extra round of reasoning improved accuracy by a few percentage points, but every gain came at an enormous cost. 

This pattern repeats across every benchmark. Adding more context or more examples helps at first, but the benefits taper off quickly. Once models reach a certain level of reasoning complexity, they begin to repeat themselves.

In the study, large language agents operating in “reflexion” mode, reasoning step by step before deciding on an output, showed steep inefficiencies. The biggest models consumed up to 1 gigawatt of power when scaled to search-level traffic, roughly the energy demand of a mid-sized city. Even the lightest versions required power equivalent to Seattle’s daily usage. The results are clear: the performance curve has stopped climbing, but the cost curve keeps rising. If one company ran a model like that for a billion queries a day, it would spend millions in energy and hardware upkeep every week. The latency would cripple usability, and the environmental impact would be staggering.

The same dynamic holds true for model size. Small models trained with precise, structured data often rival the accuracy of models ten times their size. Research has shown that a small model like Llama-3 8B using optimized reasoning methods achieved almost the same accuracy as a 70-billion-parameter model, while only consuming a fraction of the power. The smaller system ran on a single GPU, while the larger one needed eight. Beyond that, Llama-3 8B has shown improvements of 4% in accuracy when giving it more time to reason, with a latency increase from 16 seconds to 25; yet to achieve another four percent improvement, it must run for over five minutes, a 30x increase in electricity for the same gain. [1]

Interested in learning more? Subscribe to the Sapien Newsletter and get a weekly brief on the future of human intelligence in AI.

The Data Problem Beneath It All

At the root of every bottleneck lies the same problem: data quality. All AI models depend on it, yet the supply of fresh, diverse, high-quality human data is finite. The internet is saturated with repetition, duplication, and synthetic noise, meaning that large models trained on this data inherit its flaws, leading to biased outputs and hallucinations.

The gap between “knows the words” and “understands the stakes” doesn’t shrink with raw size. A frontier model can still err catastrophically without human in the loop AI oversight and tell a hospital admin to violate policy, or generate financial advice that would get a junior analyst fired. The industry’s reliance on internet-scale scraping has compounded the issue. Many public datasets contain inconsistent or outdated information, and the filtering process cannot always distinguish between credible sources and noise. As the model’s appetite for data outpaces the world’s capacity to produce new human knowledge, it begins to consume its own output.

This feedback phenomenon, sometimes referred to as “model collapse,” occurs when AI systems are retrained on data generated by other models. Each new cycle dilutes the signal of human input. The model begins to imitate itself, losing the diversity and nuance that once made its predictions accurate. Researchers warn that this process accelerates as the volume of synthetic data increases. In large systems, the decay is exponential.

The Efficiency Frontier

The same study revealed that thoughtful data strategy can outperform hardware scaling. Smaller models, when trained and guided effectively with superior data quality, achieved performance close to their larger counterparts at a fraction of the energy cost. The alternative is smaller, curated datasets built on verified human insight. When trained on data produced by domain experts, such datasets for machine learning produce meaningful reasoning depth. They understand not just what an answer looks like, but why it comes to that conclusion. [2]

The focus is shifting from “Can it answer?” to “Can it answer correctly, consistently, and affordably?” Companies deploying AI at scale are discovering that smaller models trained on specialized data outperform generic multitask systems for real-world use. In healthcare, legal analysis, and robotics, accuracy in narrow domains now matters more than broad linguistic fluency.

The age of “just make it bigger” is ending. Scaling up AI models has delivered historic breakthroughs, but the economics are turning. Every additional point of performance demands outsized investments in compute investments that few organizations can justify indefinitely.

The New Data Paradigm

A more durable strategy is becoming clear. High quality data sourced with consent, stress tested and reviewed by experts, is outperforming bulk data scraped from the open internet. Fine tuning large language models on this kind of trusted dataset produces systems that are both auditable and defensible. The same principles that built open financial systems and transparent supply chains are now being applied to the infrastructure of intelligence itself.

Sapien’s decentralized data foundry enables contributors worldwide to provide expert human input and stake their reputation on the quality of their work. Through making the human contributors a part of the value chain, this aligns incentives across the network towards quality work.

The winners in the next phase of AI won’t just be those with the largest clusters of GPUs. They’ll be the ones who can fine tune LLM using verifiable human data, and do so at scale.

Read Next:

What happens when a model collapses? - How Human Knowledge Keeps AI From Consuming Itself

How our token guarantees Proof of Quality - Sapien Tokenomics

When will we run out of data? - Exploring the Limits of Internet-Sourced Training Data

FAQ:

What does “diminishing returns” mean in AI?
Making models larger or feeding them more compute no longer improves accuracy in proportion to the resources used. Each new performance gain costs exponentially more energy, time, and money.

Why did AI scaling work before and why is it slowing now?
Early scaling delivered big leaps because models were moving from underfitting to understanding. Now, most of the low-hanging gains are gone. New data is repetitive, synthetic, and less informative, so models learn less from more.

How does data quality affect AI performance?
Poor data quality leads to bias, hallucinations, and errors. Clean, human-curated data improves reasoning depth, reliability, and interpretability, especially in domains like healthcare, law, and robotics.

What’s the alternative to scaling?
Smaller, domain-specific systems trained with expert oversight. When data is sourced from qualified humans and validated onchain, models deliver consistent accuracy without waste.

How does Sapien’s protocol address this?
Sapien’s decentralized data foundry transforms human expertise into verifiable, high-quality AI training data. Contributors stake tokens, validate peers, and earn based on performance, ensuring accuracy through aligned incentives.

How can I start with Sapien?
Schedule a consultation to audit your LLM data training set.

Sources:
‍[1] Jiin Kim and Byeongjun Shin and Jinha Chung and Minsoo Rhu, The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective, 2025 - https://arxiv.org/abs/2506.04301
‍[2] Yanxin Liu and Yunqi Zhang, Relative Overfitting and Accept-Reject Framework, 2025 - https://arxiv.org/abs/2505.07783

‍