Multimodal AI will not produce AGI. That is the central claim of this piece from The Gradient, and the author backs it with a specific target: sensorimotor reasoning, motion planning, and social coordination. Current large multimodal networks scaled on existing hardware because the hardware was available, not because the architecture was correct. The author's alternative is not incremental. Embodiment and environmental interaction should be treated as primary, with modality-specific processing as emergent, not the other way around.
The most technically precise section concerns OthelloGPT. Researchers found transformer hidden states could predict Othello board positions from legal move sequences, which supporters of the LLM world-modeling hypothesis cited as evidence. The author dissects why this fails to generalize: Othello is a purely symbolic game that happens to use physical tokens, so move sequences provably encode full board state. Physical tasks like sweeping a floor or driving a car have no equivalent symbolic closure. A separate finding reinforces this: OthelloGPT learned heuristics specific to its training distribution, such as inferring square B4 is empty if its token did not appear before A4, rules that do not hold across all legal Othello games. The next-token objective selects for whatever is easiest to learn, not for ground truth world structure.
What makes this worth reading in full is not the conclusion but the argument about the predict-next-token objective as a selection pressure. The author is not saying LLMs are useless. The claim is narrower and harder to dismiss: the objective function actively deprioritizes physical world modeling whenever a cheaper heuristic achieves equivalent token prediction loss. That framing recontextualizes benchmark performance entirely. A forthcoming MIT Press volume edited by George Konidaris, titled Designing an Intelligence, is cited as extending these arguments.
[READ ORIGINAL →]