How "more data means smarter AI" became the conventional wisdomFor the past several years, AI development has rested on one simple belief: that increasing data, increasing compute, and making models bigger will make them proportionally smarter. This belief had real grounding. In 2020, a paper on "scaling laws" by Jared Kaplan and colleagues at OpenAI showed that performance improves predictably as you increase three things: model size, the amount of training data, and the amount of compute. That paper opened the era of scaling, in which the pursuit of sheer size was expected to drive performance.In practice, the 2019 GPT-2 had 1.5 billion parameters, while its successor GPT-3 ballooned to more than 100 times that, with a dramatic jump in capability. And so the whole industry started sprinting in one direction: bigger, and more. Data centers the size of small cities were built, and the sums being poured in reached the scale of hundreds of billions of dollars (tens of trillions of yen). To build smart AI, everyone believed, was to pile up an enormous mountain of data. No one really questioned it.The surprising result that Johns Hopkins put on the tableThen, across 2025 and 2026, a string of studies appeared that shook this conventional wisdom to its foundations. The most emblematic was research led by Assistant Professor Mick Bonner and colleagues at Johns Hopkins University, published in the journal Nature Machine Intelligence.The team took the three types of network architecture most widely used in modern AI, namely transformers, fully connected networks, and convolutional neural networks, and adjusted their designs little by little to build dozens of different artificial neural networks. The crucial point is that none of these models were trained at all. They showed the untrained models images of people, animals, and objects, and compared the internal activity of the networks with the brain responses of humans and non-human primates viewing the same images.The result defied expectations. Increasing the number of artificial neurons in transformers and fully connected networks produced almost no meaningful change, but the same adjustment in convolutional networks produced activity patterns that more closely resembled the human brain. What is more, these untrained convolutional models performed on par with conventional AI systems that normally require exposure to millions or even billions of images.Bonner's words are striking. The field right now, he notes, is moving toward throwing huge amounts of data at models and building compute resources the size of small cities, while humans learn to see using very little data. Perhaps the brain design that evolution arrived at over a long stretch of time exists for good reason → and if so, starting from a brain-like design might place AI at an enormously advantageous starting point. That is the heart of this research.Why massive data can become wasted effortSo why does a vast pile of data stop being meaningful? There are several reasons.One is the phenomenon of diminishing returns. Scaling laws always meant that gains in performance would flatten out in the shape of a power curve. On a logarithmic graph it looks like a straight line, which feels like steady, healthy progress, but converted to an ordinary graph, the same curve shows large early gains that quickly level off. In 2021, doubling compute roughly doubled performance, but by 2025, doubling pre-training compute bought only a 10 to 20 percent improvement on the hardest reasoning tasks.Another is the "data wall." Training a one-trillion-parameter model optimally would require roughly 20 trillion tokens of data, yet high-quality text on the internet is estimated, depending on how you count, at only somewhere between 10 and 50 trillion tokens. In other words, the day when we exhaust the writing humanity has left behind is becoming a real prospect.More serious still is the problem known as "model collapse." There is a push to fill the gap left by dwindling data with synthetic data generated by AI itself, but here lies a trap. One study showed that even a tiny fraction of synthetic data, as little as 0.1 percent, can halt the progress of learning. It has also been found that larger models tend to overfit to imperceptible artifacts inside synthetic data, making them more, not less, vulnerable to harm. Continuing to add thin, low-substance data in pursuit of volume can drag performance down rather than push it up.The positive side: where data still worksThat said, it would be hasty to declare that data is meaningless. The accurate understanding is that what matters is not "quantity" but "quality" and "how it is used."A telling example is Microsoft's "Phi" series. The Phi-1 model, with just 1.3 billion parameters, was trained on carefully prepared, textbook-quality synthetic data and outperformed models 100 times its size on a coding benchmark. The team pursued improvement along a different axis, quality rather than quantity of data, and achieved common-sense reasoning results comparable to models ten times the size, using less than a tenth of the data. The idea that "textbooks are all you need" gave the industry a fresh jolt.But there is an easily overlooked piece of backstory here. The Phi team later stopped relying on purely synthetic data, and went back to the now-familiar method of mixing high-quality real data with synthetic data. Textbook-style synthetic data alone made the vocabulary that came out thin in diversity, and in some cases dragged performance down. The lesson researchers arrived at was to mix rather than replace, keeping genuine human-written data and appending synthetic data on top. Data is still necessary, but rather than piling it up carelessly, the design of what to mix and how decides success or failure.The behind-the-scenes story: what researchers had quietly realizedWhile the public rallying cry of "bigger, bigger" continued, the mood inside the labs had been shifting for a long time.The view that frontier models may have hit a ceiling had been quietly spreading within the industry for over a year. This was something not easy to bring out into the open. Having attracted enormous investment, it is hard to openly admit a slowdown. Even so, influential voices spoke up one after another. Ilya Sutskever, a co-founder of OpenAI, said at NeurIPS in 2024 that pre-training as we know it will end, and that the 2010s were the age of scaling, while now we have returned to an age of discovery.Researcher Sara Hooker, in a 2026 essay, documented how smaller models are rapidly closing the gap with larger ones through better training techniques. In fact, Falcon 180B, which appeared in 2023 with 180 billion parameters, was outperformed by the 8-billion-parameter Llama 3 8B that arrived just one year later. This was a vivid example showing that the era in which scale alone did the talking is already drawing to a close.The decisive change was that the protagonist of performance gains switched. Reasoning models, exemplified by OpenAI's o1 and o3, focused not on making the model bigger but on a different axis, namely how long they can think before producing an answer. DeepSeek R1's breakthrough, too, lay not in its size but in the ingenuity of its post-training recipe. Around the same time, an MIT team pointed out that the largest, most compute-intensive models may eventually face diminishing returns relative to smaller ones, warning that things are very likely to start narrowing over the next five to ten years. In short, the source of intelligence is shifting from "how much you fed it" to "with what structure, and how you make it think."What changes from herePulling these threads together, one map comes into view. The strategy of pouring in data without limit is less a mistake in itself than a strategy whose era is ending, having served its role.What the Johns Hopkins research showed was a hopeful future. If training on massive data were really the decisive factor, Bonner says, there would be no way to reach brain-like AI through architectural changes alone. The fact that it was possible means that by starting from the right blueprint and incorporating insights from biology, we may be able to dramatically accelerate learning in AI. The team is now exploring simple learning methods modeled on biological mechanisms, aiming for a new generation of frameworks that are faster, more efficient, and less dependent on massive data.This shift also matches our own intuition. A human child can learn to recognize a dog after seeing one a few times, without looking at billions of photos. The "good initial design" that evolution prepared turns a little experience into meaning all at once. AI, too, is quietly moving its center of gravity, from the stage of brute-forcing its way through with sheer volume of data, to the stage of raising efficiency through smarter structure and learning methods.Feeding in massive amounts of data has not become meaningless. But the naive expectation that this alone will make a system smart no longer holds. The question from here will shift from "how much do we give" to "into what vessel, what do we give, and how." Beyond the race to pour in money and energy without limit, a new era of competing on design wisdom has already begun.ReferencesJohns Hopkins University, "AI may not need massive training data after all," ScienceDaily, January 4, 2026Kazemian, A., Elmoznino, E., Bonner, M. F., "Convolutional architectures are cortex-aligned de novo," Nature Machine Intelligence, Vol. 7, 2025Jeong, J., "The One-Axis Era Is Over: AI's Four New Scaling Laws," Medium, May 2026"LLM Scaling Laws Explained: Will Bigger AI Models Always Win?," BuildFastWithAI, March 2026MIT FutureTech, MIT News, "Initiative on the Digital Economy," April 2026"Demystifying Synthetic Data in LLM Pre-training," arXiv:2510.01631, 2025Langlais, P-C., "Synthetic Pretraining," Vintage Data, February 2026"Context Collapse: In-Context Learning and Model Collapse," arXiv:2601.00923, 2026