The argument in one sentence
In a recent essay at The Gradient, Brown CS PhD student Benjamin A. Spiegel argues that the current strategy of building AGI by stitching together massive modular networks for separate modalities is doomed in the near term. It will not produce a system that can do sensorimotor reasoning, motion planning, or social coordination. The right path is to treat embodiment and interaction with the environment as primary, and to let modality-specific processing emerge from that foundation.
The argument cuts against the dominant 2026 consensus, which is that AGI is a multimodal problem solved by scaling language models with vision and audio adapters. Spiegel thinks this is the wrong shape of solution.
Why the multimodal path fails
The first reason is structural. To glue modalities together, you have to make implicit assumptions about how they relate. Today's AI processes images and text with completely separate architectures, even though both are visual in some sense. Humans use overlapping cognitive structures for text and image processing, and today's AI misses that overlap because the architecture forces separation.
The second reason is data. Multimodal training at scale requires aligned multimodal data, and the cleanest such dataset is internet text paired with images. That works for the kind of reasoning language models already do well. It does not work for the kind of reasoning required to navigate a cluttered kitchen, manipulate an unfamiliar object, or coordinate a handoff with another person. Embodied data is rare, and synthetic embodied data is shallow.
The third reason is computational philosophy. Multimodal models, in their current form, contradict the Bitter Lesson. The Bitter Lesson says that general methods that scale with compute eventually win over human-knowledge-encoded methods. Multimodal AI as practiced today is the opposite: it encodes a lot of human knowledge about which modality is which and how they relate. That encoding will become the bottleneck.
What the alternative looks like
Spiegel's positive case is for approaches that treat embodiment and interaction with the environment as primary. Modality-centered processing should be an emergent phenomenon, not a design choice.
This is closer to how biological intelligence works. A child does not first learn language, then learn vision, then learn motor control, and then integrate them. The child learns to act in the world, and language, vision, and motor control are all specializations of that more basic capacity. The architecture reflects the developmental priority.
For AGI research, this means less time building bigger multimodal adapters on top of language models, and more time building systems that learn to act and perceive in a unified way from the start.
The harder version of the question
The deeper question Spiegel is raising is whether scale alone is a path to general intelligence, or whether general intelligence requires a specific kind of architecture that current scaling efforts are not delivering. The answer depends on what counts as AGI. For language-and-vision benchmarks, the multimodal approach is winning. For sensorimotor reasoning, motion planning, and social coordination, it is not.
If AGI is defined as human-level performance on the latter set of tasks, then the current strategy is not just slow. It is missing the point. The fix is not more parameters or more data of the same kind. The fix is a different starting point: embodiment first, modality later.
What a research program would actually look like
If you take the embodied-first thesis seriously, the research program that follows is materially different from scaling multimodal foundation models. It means building agents that learn to act in a unified 3D environment, where perception and motor control are trained together rather than bolted together. It means investing in robotics platforms that can collect rich interaction data, and accepting that the resulting datasets will be smaller and messier than the internet-scale text corpora that have powered the LLM era.
It also means being honest about what current systems can and cannot do. A language model with vision adapters can describe a cluttered room accurately. It cannot reliably navigate that room, pick up a coffee mug without knocking it over, or hand it to a person in a way that respects their personal space. The first task is now largely solved. The second task is unsolved, and the current architecture is not close.
The empirical test
The empirical test of the embodied-first thesis is whether the next generation of robot foundation models, trained on real interaction data at scale, can match the generality of current language models on embodied tasks. If they can, the multimodal path turns out to have been a detour. If they cannot, the multimodal path is the right one and the embodied-first critics are wrong.
The answer is not in yet. The interesting bet is which research labs are positioned to find out. The labs that have invested in robot data collection platforms have a structural advantage. The labs that have invested in scaling multimodal foundation models have a different kind of advantage. The competition between the two bets is the most consequential research question in AI right now, and the answer will shape the next decade.