The brain decides earlier than we thought: the clue that explains why today's AI burns so much energy

By Tech Times · July 13, 2026.

A team at the University of Illinois Urbana-Champaign, led by Yurii Vlasov (formerly at IBM, where he worked on neuromorphic computing) together with doctoral student Alex G. Armstrong, published in PNAS a finding that corrects a textbook assumption: the primary somatosensory cortex (S1), the first cortical stop for tactile information, does not merely relay data upward for the frontal cortex to decide. In mice navigating a virtual-reality corridor using a single pair of whiskers (a deliberate experimental bottleneck to isolate the signal), the researchers saw that, during the evidence-accumulation phase, the activity of many neurons in S1 collapsed into a single latent variable that rose in synchrony —a pattern until now associated with decision regions, not with input sensory cortex. And it did so because premotor and frontal regions were sending feedback signals downward, modulating in real time what S1 encoded. It is systems-level evidence, in naturalistic behavior, for something predictive-coding theory had been proposing for years: that the brain does not process in a clean cascade, but continuously negotiates between what it expects and what it perceives.

The finding matters beyond neuroscience because the feedforward model —information rises layer by layer, without the upper layers rewriting what the lower ones have already computed— is literally the blueprint of convolutional networks and, with nuances, of the transformers underpinning today's large language models. That blueprint was built on an image of the brain that, according to this study, was incomplete. Vlasov puts it bluntly: he wants to learn from a billion years of evolution to make AI "more energy-efficient, less power-hungry and smarter," and he points to the decision level as the current weak point. The comparison the field itself uses is striking: a human brain runs on about 20 watts; training a frontier language model can consume energy equivalent to that of hundreds of homes over a year. Part of that gap is, very likely, architectural: the brain distributes the commitment of deciding across several levels of the hierarchy at once, whereas feedforward AI concentrates that commitment at the end of the chain and has no systematic mechanism for the result to modulate the early representation.

It is worth being precise about what this work does NOT demonstrate. It is a study in mice, with one sense —whisker touch— and a binary-choice task; the authors themselves insist it is not a recipe for building better AI, but a constraint: any model of the brain that ignores early modulation by feedback is, from now on, incomplete. The team itself suggests the next step is to understand the fine temporal dynamics —on a millisecond scale— of those feedback loops, and that only from there might, perhaps, emerge an architectural principle transferable to AI. In other words: we are in the mechanism-understanding phase, not in the design of chips or recurrent networks that replicate it. Distinguishing this from the eye-catching headline is exactly the kind of judgment worth applying when neuroscience promises lessons for engineering: the analogy is suggestive, the implementation is an open research problem spanning several years.

Our reading is that this finding adds to a thread we have been pointing to: energy efficiency is not an infrastructure detail, it is a design constraint that determines who can afford to train and run frontier AI. If the real bottleneck is not just raw compute but architecture —and this study suggests that feedback loops do "computational work" that feedforward stacking, however deep, cannot replicate— then the next generation of efficiency gains might come not from more parameters or more GPUs, but from rethinking how information flows within the model. That connects directly with the abundance we advocate as a long-term horizon: an AI that needs a fraction of today's energy to reason with the same or better quality brings frontier-level power within reach of more players —midsize companies, countries without megaclusters, local hardware— instead of reserving it for those who can pay for dedicated power plants. In the short term, however, the message is one of patience: neuroscience is still deciphering "a largely unknown language," in Vlasov's words, and the translation to silicon, if it comes, will take years. It is one more piece of the puzzle, not the blueprint of the next architecture.