Is scaling enough to reach human-level AI? The case for bigger models, the case for new ideas, and what the evidence shows

Every few months a new AI model arrives that is better than the last, and the same question follows it. Did it get better mainly because it was bigger, trained with more computing power on more data? Or because someone had a genuinely new idea? The answer matters well beyond research labs. If size is what counts, progress depends mostly on money, chips and electricity, and the companies with the most of each will keep pulling ahead. If new ideas count most, progress is harder to predict, and a small team with the right insight could change the picture.

This article sets out the strongest version of each side, then looks at what the evidence actually shows. You will come away knowing what "scaling" means, why the argument is less of an either-or than it sounds, and which signals will tell you which side is gaining ground.

What the two sides mean

Scaling means making the same basic recipe bigger: more parameters (the adjustable numbers inside a model), more training data, and more compute (the amount of calculation used to train or run it). The idea gained its scientific footing in 2020, when researchers at OpenAI showed that a language model's error rate, which they called its loss, "scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude"². A power law is a smooth, predictable curve. In plain terms: put in ten times more, and you can forecast roughly how much better the model will get. These smooth curves are called scaling laws.

New ideas means changes to the recipe itself: a different model design (an architecture), a different way of learning, or a different kind of training signal. The people in this camp do not deny that scale helps. Their claim is that today's recipe has limits that more of the same will not remove.

The argument is old. In March 2019, the computer scientist Rich Sutton wrote a short essay called "The Bitter Lesson." It begins: "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin"¹. The "bitter" part is that clever, hand-built knowledge has repeatedly lost to simpler methods given more computing power. Three years later, the cognitive scientist Gary Marcus argued the opposite, that "hybrid AI, not deep learning alone (nor symbols alone) seems the best way forward," and that "there are serious holes in the scaling argument"⁸. Both positions have only sharpened since.

The case that scaling is enough

Put at its strongest, the scaling case says that the curves have held for years, that the ingredients keep growing, and that each predicted wall has so far failed to appear.

The inputs keep growing fast. Epoch AI, an independent research group that tracks AI trends, found that the compute used to train notable models grew about 4.1 times a year from 2010 to May 2024 (90% confidence interval 3.7 to 4.6), and concluded that recent growth is "best described as increasing by a factor of 4-5x/year"⁴.

Scale explains most of the gains. A 2024 study by Epoch researchers estimated how much of the progress in language models came from bigger training runs and how much from better methods. Better methods mattered a lot: the compute needed to reach a given level of performance "halved approximately every 8 months, with a 95% confidence interval of around 5 to 14 months," faster than chip improvements under Moore's law. Yet "the increase in compute made an even larger contribution"⁵. Epoch summarises this as about two-thirds of the improvement over the last decade coming from scale⁴. The study rests on two older benchmarks and data from 2012 to 2023, which its authors call "noisy," so the split is an estimate, not a precise measurement.

Even the improvements made scaling work better. In 2022, DeepMind's Chinchilla study found that earlier models had been too big for the amount of data they saw. Model size and training data "should be scaled equally." A 70-billion-parameter model trained on four times more data, with the same compute budget as the 280-billion-parameter Gopher, outperformed Gopher and several larger models³. That was a new idea, but its effect was to show how to scale more efficiently.

The output curves are still climbing. METR, a nonprofit that tests AI systems, measures the length of software task, in skilled-human working time, that a model can finish half the time. In its January 2026 update, that length doubled about every seven months from 2019 to 2025, and faster for recent models: every 131 days for models since 2023 and every 89 days since 2024¹⁶.

The scaling camp's challenge to its critics is simple: name the wall, and say when it arrives.

The case that new ideas are needed

Put at its strongest, the new-ideas case says that the models are impressive but learn in a fundamentally inefficient way, that the cheapest fuel is running low, and that many of the people who built the current wave now agree.

Models generalise poorly. To generalise is to handle situations unlike the ones seen in training. Ilya Sutskever, a co-founder of OpenAI who now runs Safe Superintelligence, described 2020 to 2025 as "the age of scaling" and said the field is going "back to the age of research again, just with big computers," because models "generalize dramatically worse than people"¹².

Some tests remain stubbornly hard. The ARC-AGI benchmarks, designed by the researcher François Chollet, use small visual puzzles that most people can solve but that require learning a new rule from a few examples. When the second version launched in March 2025, "pure LLMs score 0%" and public reasoning systems achieved "only single-digit percentage scores," while every task had been solved by at least two people in a human test panel, whose average score was 60%¹⁰. Chollet's own verdict, written in December 2024: "Simply scaling up the things we were doing from 2019 to 2023 -- take the same architecture, train a bigger version on more data -- is not enough. Further progress is about new ideas"⁹.

The fuel is finite. A peer-reviewed estimate projected that models would be trained on datasets as large as the whole stock of public human-written text sometime between 2026 and 2032, with a median of 2028¹⁴. Epoch AI's analysis of the physical limits found that training runs of about 2×10²⁹ operations will likely be feasible by 2030, but that power is the constraint "likely to bind first," followed by chip manufacturing¹⁵. Neither stops progress, but both raise the price of each further step of brute-force scaling.

Researchers are sceptical, and serious money agrees. In a 2025 survey by AAAI, one of the field's oldest scientific societies, 76% of 475 respondents said that "scaling up current AI approaches" to yield AGI (artificial general intelligence, broadly human-level ability across tasks) is "unlikely" or "very unlikely" to succeed¹¹. In March 2026, AMI Labs, a Paris-based start-up co-founded by the Turing Award winner Yann LeCun after he left Meta, raised US$1.03 billion to build world models, AI that learns from the physical world rather than mainly from text, based on an architecture LeCun proposed in 2022. Its chief executive was candid that "it starts with fundamental research," and that turning it into products could take years¹³.

What the evidence shows so far

The most useful finding is that the two camps are describing the same events in different words.

Take the biggest change of the past two years: reasoning models, which are trained to work through a problem step by step before answering. That was a new idea. DeepSeek's R1 paper, later published in Nature, showed that "the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories"⁷. Reinforcement learning means learning by trial and reward rather than by copying examples. But reasoning models also opened a new thing to scale: test-time compute, the amount of calculation a model spends while answering. A 2024 study found that, in a test where total compute was held equal, a smaller model allowed to think longer could outperform a model 14 times larger, on problems where the smaller model already had some success⁶.

ARC-AGI tells the same two-sided story. In December 2024, a preview of OpenAI's o3 reasoning model scored 75.7% on the first ARC-AGI test within the prize's compute limit, and 87.5% when given 172 times more compute. Chollet called it "a genuine breakthrough" and also said, "I don't think o3 is AGI yet"⁹. The model had been trained on part of the test's public training set, and its successor test then reset scores close to zero⁹¹⁰. A scaling advocate reads this as more compute winning again. A new-ideas advocate reads it as a new method that needed enormous compute and still fails at the next test.

Both readings fit the data, which is the point. Most progress so far has come from a mix: scale doing most of the work, and new ideas, roughly every year or two, changing what is worth scaling.

Evidence, opinion and speculation

Evidence (measured): scaling curves have held across many orders of magnitude²³. Training compute has grown four to five times a year⁴. Algorithms have roughly halved compute needs every eight months, with scale contributing more overall⁵. Test-time compute can stand in for model size on some problems⁶. Some reasoning tests still defeat current systems¹⁰.

Informed opinion (judgment from experience): Sutton's lesson¹, Marcus's case for hybrids⁸, Chollet's call for new ideas⁹, Sutskever's "age of research"¹² and the AAAI survey¹¹. These are valuable, and the people giving them have strong track records, but they are views, not measurements. Note also that the AAAI respondents were mostly academics, a group with fewer of the largest training runs to look at than people inside frontier labs.

Speculation (no direct evidence yet): whether today's methods hit a ceiling before human-level ability, and whether a different architecture such as world models would do better at the same cost. No one has yet shown either at scale. The most decisive test is not a debate but a result: a new approach that beats scaled-up models on hard tasks for less compute, or scaled models that close the gap on tests like ARC-AGI-2 without a new method.

What this means for you

1. Treat "it is just scale" and "scaling is dead" as slogans. When you hear either, ask which kind of scale is meant: model size, data, training compute or thinking time. The answer usually turns a slogan into a checkable claim.
2. Expect steady gains, with occasional jumps. The trend lines suggest current tools will keep improving through the next product cycle. Plan for that, but leave room for sudden shifts when a new method lands.
3. Watch three signals. Scores on deliberately hard reasoning tests such as ARC-AGI-2. Whether frontier labs keep growing training runs at four to five times a year or shift spending to thinking time. And whether research-first start-ups like AMI Labs publish results that beat larger models.
4. Use thinking time when quality matters. If your AI tool offers a slower "reasoning" or "extended thinking" mode, test it on your hardest tasks. The evidence says extra thinking can matter as much as a bigger model on some problems.
5. Do not bet your plans on one camp. Choices that pay off either way, such as clean data, clear processes and staff who know how to test AI output, are the safest investment.

Sources

1. Sutton, R. (2019, March 13). The bitter lesson. incompleteideas.net/IncIdeas/BitterLesson.html
2. Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020). Scaling laws for neural language models. arXiv:2001.08361. arxiv.org/abs/2001.08361
3. Hoffmann, J., Borgeaud, S., Mensch, A., et al. (2022). Training compute-optimal large language models. arXiv:2203.15556. arxiv.org/abs/2203.15556
4. Sevilla, J., & Roldán, E. (2024, May 28). Training compute of frontier AI models grows by 4-5x per year. Epoch AI. epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year
5. Ho, A., Besiroglu, T., Erdil, E., Owen, D., Rahman, R., Guo, Z. C., Atkinson, D., Thompson, N., & Sevilla, J. (2024). Algorithmic progress in language models. arXiv:2403.05812. arxiv.org/abs/2403.05812
6. Snell, C., Lee, J., Xu, K., & Kumar, A. (2024). Scaling LLM test-time compute optimally can be more effective than scaling model parameters. arXiv:2408.03314. arxiv.org/abs/2408.03314
7. DeepSeek-AI (Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., et al.). (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv:2501.12948, published in Nature (2025). arxiv.org/abs/2501.12948
8. Marcus, G. (2022, March 10). Deep learning is hitting a wall. Nautilus. nautil.us/deep-learning-is-hitting-a-wall-238440
9. Chollet, F. (2024, December 20, updated April 16, 2025). OpenAI o3 breakthrough high score on ARC-AGI-Pub. ARC Prize. arcprize.org/blog/oai-o3-pub-breakthrough
10. Kamradt, G. (2025, March 24). ARC-AGI-2 + ARC Prize 2025 is live! ARC Prize. arcprize.org/blog/announcing-arc-agi-2-and-arc-prize-2025
11. AAAI. (2025, March). AAAI 2025 Presidential Panel on the Future of AI Research (community survey, 475 respondents, p. 63). aaai.org/about-aaai/presidential-panel-on-the-future-of-ai-research
12. Patel, D. (2025, November 25). Ilya Sutskever: We're moving from the age of scaling to the age of research [Interview transcript]. Dwarkesh Podcast. dwarkesh.com/p/ilya-sutskever-2
13. Heim, A. (2026, March 9). Yann LeCun's AMI Labs raises $1.03B to build world models. TechCrunch. techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models
14. Villalobos, P., et al. (2024). Will we run out of data? Limits of LLM scaling based on human-generated data. Proceedings of the 41st International Conference on Machine Learning (ICML 2024). arXiv:2211.04325. arxiv.org/abs/2211.04325
15. Sevilla, J., et al. (2024, August 20). Can AI scaling continue through 2030? Epoch AI. epoch.ai/blog/can-ai-scaling-continue-through-2030
16. METR. (2026, January 29). Time Horizon 1.1. metr.org/blog/2026-1-29-time-horizon-1-1