I believe P2P rose to prominence two decades ago as a response to the cost of bandwidth. I wonder if similar methods could effectively overcome the cost of compute for LLMs. Here are two projects I found from a quick search:
Serving: https://petals.dev/
Im using it for Payload[1] in for LAN and WAN transfers (if possible). Reduce operational costs (especially if you run on public clouds and have to pay extortion rates for egress) and also you must use it to capitalize on latency/throughput in LAN. Moving data from A->server->B means your need multiple servers on the edge, which means you kinda need to depend on mega-corps. If your destination is closer it’s easier for your application infra. I’d like to reverse the question, why send all data through another machine in the cloud if you don’t need to?
That said, p2p being flaky and bad is real. It’s a self-fulfilling prophecy, because middlebox engineers say ”let’s add these layers of garbage and nobody will notice unless they use p2p but its so bad who uses it anyway”. Well, yeah. It’s worse because of you! Philosophically, I also think p2p is a necessary precondition to a decentralized internet without tiers (ie client and server separation).
Anyway, rant aside, you have to currently have a relay backup if you need availability. P2P will fail often even with the smartest hole punching algorithms. This makes things more complicated, because you need a hybrid solution. However, it’s not as complicated as WebRTC, that thing is an overengineered mess. It works, but I don’t like the complexity it brings.
[1]: https://payload.app/
The loss of a "central server" or whatever never matters.
There are a lot of things that do P2P under the hood, usually with cloud relay fallback so it always works. You just don’t hear much about it because it’s not a selling point, just an under the hood detail.
*shudders*
I'm aware that with things like this you're supposed to use the latest and greatest like Wireguard or whatever, but nothing really does the p2p thing as easy as Tinc, and given secondary encryption measures (e.g. I'm sshing and httpsing to those machines) I'm just not worrying much about it right now.
Tinc will also do a layer 2 tunnel if you want (tap) while wireguard is only layer 3 (no broadcast/multicast).
The big con of tinc (and most VPN solutions) compared to wireguard is performance. Wireguard is a small kernel module which can process traffic very fast.
# Zhejiang ICP No. 2024137671-1
It takes you to some government website but it is not clear whether this is a business registration or something else.
I'm not sure if I'd be up for that, to be honest...
It's useful when you have computers that talk to each other over the internet, likely without public interfaces, and using protocols that may or may not be secure.
this would mean, for instance, torrents that are wireguarded between peers by default. sure you will see tons of IPs connected via wireguard but who is going to bother intercepting them?
How does it square up against DPI censorship techniques that successfully block WireGuard?
Ironically, it was American companies that sold firewall tech to the CCP: https://www.cfr.org/backgrounder/us-internet-providers-and-g...
I don't know why people focus on Tor and censorship associations. The meaning of a VPN is just a virtual network between devices, not anonymization.
If every node is both a server and a client then will a lot of traffic use my node/server as an exit node?
I see there is a separate list of public servers. Presumably, these are people running EasyTier nodes/servers who are willing to allow strangers in?
If I start my own node and I wish to connect to the mesh is that part of the reason for pubic nodes?
Rust vs Go is one difference. What else?
Tailscale: https://github.com/tailscale/tailscale
Do I need to run the service first?
