The Core Thesis
The recent successes of generative AI have convinced some that AGI is imminent. The author argues this is a mistake. These models have emerged not because they are thoughtful solutions to the problem of intelligence, but because they effectively memorize and interpolate on massive datasets. The most emblematic case of this misplaced confidence is the multimodal approach, in which massive modular networks are optimized for an array of modalities. The author argues this strategy will not lead to human-level AGI capable of sensorimotor reasoning, motion planning, and social coordination.
Instead, we should pursue approaches that treat embodiment and interaction with the environment as primary, and see modality-centered processing as emergent phenomena.
The World Model Debate
Proponents of the LLM world-modeling hypothesis cite the prowess of state-of-the-art models on benchmarks and the idea that "language mirrors the structure of reality." The author is skeptical. Evidence from OthelloGPT, where researchers predicted game boards from transformer hidden states, does not generalize to natural language. Othello fundamentally resides in the land of symbols and can be played with pen and paper. One cannot sweep a floor, do dishes, or drive a car with pen and paper.
The author conjectures that LLM behavior comes not from a learned world model but from brute-force memorization of incomprehensibly abstract rules governing symbol behavior, a model of syntax rather than semantics or pragmatics. Syntax studies sentence structure. Semantics boils language down to literal meaning. Pragmatics involves interpreting speech while reasoning about the environment and the intentions of other agents. It is the fusion of these disparate cognitive abilities that produces human language understanding, and LLMs may be reducing semantics and pragmatics to syntax.
The Bitter Lesson Misread
Sutton's Bitter Lesson, that methods leveraging computational resources will outpace those that do not, has been misinterpreted by scale maximalists as meaning that making assumptions about the structure of AI is a mistake. The author counters that it is precisely when humans think deeply about the structure of intelligence that major advancements occur: convolutional layers assumed translation invariance and kickstarted modern computer vision; the Transformer assumed long-distance relationships between symbols and made ChatGPT possible; 3D Gaussian Splatting assumed the solidity of physical objects.
The real question is how to heed Sutton's lesson in AGI development. Scale maximalism worked for LLMs and large vision models because we had natural deposits of text and image data. An analogous application to AGI would require forms of embodiment data that we simply do not have.
Why Multimodal Scale Maximalism Fails
There are multiple issues with stitching modalities together. First, each new modality increases the dimensionality of the concept space, making concept synthesis ever more difficult. Modern approaches encode percepts from all modalities into the same latent space, but these spaces do not cogently capture all information relevant to a concept. Meaning resides not in the vector but in the way modality-specific decoders process it. As long as encoders and decoders are subject to modality-specific training objectives, "meaning" will be decentralized and potentially inconsistent.
Second, it is not clear that today's modalities are an appropriate partitioning of observation and action spaces for an embodied agent. The human capacities for reading, seeing, speaking, and moving are ultimately mediated by overlapping cognitive structures. Making structural assumptions about how modalities ought to be processed is likely to hinder the discovery of more fundamental cognition responsible for processing data in all modalities.
Third, the learn-from-scale approach trains models to copy the conceptual structure of humans instead of learning the general capability to form novel concepts. Humans spent hundreds of thousands of years refining concepts and passing them memetically through culture and language. Today's models are trained only on the end result of this process. The flexibility to form new concepts from experience is a foundational attribute of general intelligence.
The Alternative
Instead of pre-supposing structure in individual modalities, we should design settings in which modality-specific processing emerges naturally. Research on visual theory of mind saw abstract symbols naturally emerge from communication between image-classifying agents, blurring the lines between text and image processing. Eventually, we should hope to reintegrate as many features of intelligence as possible into a unified cognitive process.
One path forward is to process images, text, and video using the same perception system and produce actions for generating text, manipulating objects, and navigating environments using the same action system. What we lose in efficiency, we gain in flexible cognitive ability.
Conclusion
The most challenging mathematical piece of the AGI puzzle has already been solved: the discovery of universal function approximators. What remains is to inventory the functions we need and determine how they ought to be arranged into a coherent whole. This is a conceptual problem, not a mathematical one. The overall promise of scale maximalism, that a Frankenstein AGI can be sewn together using general models of narrow domains, is extremely unlikely to yield an AGI that feels complete in its intelligence.