Large language models (LLMs) such as ChatGPT, Claude, and Gemini dominate AI headlines, but building AGI beyond language models is where the most serious research energy is now flowing. Other approaches are progressing steadily, but they get far fewer headlines.
That accessibility is not the same thing as comprehensiveness. LLMs are remarkably easy to interact with because they speak our language: you type, they respond. The interface feels natural because it mirrors the way humans already communicate. Conflating accessibility with comprehensiveness, however, is one of the most common misconceptions shaping public understanding of AI today.
Why LLMs Fall Short of AGI
LLMs are fundamentally statistical engines. They predict the next most likely token in a sequence, trained on enormous quantities of text. They are extraordinarily capable at tasks that map well onto that structure: writing, summarizing, translating, and reasoning through language-shaped problems.
Nevertheless, intelligence, in the broad and general sense we aspire to build, involves far more than language. It involves perceiving the physical world, learning from sparse experience, planning over long time horizons, forming causal models of reality, and adapting to entirely novel environments without access to a training corpus. As a result, the gap between today’s LLMs and artificial general intelligence (AGI), a system capable of performing any cognitive task a human can, remains substantial.
The Generalization Gap
By early 2026, that gap had become a defining topic of conversation among the very researchers who built the current generation of models. Ilya Sutskever, the co-founder of OpenAI, noted in a widely discussed interview that today’s models “still don’t generalize well,” and that there is a wide gulf between impressive benchmark evaluations and real-world reliability.
Reflecting on the arc of the field, Sutskever suggested that 2026 onward will be a new “age of research,” one defined by new architectures and learning paradigms rather than the sheer scaling of existing ones (Singularity University, 2026).
The End of Scaling
Yann LeCun, who won the Turing Award for helping invent deep learning itself, went further, declaring LLMs a dead end for AGI and founding a new company to pursue a fundamentally different approach. Both researchers have consequently put their reputations and capital behind bets that the next leap in intelligence will not come from making language models bigger.
With that context in place, it is worth examining the companies and researchers building AGI through fundamentally different paradigms, as well as those combining those paradigms with language models in productive ways.
Five Companies Building AGI Beyond Language Models
1. Safe Superintelligence Inc.: Generalization Over Scaling
Safe Superintelligence Inc. (SSI) was founded in June 2024 by Ilya Sutskever, alongside Daniel Gross and Daniel Levy, with a singular stated mission: to build superintelligent AI that is safe by design, not an afterthought. By 2026, the company had raised over $3 billion and reached a $32 billion valuation, extraordinary figures for a lab that has released no products (Singularity University, 2026).
The best SSI is making is philosophical as much as technical. Sutskever’s central argument is that the core limitation of current AI is not raw capability but generalization, meaning the ability to perform well outside the training distribution. A model that passes a medical licensing exam but then gives a patient dangerous health advice has not generalized; it has pattern-matched.
SSI’s approach involves researching new architectures and training methods centered on continual learning, in which a system continually improves through interaction with the world rather than a fixed pre-training pass. Sutskever has described the vision as a superintelligent 15-year-old who starts without domain expertise but learns any trade rapidly through lived experience, not an omniscient oracle, but a generalist learner (Singularity University, 2026).
2. Advanced Machine Intelligence: World Models and JEPA
After years of publicly arguing that LLMs are the wrong path to AGI, Yann LeCun left Meta in late 2025 and founded Advanced Machine Intelligence (AMI). The company is built around his theory of world models and the Joint Embedding Predictive Architecture (JEPA) he developed during his time at Meta.
Where LLMs predict the next word in a sequence, world models predict the next state of the world, building internal representations of physics, causality, space, and time from video and sensory data rather than from text alone. As LeCun put it, the goal is for AI that “learns the underlying rules of the world from observation, like a baby learning about gravity,” treating that as the foundation for common sense and real-world reasoning (Singularity University, 2026).
This approach targets the exact capabilities that LLMs struggle with most: understanding that a glass will fall if pushed, that actions have consequences, and that cause precedes effect. Furthermore, it addresses data efficiency, as children recognize objects with far fewer examples than deep learning models require, suggesting that the brain uses inductive learning principles that current architectures still lack (Potkalitsky, 2026).
3. Google DeepMind: Reinforcement Learning and World Modeling at Scale
Google DeepMind is arguably the most prolific producer of non-LLM AGI research worldwide, with a body of work spanning reinforcement learning, neuroscience-inspired architectures, and generative world models. Their AlphaZero system mastered chess and Go through self-play alone, starting from nothing but the rules and rapidly surpassing all prior world-champion programs.
AlphaFold used deep learning combined with evolutionary data to solve protein structure prediction at a level that earned its founders a Nobel Prize in Chemistry in 2024. More recently, Genie 3, their generative world model, creates realistic, interactive 3D environments on demand from simple prompts, with applications ranging from robotics training to game development.
The significance of Genie 3 for AGI research is that it represents a system learning the structure of physical reality from visual experience rather than from textual descriptions, which constitutes a fundamentally different form of world knowledge (Websolutioncentre, 2026). DeepMind’s overarching approach treats intelligence as emerging from structured interaction with environments, reflecting a consistent belief that language is one modality among many, not the foundation of general cognition.
4. World Labs: Spatial Intelligence as a Path to Generality
World Labs was founded in 2023 by Fei-Fei Li, the researcher best known for building ImageNet, the dataset that catalyzed the deep learning revolution. The company’s thesis is that spatial intelligence, meaning the ability to understand, model, and reason about three-dimensional environments, is one of the most critical missing pieces on the path to AGI.
In 2025, World Labs released Marble, their first world model capable of generating and manipulating 3D environments. The underlying argument is that purely text-based training gives an AI knowledge about the world as described in human language, which is inherently secondhand and incomplete.
Spatial reasoning from direct perceptual experience provides a complementary form of understanding more analogous to how biological intelligence actually develops (Websolutioncentre, 2026). Moreover, any AI system operating in the physical world, through robotics, autonomous vehicles, or augmented reality, requires spatial competence as a prerequisite, making World Labs’ work foundational to the broader AGI agenda.
5. Numenta: Intelligence Through Neuroscience
Numenta has spent over two decades building AI not by scaling neural networks but by reverse-engineering the neocortex. Their Hierarchical Temporal Memory (HTM) framework is derived from theoretical models of how the brain’s cortex forms predictions, stores sequences, and builds sparse, distributed representations of the world.
Rather than training on labeled datasets in discrete phases, HTM systems learn continuously and online, updating their internal representations as new data arrives, much as biological memory does. Numenta’s research has directly influenced thinking about continual learning, the property by which a system can acquire new knowledge without catastrophically overwriting what it previously learned.
This problem persists in conventional deep learning and is one of the clearest illustrations of why biological intelligence and current AI architectures operate on fundamentally different principles (Potkalitsky, 2026). Numenta’s path is slower and less visible than that of the frontier labs, but its conceptual contributions on sparse representations, sequence memory, and neocortical theory inform a growing body of neuroscience-inspired AI research that many believe will be essential for any system that truly learns the way humans do.
The Combination Approach: Building AGI Beyond Language Models
As it turns out, some of the most capable AI systems in 2025 and 2026 are not the product of a single paradigm but of two or more working in concert. This hybrid strategy is increasingly seen as a practical near-term path for building AGI beyond language models, bridging the strengths of language models with the capabilities they lack on their own.
DeepMind’s AlphaGeometry: Neural Plus Symbolic Reasoning
AlphaGeometry is one of the clearest demonstrations of what happens when LLMs are paired with formal symbolic reasoning. It solved International Mathematical Olympiad-level geometry problems by combining a neural language model that generated intuitive geometric constructions with a symbolic deduction engine that formally verified each logical step.
Neither component could solve the problems on its own. The language model provided creative leaps that pure symbolic search would miss; the symbolic engine provided verifiable, step-by-step rigor that the language model lacks on its own (Robison, 2025). Together, they solved 25 of 30 Olympiad problems, approaching the performance of a human gold medalist, an outcome that illustrates precisely the kind of complementary hybrid architecture many researchers have argued is necessary for robust, trustworthy general intelligence.
OpenAI’s o1 and o3: Language Plus Reinforcement Learning
OpenAI’s o1 and o3 reasoning models combine transformer-based language generation with reinforcement learning that teaches the model to allocate reasoning effort to hard problems, searching through chains of thought and self-correcting in ways that basic next-token prediction cannot produce.
The result is a system capable of earning gold medals at the International Mathematical Olympiad and performing at elite levels in competitive programming, capabilities far beyond those of a purely supervised language model (Potkalitsky, 2026). The RL component is doing something categorically different from the language modeling component, and the combination is demonstrably more powerful than either in isolation.
Figure AI: Language Planning Plus Motor Control
Figure AI, which incorporates research originally developed at Vicarious, uses LLMs for high-level task planning and language-based instruction-following, while deploying separate neural architectures for low-level motor control and physical interaction.
When a user instructs a Figure humanoid robot to perform a household task, the language model interprets the instruction and decomposes it into a task plan, while a distinct control stack handles the real-time physical execution. This architectural division mirrors the way neuroscientists describe the human brain’s own organization, a deliberate, slow prefrontal reasoning system working alongside faster, more automatic motor and sensory systems, and points toward how physically embodied AGI might eventually be structured (Websolutioncentre, 2026).
Amazon: Neural Perception Plus Symbolic Constraint
Amazon has applied neuro-symbolic techniques to both its Vulcan warehouse robots and its Rufus shopping assistant, adding symbolic reasoning layers to reduce the hallucinations and logical errors that arise when pure neural systems operate in high-stakes, rule-governed environments (Robison, 2025).
The emerging pattern of neural perception combined with symbolic constraint is becoming a practical engineering solution for organizations that need AI to be reliable, not merely impressive.
Why Building AGI Beyond Language Models Requires a Pluralistic Approach
The broader picture here is one of a field in the early stages of discovering what general intelligence actually requires. LLMs demonstrated, definitively, that language-shaped reasoning can scale to remarkable levels and, in doing so, made AI accessible and useful to hundreds of millions of people. Furthermore, they surfaced the next set of challenges with equal clarity: planning over long horizons, learning efficiently from limited data, grounding knowledge in the physical world, reasoning causally, and operating reliably in genuinely novel situations.
LLMs as a Transitional Technology
As Potkalitsky (2026) summarized the emerging consensus at the start of this year, the path forward requires solving fundamental research problems centered on new paradigms, specifically systems that learn from direct experience with the world rather than from pre-digested human text. LLMs are, in that framing, a transitional technology: they used computation to absorb human knowledge at scale, and the next phase will use computation to learn directly from experience.
A Pluralistic Path Forward
As Adegoke et al. (2025) concluded in their systematic review, “symbolic reasoning, neuromorphic computing, or hybrid AI architectures may offer more viable pathways to AGI,” underscoring the importance of exploring multiple paradigms in parallel rather than concentrating the field’s resources on scaling a single one.
The most compelling paths to general intelligence, therefore, are those that treat cognition as a multi-layered, multi-system challenge, and building AGI beyond language models by combining reinforcement learning, world models, neuroscience-inspired architectures, and symbolic reasoning is precisely how the field is beginning to do so.
References
Adegoke, B., Nwoye, C., Okonkwo, F., & Eze, H. (2025). Navigating artificial general intelligence development: Societal, technological, ethical, and brain-inspired pathways. Scientific Reports, 15, Article 8821. https://doi.org/10.1038/s41598-025-92190-7
Potkalitsky, N. (2026, January 6). Understanding AI in 2026: Beyond the LLM paradigm, or what’s actually required for progress. Public Services Alliance. https://publicservicesalliance.org/2026/01/06/understanding-ai-in-2026-beyond-the-llm-paradigm-potkalitsky/
Robison, G. (2025, November 19). Neuro-symbolic AI: A foundational analysis of the third wave’s hybrid core. Medium. https://gregrobison.medium.com/neuro-symbolic-ai-a-foundational-analysis-of-the-third-waves-hybrid-core-cc95bc69d6fa
Singularity University. (2026, February 2). The Singularity monthly: Chasing AGI. https://www.su.org/resources/the-singularity-monthly-chasing-agi
Websolutioncentre. (2026, February 11). AI world models revolution 2026: Beyond language models. https://www.websolutioncentre.com/blog/2026/02/11/ai-world-models-revolution-2026-beyond-language-models/
