Webbrain is an initiative to develop an emergent AI using the infrastructure provided by the internet. The brain is a marvelous biological machine. With billions of neurons and trillions of connections, it provides the substrate from which our behaviours emerge. Our brain can help us move in multitudes of environments, set goals, solve practical and abstract problems, use languages, vision or pattern matching. Could we use computers on the Internet to create a global thinking machine with similar capabilities?
Neurons as Nodes or Processes:
An adult human brain posesses around 85bn neurons (≈ 8 × 1010). Each neuron has thousands of connections leading to estimates of 2 trillion cortical synapes (= 2 × 1014) or 2 trillion internconnections between neurons in the cortex (which contains around 20% of all the neurons).
Is it possible to create an artificial system of a similar scale? If we add the servers from the internet giants like Google, Amazon, Facebook, Microsoft, Baidu, Tencent, Taobao there are between 80 - 150 million servers worldwide. If we could run 1,000 tiny programs in each one of them (say lighweight processes), we would have ∼ 100bn nodes (or agents). Of course other architectures are possible: we could use 5 million servers, 10 million PCs, and run 6,000 processes in each one of them giving us 9 × 1010 neurons.
Synapses as Connections or Messages:
The next element in our analogy are synapses and the latency of the communication between neurons. To reach the equivalent of 2 × 1014 synapses we would need around 2,000 connections per node. But could these tens of thousands of processes communicate with each other fast enough? Could they send messages within a machine and to other machines fast enough? The neuroscience literature shows latencies of 10ms - 50ms for a neuron and 500ms for a group of neurons creating a reaction. This is not unlike some of the latencies found on the internet which range from 1ms - 500ms. Of course the time-domain topology of the network is vastly different. But what matters is that the order of magnitude is comparable.
Certain programming languages are very well suited for the simulation of neurons. This project uses Erlang / Elixir to create the nodes of the WebBrain. This language can send tens of messages per milisecond between neurons. Again the details of the topology will matter. But speed may not [yet] be the key ingredient for this intelligence.
