The Future of Large Language Models: From Scaling to Trustworthy Intelligence
A breakthrough, not a final destination
Large language models have already changed how people write, code, search, translate, summarize, teach and organize knowledge. Their success rests on a powerful empirical insight: when models, data and compute grow together, new capabilities appear. This is the core intuition behind the scaling hypothesis, and it explains much of the progress of the last decade.
But the scientific debate has moved beyond the simple question of whether LLMs work. They do. The more important question is whether bigger language models, trained on more data with more compute, are sufficient for reliable, general and socially useful artificial intelligence. The strongest answer from current research is nuanced. Scaling remains important, but it is not enough as a complete theory of intelligence or as a public strategy for AI.
The future of LLMs will not be defined only by size. It will be defined by data quality, reasoning reliability, grounding, architecture, governance and the ability of societies to control the infrastructures on which they depend.
The data bottleneck
The first constraint is data. Modern LLMs have already consumed enormous amounts of publicly available digital text. Future progress cannot simply rely on “more internet”, because high-quality human-generated text is finite. This changes the economics and politics of AI. Companies increasingly seek private datasets, scientific corpora, user interactions, code repositories, legal materials, books, audiovisual data and licensed archives.
A second path is synthetic data, generated by models themselves. Synthetic data can be useful, especially for controlled tasks, rare cases, simulation and instruction tuning. But it also carries a serious risk. If future models are repeatedly trained on outputs produced by earlier models, they may gradually lose diversity, distort rare patterns and drift away from the original human distribution. The result can be more fluent but less grounded systems.
This risk matters especially for smaller languages, public administrations and scientific communities outside the dominant English-speaking data economy. If Greek, Finnish, Czech, Arabic or Swahili are represented mainly through weak, filtered or synthetic traces, future AI systems will reproduce linguistic inequality. A democratic AI strategy therefore requires high-quality public datasets, open corpora, domain-specific knowledge bases and transparent data governance.
Fluency is not understanding
The most important conceptual criticism of LLMs is that language fluency should not be confused with understanding. LLMs are extraordinary pattern-learning systems. They can produce coherent prose, summarize complex material, write code, simulate dialogue and support research. But they do not understand in the human sense. They do not have intentions, responsibility, lived experience or an intrinsic grasp of truth.
This is why they can produce confident errors. Hallucination is not merely a bug that can be patched away by making models larger. It follows from the fact that the system is optimized to generate plausible continuations, not to guarantee truth. Better training, retrieval, evaluation and tool use can reduce the problem, but they do not eliminate the need for institutional design.
The practical conclusion is clear. LLMs should be treated as powerful assistants, not autonomous authorities. In public administration, health, justice, education and journalism, the final responsibility must remain human. Every critical system should include source-grounded retrieval, audit logs, error reporting, human review and periodic reassessment. The right question is not whether an LLM can answer. It is whether its answer is verifiable, contestable and safe to use in context.
World models and embodied intelligence
A second major research direction argues that intelligence requires more than language prediction. Human beings do not learn only by reading text. They learn through perception, movement, manipulation, social interaction, failure and correction. They build models of the world: objects persist, actions have consequences, plans can fail, bodies are constrained by physics, and other agents have goals.
This is why many researchers argue that future AI systems will need stronger world models. Such systems should not merely predict the next token. They should learn structured representations of the world from video, sensors, interaction, simulation and action. This is especially relevant for robotics, scientific discovery, planning, engineering, autonomous systems and any domain where the model must anticipate the consequences of interventions.
This does not mean that LLMs will disappear. It means they will become part of larger architectures. Language is a powerful interface, but it is not the whole of intelligence.
The hybrid future
The most plausible future is hybrid. LLMs will be combined with retrieval systems, symbolic reasoning, formal verification, calculators, code execution, knowledge graphs, simulation environments, smaller specialized models, multimodal perception and human oversight. This is already visible in practical systems: retrieval-augmented generation for legal and administrative knowledge, tool-using agents for software engineering, verification loops for mathematics and code, and domain models for medicine, climate, energy and public policy.
This hybrid path is also the most relevant for Europe. The strategic issue is not only who builds the largest model. It is who controls the stack: data, models, evaluation, deployment, APIs, hardware, cloud infrastructure, governance and public procurement. If governments and universities rely only on closed foreign systems, they may gain short-term convenience but lose long-term capacity.
Open models, open standards, open datasets and public AI infrastructures are therefore not ideological luxuries. They are prerequisites for scientific reproducibility, democratic accountability and digital sovereignty. Public money should build public knowledge: reusable datasets, transparent benchmarks, open-source tools, auditable models and local expertise.
The real future of LLMs
LLMs are neither a passing trend nor a guaranteed road to artificial general intelligence. They are a major general-purpose technology with enormous value and real limits. The most scientifically grounded position is to take both seriously.
Scaling will continue to produce improvements, especially when combined with better data, longer training, better inference strategies and specialized architectures. But reliable intelligence will require grounding, memory, reasoning, verification, embodiment in some domains, symbolic structure in others and accountable human institutions around all of them.
The next phase of AI will belong not simply to those who build the largest models, but to those who build the most trustworthy, efficient, open and socially governed systems. For small countries, public administrations and research communities, this is the central lesson: do not become passive consumers of black-box intelligence. Build capacity, build commons, build open infrastructures, and use LLMs as instruments of knowledge rather than substitutes for judgment.
Article sources:
DeepMind, “Training Compute-Optimal Large Language Models”: The Chinchilla paper shows that LLM performance depends not only on parameter count but on the balance between compute, model size and training tokens, demonstrating that smaller, better-trained models can outperform much larger undertrained ones: https://arxiv.org/abs/2203.15556,
Epoch AI, “Will we run out of data? Limits of LLM scaling based on human-generated data”: This study documents the likely constraint posed by finite high-quality human-generated text and explains why continued scaling cannot rely indefinitely on public web text alone: https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data,
Nature, “AI models collapse when trained on recursively generated data”: This peer-reviewed paper documents model collapse, the degenerative process that can occur when new generative models are repeatedly trained on data produced by earlier models: https://www.nature.com/articles/s41586-024-07566-y,
Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, Margaret Mitchell, “On the Dangers of Stochastic Parrots”: This foundational paper argues that large language models can produce fluent language without guaranteed understanding, while raising risks concerning bias, opacity, environmental cost and concentration of power: https://dl.acm.org/doi/10.1145/3442188.3445922,
Yann LeCun, “A Path Towards Autonomous Machine Intelligence”: LeCun’s position paper argues that future AI needs architectures capable of learning world models and planning beyond next-token prediction, laying the groundwork for objective-driven AI and JEPA-style approaches: https://openreview.net/forum?id=BZ5a1r-kVsf,
Artur d’Avila Garcez and Luis C. Lamb, “Neurosymbolic AI: The 3rd Wave”: This paper explains why future AI systems may need to combine neural learning with symbolic knowledge representation, logical reasoning, explainability and accountability: https://arxiv.org/abs/2012.05876,
François Chollet et al., “ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems”: This benchmark paper shows that frontier AI systems still struggle with abstract reasoning tasks that are accessible to humans, supporting the view that linguistic fluency is not equivalent to general intelligence: https://arxiv.org/abs/2505.11831,
Rich Sutton, “The Bitter Lesson”: Sutton’s essay explains why general methods that scale with computation have historically won in AI, while also helping frame the current debate on whether scaling must be complemented by better architectures, grounding and real-world experience: https://www.incompleteideas.net/IncIdeas/BitterLesson.html.