From 2022 to 2024 we will see a new surge in the term artificial intelligence, driven by recent developments in language models such as GPT, which some call generative artificial intelligence, others copiative. Some academics say the next two years will have reinforcement learning as the new fashion, which also makes people understand how animal behavior works, which will also end up under the artificial intelligence umbrella. Enterprise Statistics usually talks about statistical learning, for statistical models, to refer to what some people call artificial intelligence, also in the context of neural networks, a model inspired by biology, then natural intelligence. Is the term artificial intelligence becoming an empty pipe? With this article I want to tease out a thought using different contexts.
Most readers of this blog have played at least one video game. Game difficulty depends not only on character and enemy parameters such as amount of life (HP), amount of damage, but on enemy behavior. This may in turn depend on a priori parameters, so something completely unintelligent, such as vision range, in the case of a war game. In more advanced games the enemy can change strategy according to our strategy (Half-life, Halo, Alien: Isolation, etc.), through a loop that keeps monitoring the environment and acts according to maximizing its reward: killing us. In other words, a series of “a priori” features bring out behavior that we can identify as intelligent, but not in the manner defined in another article.
Part of the readers of this blog have surely witnessed the public debate on nuclear energy post-Fukushima accident, or in the middle school exam they were presented with a model of a nuclear power plant, like yours truly. So they know that these are glorified steam engines to date. At a production site like this, we have various automatic monitoring and actuations. For example, the temperature of the core, or reactor, can be regulated mainly by cooling water flow (e.g., diverted river). If the core temperature increases, the automation will increase the water flow to cool it. More precisely, the temperature probe sends the signal to the pump, which will increase the volumetric water flow rate. If the temperature of the core decreases, the automation will reduce the water flow to avoid overcooling it and decrease the volumetric flow rate of steam, which will affect the speed of the turbine, and this speed will lower the energy output from the generator, depending also on the demand of the grid. In this case we have a series of “ifs” that bring out sensible behavior, from a data stream. But can we call the production site, with its set of probes, pumps, valves, engine room, an artificial intelligence? If we see the production site as an agent that optimizes the economic performance of the company according to the market environment (the demand for electricity), yes. But it really seems like a stretch to me.
Part of the readers of this blog have understood what statistical modeling means and when it is needed. If in statistical consulting I explain a phenomenon such as the propensity to buy for a company’s product, I will end up with a probability that I can turn into “buyer/non-buyer,” a prediction. I can notify a salesperson of this label for each incoming user, who will then proceed or not to the negotiation. I can create this model using the same technique they use in some games, namely a decision tree, although in that context it is called a behavioral tree. The tree contains automations that can be constructed in different ways, such as through the Gini index, which is also used in economics.
I think by now you may have thought that the quantity of automations leads to intelligence understood as quality. In fact, there are various branches that investigate this classification.
In our body, in fact, we have various chemical cycles, or automations: from the one for energy (Krebs) to the one for managing the by-products of the energy cycle (urea). So these cycles also have sub-cycles, like the nested cycles in computer science. Looking at them separately, you can’t talk about intelligence; at most, you can be amazed by their complexity and the poetry of their mechanism. But the same is true if you look at the cycles of a production site or a statistical model separately, even if they are much less complex than those of the human body. Statistical learning is even accused of making simple conditional logic through “if/else“, this is because of start-ups and companies that wanted to attract a few more chickens by inserting the magic word “artificial intelligence”.
Every day, millions of people get out of bed with a beep, take a bus, go to work, come home, have dinner, relax and go to sleep. They repeat this cycle almost every day, and each part of the day has its sub-cycles. Seen this way, could you indicate, with your index finger, the intelligence on those days?
Is there a spectrum where the amount of automation becomes intelligence? I don’t know, but if you are interested in filtering the artificial intelligence that can serve your company from the fluff, we can do it by getting to know each other in a free call of about 30 minutes.