Where does a decision begin?
I read the study published in PNAS and saw it looking less at neuroscience than at AI engineering. Yurii Vlasov and Alex Armstrong of the University of Illinois recorded brain activity in mice navigating a virtual reality corridor and making perceptual decisions. Classical theory says information flows in a strict hierarchy from sensory regions up to the frontal cortex, where the decision is made last. The finding disrupts that: decision-related activity shows up in the primary somatosensory cortex (S1), one of the brain's earliest sensory areas. More, S1 doesn't just pass information forward; it's shaped by feedback loops from higher regions. There's a continuous, two-way conversation.[1]
Why in an AI column? Because many of today's neural networks — including convolutional ones — were inspired by exactly that assumption of one-way processing in the brain. Vlasov's line could be this column's motto: 'We want to learn from a billion years of evolution... Can we learn from the architecture of the brain and emulate it to make AI more effective, less power hungry, and more intelligent? At the level of decision-making, that's where current AI is lacking.'[1]
Humble science, an ambitious book
I liked the scientists' honesty: the study 'does not provide a blueprint for building better artificial intelligence,' only a clue that might inspire future architectures. The next steps are modest too — measuring signal timing more finely, developing new tools to understand how feedback loops emerge. No promise of a miracle; a method. That tone is exactly what I look for in this column.[1]
A book out the same week is the coin's other face. Computer scientist Peter Denning argues in 'Turing's Mistake' that Turing's two 1950 assumptions — that intelligence can exist independent of a body, and that a machine proves intelligence by imitating a human in conversation — sent AI down the wrong path for 75 years. To Denning, large language models 'only manipulate words, they cannot know or understand the meaning of what they are saying.' As evidence he cites the Cyc project: four decades of work, roughly 25 million entries, yet not enough common sense to make expert systems truly expert. Denning's thesis is contested and bold — I don't share it, because even without 'meaning,' models solve measurable tasks measurably well. But putting the counter-voice on the table resonates nicely with this week's humble mouse-brain lesson: we've fully cracked neither the brain nor the machine.[1], [2]