What Backpropagation Leaves Open About How Intelligence Learns

NotebookLM text comparing synaptic plasticity in biological neural networks with global calculations — My AI teacher teaching me about AI explains why current AI isn't the future.

This is from my NotebookLM summary of Harvard's open-source AI course (it's on GitHub and it's FREE — look it up).

The course focuses on the transformer era (post-2017), but buried in the material is something that I think is being massively underappreciated: backpropagation — the training method behind every major LLM — relies on global calculations across the entire network.

Biology doesn't work that way. The brain learns through local chemical processes at the synapse level, continuously, in real time.

That gap isn't a minor implementation detail. It might be the fundamental reason why transformers, no matter how big we scale them, can't self-regulate — and self-regulation is arguably the core requirement for AGI.

Worth thinking about.

#artificialintelligence
#machinelearning
#neuralnetworks

Kendrick is comparing two ways a system can change with experience. In today’s machine-learning practice, backpropagation computes how much each adjustable connection contributed to an output error. A gradient is the direction and amount those connections should change to reduce that error. The error signal is the information carried backward through the network to make those updates possible.

This is not a failed or merely theoretical technique. Backpropagation is proven engineering machinery. The 1986 paper by Rumelhart, Hinton, and Williams showed that it could help networks learn useful internal representations rather than requiring a person to specify every feature by hand.[1]

Where the biological comparison gets interesting

A biological synapse changes through plasticity: activity and local cellular processes alter how strongly one neuron influences another. That makes strict backpropagation look biologically awkward. The textbook algorithm appears to require information and coordination that a synapse may not have available locally.

But “awkward” is not “disproved,” and the neuroscience question is not settled. A major review describes strict backpropagation as historically difficult to reconcile with biology, then surveys feedback mechanisms that could provide workable local approximations rather than a literal copy of the engineering algorithm.[2] Lillicrap and colleagues are useful here because they keep both facts in view: the biological objections are real, and possible bridges remain under active study.

Why you should care

The practical question is not whether a larger transformer is doomed. Kendrick’s post raises that possibility; the evidence cited here does not establish it. Nor does the research establish that self-regulation is a required definition of artificial general intelligence.

The useful business lesson is narrower. A model trained before deployment and a system expected to learn safely while operating are different products. If a proposed system must adapt continuously, preserve stability, respond to local feedback, or explain when it should stop changing, model scale alone does not answer the operating-design questions. You still need explicit feedback, permissions, evaluation, rollback, and human accountability around the model.

The course behind Kendrick’s note is Harvard’s official CS50 Introduction to Artificial Intelligence with Python, which covers foundational AI topics and is available as open course material.[3] The post is best read as a research question prompted by that material—not as a report of experimental results.