The essays are often created in question-and-answer form. Answers are structured so that when a user asks an AI one of the key questions, the user gets the answer the advertiser has created. But instead of getting it delivered as an advertisement, which is received with a note of skepticism, it is received in an authoritative fashion, likely to be accepted by the user as ground truth. In today’s world of intelligent agents acting for people, the user may be an intelligent agent just as susceptible (and maybe more so).

Thus, today advertising effectiveness is measured by how much of the copy has been successfully inserted into training data sets, with the resulting probability that the content will be delivered by an AI in an inference response.

Traffic Evolving 

Architecting to influence AI behavior through impact on training data is just one example of a larger underlying sea change. Today, more than 50% of the traffic on the internet is machine-generated. It is likely that much of that has neither a human originator nor recipient. More than 50% of the compute/storage is AI. Those proportions will increase, as will the overall volume. Intelligent agents don’t sleep. They will proliferate. The resulting factories, transportation systems, workspaces, marketplaces, political organizations, forms of art and entertainment, etc. will continue to evolve.

Blurring of Communication, Processing and Storage 

Processing and storage are going through big changes as well. The building boom in AI data centers is driving us to hub-and-spoke networks (whether public or private). Traffic flowing to and from these big AI data centers will dominate communications, electrical power generation/distribution, processing, and storage. From an AI perspective, networks will be relatively dumb. However, this too will change.

Limitations in memory semiconductor fab capacity and scale advantages in current AI implementations are currently creating economic incentives that favor these large AI data centers. Over time, as fab capacity increases and memory prices come down, incentives will move to more AI processing/storage at the Edge, near the Edge, and in the network. 

One can think of the human nervous system as a model for the next-generation architecture. In humans, there is the brain (centralized AI data center) and a network of ganglia (Edge and near-Edge AIs ). The ganglia get smaller and more specialized as they get closer to the Edge in operation. The ganglions are generally part of our autonomic system. That is, we are generally not aware of their functioning.

For example, our brain makes a conscious decision to walk in a certain direction at a certain pace. That decision triggers a very large network of ganglia that control each of our many muscle systems. Some control our posture. Others get our legs to move. Still others control our toes to keep us balanced and standing upright. The same basic kind of tissue builds the brain, the ganglia, and the connections.

As memory prices come down and AI processors become more power efficient, we are likely to see hybrid AI systems that function in a way similar to the human nervous system. Inference requests and responses will be autonomously broken up and distributed over a network that includes Edge AI processing, modest-sized servers near the Edge, AI larger processors in networks between the servers and the large data centers, and the data centers themselves.

Because of the proliferation of Intelligent Agents interacting and cooperating with each other, the resulting network will transition from a hub-and-spoke architecture to more of a mesh-like architecture. This will make it harder to distinguish pure communications from pure processing and pure storage.