Listening in on https://carboncopies.org workshop, and realizing that one way of expressing the goal of whole brain emulation is: to apply the Bitter Lesson in AI to making neural models.
Comments
Log in with your Bluesky account to leave a comment
I don't really get what you have in mind. But if the idea is to create an "emulated brain" based primarily on training to match human behavior, that is very dangerous. We will have no way of knowing whether those people are conscious
I'm confident that a sufficiently accurate simulation would be conscious. But we don't know what level of accuracy is needed. If we work bottom-up, based on reconstructing known structures, then when it wakes up we know we succeeded
But if we put in the desired behavior as training data, then we don't know - and likely will never know - whether the internal emulation is good enough to have qualia
Sure. My point is to emphasize that with the old-fashioned "simulation" approach, this was much less of a danger - if we didn't reproduce whatever the brain is doing, including consciousness, then it just won't work. Whereas a "training" approach gives you the output no matter what
The Bitter Lesson is the observation that brute force, or rather the power of general purpose methods that scale with computation, tends to beat building in human knowledge in the models... eventually. http://incompleteideas.net/IncIdeas/BitterLesson.html
Listening to the great talks on how the currently best models are made, they are clearly to a large degree handcrafted (but by no means as much as past models!) But we are moving away from this as more data arrives and optimization of unknown parameters become feasible.
Thinking about the whole pipeline from biology to model, for WBE to be doable we need not just the right scanning tech, but reconstruction, model tuning, and validation. Right now this is heavily tied to human understanding of neuroscience and tools. That is unlikely to scale.
I imagine a successful WBE project to include a lot of AI in the pipeline steps, and AI supported science (the as yet little implemented part) to perform the iterative feedback loop of updating the steps - if we need many people there, it will be too slow and bottlenecked.
Comments