When bigger stops being better
For years, the slogan “scale is all you need” captured the dominant mindset in AI. The recipe sounded simple: more data, more compute, larger models, and general intelligence would somehow emerge. That story is now visibly cracking. Even top conferences highlight how performance gains are slowing down while fundamental weaknesses remain in reasoning, robustness and real-world reliability.(patmcguinness.substack.com)
Empirical studies of scaling laws show clear diminishing returns. Doubling training costs often buys only marginal quality improvements, while standard benchmarks start to saturate. Behind the scenes, researchers and practitioners report a growing gap between impressive demo videos and everyday deployments that must meet strict constraints of cost, latency and safety.(arsturn.com)
Energy, infrastructure and concentrated power
Endless scaling also collides with physical reality. Frontier models demand enormous data centers, specialized chips and electricity on the scale of entire cities. Meta-analyses of training trends document huge increases in computational cost from early deep learning systems to modern large language models, raising deep concerns about long term technical and economic sustainability.(ResearchGate)
This infrastructure push is not neutral. A handful of tech giants now pour unprecedented capital into AI clusters, while lobbying for favorable regulation, access to energy and public subsidies. Reports show how this “too big to fail” build out locks the field into a path where control over models, data and hardware is concentrated in very few hands. The environmental footprint and the political power of these infrastructures are social questions as much as technical ones.(AI Now Institute)
Human needs, not just benchmarks
At the same time, the business case for brute force scaling looks weaker than the hype suggests. Studies with real companies find that many struggle to obtain solid returns from generic generative AI tools. Hallucinations, compliance risks, integration costs and the need for human oversight often erode promised productivity gains. What matters in practice is not abstract benchmark scores, but whether systems actually solve concrete problems in reliable, accountable ways.(arsturn.com)
A human centered perspective asks different questions. Which communities benefit from AI systems and which are left paying the energy and environmental bill. Whether workers gain meaningful support and new capabilities or face opaque automation. Whether scientific understanding advances, or whether we simply stack larger black boxes on top of fragile assumptions.
Alternative directions: efficiency, structure, openness
Rethinking AI after the end of naive scaling means working with new design goals. Instead of maximizing parameter counts, we can aim for maximum capability per joule, per token and per minute of human attention. That points towards smaller, domain specific models trained on high quality, well documented and often open data, where scientific transparency is possible.
Architecturally, hybrid approaches are regaining momentum. Neuro symbolic systems, causal models and world modeling architectures try to encode structure, not only statistics. Biomimetic work revisits insights from biological brains, exploring mechanisms like sparse activation, continual learning and robust perception in changing environments. These lines of research promise both energy efficiency and more stable behavior under real world conditions.(ResearchGate)
On the socio technical side, open models and public compute can rebalance power. Public research institutions, civic infrastructures and cooperatives could host efficient models tuned to local languages and needs, under democratic oversight. Evaluation practices can shift from leaderboard worship to multi dimensional assessments that include social benefit, environmental impact and alignment with scientific norms.(AI Now Institute)
Towards an AI that serves the many
Declaring that “scale is all you need” was always an ideological move, not a scientific law. It suggested that enough data and compute would magically solve problems of understanding, justice and sustainability. The emerging evidence from scaling slowdowns, energy constraints and disappointing real world deployments makes clear that this vision has run its course.
The alternative is not to abandon AI, but to redirect it. Towards systems that are smaller but smarter in how they use resources. Towards designs that learn from human and animal cognition without trying to clone them. Towards infrastructures that are open, accountable and oriented to the public good. An AI field built on those principles will be less spectacular in parameter counts, but far more promising for people, the planet and the future of science.
—
Sources for this article:
1. Gary Marcus – “Breaking News: Scale Is All You Need Is Dead” – garymarcus.substack.com -It is one of the most systematic and influential critiques of the scaling doctrine, capturing the overall climate at NeurIPS 2025, the failure of scaling laws as a one-dimensional strategy, and the need for cognitively inspired models.
2. LogicBombs – “The End of Scaling and What’s Next” – logicbombs.substack.com – It offers an in-depth critique of scaling laws, presents the positions of Mitchell, Sutton and Choi at NeurIPS, and discusses emerging directions such as world models, causality and meta-learning.
3. TechCrunch – “AI scaling laws are showing diminishing returns”- techcrunch.com -It captures the techno-economic dimension of the scaling crisis: slowing improvements, massive energy costs, lack of business ROI and strategic shifts among major AI companies.
4. Epoch AI – “Can AI Scaling Continue Through 2030?”- epoch.ai – It provides quantitative projections on where scaling halts due to limits in data, energy and hardware, using technical models such as compute forecasts and spectral constraints.
5. NeurIPS 2025 Conference – Invited Talks & Papers – neurips.cc -This source documents the paradigm shift through official talks by leading scientists (Mitchell, Choi, Sutton) highlighting:
• the cognitive limits of LLMs
• the energy and practical constraints of scaling
• the need for new architectures such as neuro-symbolic systems and model-based reinforcement learning.