Best, Alex
I punch send, then go outside and chain-vape for half an hour. When I come back, there's already a reply. Dear Alex,
Thank you for your interest. I am available for a brief video call Friday 17:00 CET (8am your time). Please send your questions in advance if possible. Regards, Prof. Marchetti
I lose sleep for two nights prepping for the call. I read the preprint three times, annotate every line, and try to preempt every question I could be asked. I also rewatch a YouTube interview with Marchetti, who is even more Italian than I expected and has a way of talking with his hands even when they are not visible on camera. The call starts awkwardly. I manage not to say "um" for the first twenty seconds but then fall apart and revert to my usual stammering. Marchetti is polite, but I can tell he's waiting for me to say something original. "So, Alex, you believe entropy flow is the missing link between symbolic reasoning and distributed models, yes?"
"Yeah. I mean, yes. If you look at the spectrum of learning rates in current deep nets, there's always this bottleneck that mimics the information loss in classical thermodynamic systems. But nobody is formalizing that loss as a tool. It's always treated as noise."
Marchetti leans forward. "And you think this noise could be harnessed?"
"Exactly!