A / CURRENT ASSUMPTION
Perfect infrastructure is presumed.
Most machine-learning systems begin with fast links, stable clusters, and centralized coordination.
00 / Premise
Centralization is an architectural choice, not a law.
B / RESEARCH QUESTION
What if the network is the experiment?
We study coordination where peers are unreliable, bandwidth is finite, and no machine controls the whole.
01 / Active project
One attempt to make collaborative learning Internet-native.
NeuronWire
An experimental distributed-systems runtime exploring collaborative machine learning over peer-to-peer networks.
● ACTIVE PROTOTYPE
Rust / peer-to-peer / experimental
Rust / peer-to-peer / experimental
Can useful coordination emerge without a datacenter at its center?
View source on GitHub↗02 / Evidence surface
Signals of work, not proof of completion.
Approximate lines of Rust in the current prototype.
Regression and component tests. Count is not a quality claim.
Dashboard surfaces should link each result to method, configuration, hardware, and source.
Experimental — early evidence exists; broader validation is still required.
Deterministic harness and synthetic distributed simulations.
Configurations, tooling, CI workflows, and reports should remain inspectable.
03 / System anatomy
Independent subsystems. One research runtime.
#1 transport / foundation
Transport Layer
UDP and TCP today. NAT traversal and encrypted transport are planned, not shipped.
#2 discovery / experimental
Peer Discovery
Kademlia DHT with latency-aware routing and adaptive, trust-scored peer selection.
#3 exchange / research
Gradient Exchange
Gossip-based aggregation tuned for convergence under constrained bandwidth.
#4 runtime / prototype
Runtime Engine
Event-driven, largely lock-free, deterministic scheduling, low memory overhead.
#5 experiments / instrument
Experiment Framework
Configurable topologies, synthetic datasets, failure injection, deterministic replay.
#6 metrics / evidence
Metrics
Convergence, latency, bandwidth, routing efficiency, packet loss, utilization.
04 / Known limits
Credibility begins where the polished story ends.
The prototype is not a finished framework. These are open technical constraints, not footnotes.
×Internet-scale deployment has not been completed.
×End-to-end learning convergence is not fully verified.
×NAT traversal remains incomplete.
×Security mechanisms and Byzantine tolerance require further research.
×Large-scale heterogeneous hardware has not been validated.
×GPU acceleration is experimental; mobile participation is unexplored.