It took me (using Claude code and some codex), about 3 hours to put it together
And even though it was a cool demo, it seemed so easy, that it also felt like it wasn’t worth sharing
0: ChessFly (not mine), uses the FlyWire connectome (the fly’s brain’s weights) to play chess https://huggingface.co/spaces/mlabonne/chessfly
More generally, anything can be said about anything.
In the end, nothing can be said about many things and many things say nothing. Those things are then left to interpretation.
/s
It's a small machine, so it might get bogged down
And agreed that this isn’t a faithful simulation of a fly brain. I’m using the connectome as the network structure/parameters for a computational model, then using its outputs as the teacher for the classifier. Not sure how the ChessFly uses it
Edit: in any case, these are just fun demos, they aren’t research papers trying to claim accurate physiological fly brain software simulations
That sounds like you're saying that they got you on a minor technicality only. That's not the case; they're right, you're wrong – you did not "download a real fly's brain's weights".
There's a lot more going on at synaptic clefts, so just knowing the number of synapses doesn't tell you enough to model anything. You still need to know which neurotransmitters are used, how much, second-order effects like G proteins, basal firing rates, distance to the axon hillock, the shape of the neuron's effect on potential decay, etc.
And that's all to predict whether one neuron will fire. You could get very different behavior between different two neuronal pairs having the same synaptic count.
As far as I understand, you downloaded the structure of a neural network, ran it with essentially arbitrary weights, trained a classifier on the essentially arbitrary behavior, then simulated an approximation of that arbitrary behavior.
You could argue that there might be biases towards certain behaviors encoded in the connectivity, and I'm sure you'd be right, but your experiment is incapable of differentiating between those interesting behaviors and random noise. Especially because flies sorta act random anyway.
I have no idea about those actual models, but there are layers of models that you can create, and for each, you can explore with data and compute
It's not a free lunch though. Depending on the task, you might need to collect a lot of the data, or review it manually, or pay a lot for compute, or wait a lot for compute. And still have to iterate a lot on the results, and do your own explorations as a human operator/driver of the whole thing. And then create the materials, test them, get funding to do the whole thing... so theoretically, I think we are in a place where we can successfully apply models to a lot of things, but realistically, we won't be applying all the resources to everything
The story about OpenAI's Navier-Stokes solution is a good example of what I mean. I don't think it would have been possible without computer assistance because that proof is long and complicated. I'm also not sure that it would have been possible without a human proposing a new approach to the problem, because by all accounts that's exactly what led to the absurd amount of spending that OpenAI did to solve the issue.
I feel like that at least implies that there's some room left for humans in the new world.
But stop thinking that just because they were initially simpler solutions that it stayed that way. We are in the singularity now!
It took a human being to see the beauty of an alternate approach and purpose it, and only then could OpenAI spend millions of dollars to squeeze out a result.
I am very doubtful that we will see models capable of making intuitive leaps anytime soon, or being "creative." Neither of those are straightforward objectives to train against.
Thus, you can build ai/ml models+agents to explore those spaces, at a speed and scope much larger than what any human can do
I can imagine findings like these are going to keep increasing in frequency to a point in which the bar for novelty goes a lot higher