אִם יִרְצֶה הַשֵּׁם
🔗 arXiv:2604.18637v1 — NeuroAI and Beyond: Bridging Between Advances in Neuroscience and Artificial Intelligence (Zador, Fellous, Sejnowski, et al., April 2026)
TL:DR
Harnad won. This paper is the 20-year surrender terms, written by the victors so they can also claim the grant money.
Strip away the neuromorphic buzzwords, the 20-year roadmap, and the very polite NSF workshop prose, and here is what 26 of the smartest people in neuroscience and AI are actually telling you:
We built brains out of text, and now we’re shocked they don’t know what an egg is.
That’s it. That’s the paper.
They call it three "fundamental capability gaps":
- Physical incompetence. GPT-4 passes the bar exam but cannot clear a dinner table.
- Brittleness. Shift the distribution slightly and the system face-plants.
- Energy waste. Training runs cost megawatt-months. Only five companies on Earth can afford them.
They map these to five "neuroscience principles" that will save us: co-design of body and controller, prediction through interaction, multi-scale learning with neuromodulation, hierarchical distributed architectures, and sparse event-driven computation.
Sounds new. Sounds interdisciplinary.
It is also, point for point, what Harnad, Clark, and the embodied cognition people have been saying since the late 80s. You cannot get semantics from syntax alone. A symbol that has never been tied to sensorimotor consequence is just a dead token in a lookup table. An LLM's idea of "heavy" is a statistical ghost. A toddler's idea of "heavy" is scar tissue on their proprioception.
The paper knows this. It even cites the child picking things up before hearing the word. But it wraps the argument in connectomics, Neuropixels, and fruit-fly digital twins so it sounds like a forward-looking engineering agenda instead of a philosophical concession.
So what does the roundabout buy them?
Here is where I am not cynical. The philosophical victory is worthless without an engineering spec. Harnad told us what was impossible. This paper, for all its institutional throat-clearing, tries to tell us what is buildable.
And the spec is genuinely useful:
Event-driven (neuromorphic) sensing instead of frame-based cameras. The retina does not stream 4K video to the brain. It only reports changes. Like a security guard who sleeps until a door opens.
Co-design of morphology and control. Your tendon is a spring. Your fingertip has 240 mechanoreceptors per square centimeter. The body does computation the brain never has to. Stop designing robots like rigid CAD nightmares.
Neuromodulatory gating. Dopamine, acetylcholine, serotonin act as dynamic learning-rate schedulers. Not backprop everywhere. Backprop in context.
Digital twins with real biomechanics. The fly connectome exists. 140,000 neurons, 50 million synapses. Mouse is next. Human in 20 years, IyH.
These are not metaphors. These are testable mechanisms. The paper turns a negative constraint (you cannot do it without grounding) into a positive build order.
That is progress. Turning philosophy into engineering is what separates thinkers from builders.
But let us be honest about why this paper exists now.
The transformer and GPU scaling story is hitting the data wall. The curves are flattening. The energy bills are becoming a political problem. Industry poured hundreds of billions into a single architecture, and now that architecture smells like a local optimum.
So the field retreats into embodiment. It does so with enough NSF gravitas, roadmap milestones, and cross-disciplinary author lists to rebrand it as "the next frontier" instead of "we should have listened to the roboticists in 1990."
This is not a complaint. This is strategy. Smart strategy. When the money starts looking for the next thing, you want to be the person holding the map.
The paper is as much a funding brief as science. It asks for bilingual researchers (fluent in both spikes and backprop), community chip-design projects, open neuromorphic hardware, and benchmarks for continual learning and efficiency, not just accuracy.
ROI horizon: 10 to 20 years. Near-term wins: edge robotics, prosthetics, wearables.
Does the paper prove grounding is the causal ingredient? No. It shows embodied architectures beat text-only models on physical tasks, but that confounds grounding with better priors, richer data, and different compute. It does not run the ablation where you hold scale constant and only vary sensorimotor coupling.
Does it admit that grounding is necessary but not sufficient for human-level semantics? Barely. Social learning, language games, pragmatic inference get a nod in the long-term vision but not the analytical core.
And the "sub-kilowatt supercomputer matching today's clusters" is aspirational, not shown. Beautiful, but aspirational.
Read this paper as what it is: a restatement of the symbol grounding thesis with a biological engineering manual attached. Its real contribution is translating Harnad-style constraints into hardware requirements. That is genuinely valuable.
But the empirical case is still open. Better performance on embodied tasks does not, by itself, isolate grounding as the necessary mechanism, nor does it show sufficiency for semantic competence.
What it does show is that the smartest people in the room have stopped pretending that scaling a language model will spontaneously grow a body. The brain-in-a-vat era is ending. The era of the digital twin, the event-driven sensor, and the co-designed morphological computer is beginning.