Thank you!
My intuition, and what the neuroscience suggests, is that specialized brain regions are each providing specific services to the overall network which allow the sum of behavior to be flexible, persistent, and efficient to train. It seems to me that we have so far done a pretty good job of building ML networks that do a good job of doing the individual tasks of specific subsets of the overall brain -- CNNs are great at visual perception, we have fantastic speech transcription models, and we're increasingly good at language processing, obviously. The thing I think is missing is a network approach where we try to understand what each region of the brain is doing, how that service integrates to the broader picture, and how those interactions can be meaningfully captured in code. It's worth pointing out that the proof of concept can be made at arbitrarily small scales. Some species of parasitoid wasps have fully functional brains, capable of navigating them in flight, with just a few thousand neurons. And, of course, C. Elegans, everyone's favorite model worm, with its ~300 neurons, is a perfectly functional organism. It should be possible to prove cohesive integration of all brain regions into a useful and persistent entity, at a very small scale. But maybe I'm wrong, and the answer really is "Just keep scaling deep, amorphous networks."
You're right that I am really just taking a bet on the direction that I think has the highest chance of making an impact. The reasoning behind my belief is that there is essentially an infinite amount of value locked behind robotic AI, and obviously all the labs want it. We have a ton of companies working on building robots, and they're all trying to run them but struggling. If I can build a network which does even marginally better at practical robotics than the other architectures, I think that would get attention. If it does substantially better at robotics, especially running on edge hardware, it seems to me that it would be guaranteed to get a *lot* of attention. Again, we have the robots, and all the decent people want the infinite production, post-scarcity future ASAP. That future is quite literally held up on practical robotics, and I'm hoping to take steps in that direction, in a way that helps the whole industry head that way.
Yep, the failure mode is fundamentally that I'm not able to make any gains compared to existing architectures, in video game playing, drone navigation, or general environment interaction. If I try my best for a good while, and I just can't figure out how to do any of the things I'm envisioning, I will have failed, and I'll have to move on. Though, for the last case, there's actually a very different story - if Atari 100k does actually go quite well, and nobody cares, then I just turn evil and unleash an army of AGIs on the world. Look out, Will Stancil. In seriousness, though, if the architecture does well at video games, I'm fully confident it will scale to robotics, and I'll start buying robots, uploading useful brains into them, and selling them for 20x. Gardening robot? Cooking/washing dishes robot? Cleaning robot? People would pay crazy money.
Thanks for the feedback!