LessWrong essay: the path to a computer copy of the brain may first give AI fragments of human intelligence
LessWrong essay: the path to a computer copy of the brain may first give AI fragments of human intelligence
On September 5, an author writing under the name TsviBT published an essay on LessWrong about whole-brain emulation, a computer model that could reproduce a person over an extended period. He proposes evaluating the path to such a model by examining its intermediate results: which data and methods will become available first, and who will be able to use them.
Whole-brain emulation would reproduce many brain processes working together. TsviBT uses the term partial emulations for models, data, methods, and simulations that reproduce specific capabilities of the brain without reproducing the whole person. He suggests that developers of more powerful AI may be able to use some of these results before whole-brain emulation becomes possible.
“Extracting some capabilities from the brain and transferring them to a computer is much easier than transferring everything,” TsviBT writes.
TsviBT calls this interval the “bad knee”: parts of the research may already accelerate AI development, while a digital person does not yet exist. In his strategic model, the harm caused by this acceleration may outweigh the future benefits of whole-brain emulation.
The discussion below the essay shows that the risk depends on the research path. One comment compares a model trained on recordings of brain activity with connectomics, which maps neurons and the connections between them. The first method directly learns to reproduce how the brain works. The second describes its wiring. These methods may therefore produce different kinds of knowledge, although both contribute to emulation research.
A technical 2025 review explains this distinction: a connectivity map shows the brain’s wiring, while a model of brain function requires activity data. The authors propose training models to infer function from structure in organisms for which both types of data are available, then extending the approach to the larger brains of mammals.
TsviBT proposes evaluating each emulation program by identifying which intermediate models, data, and methods it makes available before whole-brain emulation is achieved.