Act III — Then the machines started promising

The root cause In the 1990s my father’s health deteriorated as a result of what was later thought to be some immune system related disease. He found out that living in warm climates with little pollution and no need to stay indoors for very long helped him maintain his health. So he took a bold move, sold his house, and became a traveling academic. In the summers he would stay in the south of France where we visited him, in the shoulder seasons he lived on one of the Canary Islands, and in the winter months he traveled to the southern hemisphere (Chile, Africa, Australia). Jokingly he called himself a “refugié pollutique”. ...

As we mature processes, they come within reach of actual automation. Especially IT management processes. To paraphrase General Carl von Clausewitz, who said “War is the continuation of diplomacy with different means”, we can say that automation is the progression of process maturity with different means - replacing human effort with software and systems. My favorite historical example is IP address allocation. In the past, whenever there was a new computer, you would walk up to the head of the lab or data center and ask for a new IP address for that machine. You may remember that. ...

There are many ways to influence and control a system, but one of the most controlled (pun intended) ways is to use feedback loops. Let’s start with an analogy first: driving a car. If you want the car to stop as fast as possible, just push the brakes as hard as you can. But that does not give you much control. For example, your passengers might not like it, and it leads to wear on the tires. If you want the car to stop at an intended location, for example just in front of a traffic light, you apply a feedback loop. You adjust the pressure on the brake continuously, based on the remaining distance. ...

Automation is a big part of IT, but not all automation brings value. Automation takes time and effort, upfront, and any benefits come later. Think of the IT projects you were involved in, most of those aimed at providing a measurable benefit somewhere. This diagram (from xkcd, a series of webcomics) illustrates when automating repetitive tasks, or even just parts of them, brings benefit. In other words, is there a business case for such an intervention? ...

How to fix the WiFi? How to find a new phone for grandma? Applying technology follows certain rules. However, many people only have an intuitive understanding of these rules. As a result, technology is not optimally applied. By understanding how applying technology really works, you can be more effective, more efficient, reduce waste, and overall do a better job. It takes just a few simple steps to improve any attempt at applying technology. We’ll focus on information technology here, though most principles have wider applicability. ...

Cloud-native software development enables new practices. But it also requires them. It is a new level of working. However, putting all these new practices together requires integrating a lot of pieces. To illustrate this new approach, I have started to develop a minimal application. Although minimal, I run it in production. Its basic function is to regularly pull out data from an air quality sensor into a cloud-based database. Together it is a few hundred lines of code. Here are the major features of the example: ...

The AI landscape has many digital infrastructures. Let’s explain this step by step and focus on which data is stored where and how it is processed. A core element of AI systems is a trained model. This is especially true for the dominant form of AI these days: deep learning neural networks. A trained model is the result of processing a lot of training data by a specific neural network. These models are fixed in size, but typically very big. The smallest useful models are close to a gigabyte, while recent public chat models run into multiple terabytes. LLM training costs hundreds, thousands, or more hours of compute time on specialized hardware. ...

AI systems, like most cloud systems, are composed of models, servers, software, and more, each of which can be provided by a different role. We follow the roles outlined in the CSA AI Controls Matrix (AICM). Each role also has promises associated with them (read more on promise theory here). Together these run AI systems, from image recognition systems to Large Language Model based customer support systems. AI Customer (AIC) End users of AI applications. This includes the actual users as well as their organizational units. Even though they are only providing a service to themselves, they are still responsible for certain security and control functions. Some of these responsibilities are in the interest of their providers, and those are typically included in an AUP (Acceptable Usage Policy). For example, customers should maintain appropriate authorizations of users, and refrain from overloading their providers. Customers are also responsible for the business use of the data that is returned by the AI application. ...

Let’s look at a practical example from one of my clients. It handles voice calls using an AI-supported chatbot. architecture-beta group frontend(cloud)[Telecom operator] service telco(cloud)[Telco switches] in frontend service sips(internet)[SIPS Gateway] in frontend group backend(cloud)[Application Provider] service app_server(server)[Azure voice bot] in backend service twilio(internet)[Twilio] in backend service database(database)[FAQ Database] in backend service llm(server)[LLM] in backend group configuration(cloud)[Client] service faq(database)[FAQ mgt] in configuration telco:R -- L:sips sips:R -- L:twilio twilio:R -- L:app_server app_server:B -- T:database llm:B -- T:app_server faq:R -- L:database Who is actually responsible for what? This is important for a number of reasons: ...