|
| Software Engineering Series |
With AI on the verge, blogging and bloggers can easily become obsolete. Why bother navigating through the many blogs to get a broader perspective when the same can be obtained with AI? Just type in a prompt of the type "write a blogpost of 600 words on the importance of AI in society" and Copilot or any other similar AI agent will provide you an answer that may look much better than the first draft of most of the bloggers out there! It doesn't matter whether the text follows a well-articulated idea, a personal perspective or something creative! One gets an acceptable answer with a minimum of effort and that's what matters for many.
The results tend to increase in complexity the more models are assembled together, respectively the more uncontrolled are the experiments. Moreover, solutions that tend to work aren't necessarily optimal. Machines can't offer instant enlightenment or anything close to it. Though they have an incomparable processing power of retrieval, association, aggregation, segregation and/or iteration, which coupled with the vast amount of data, information and knowledge can generate anything in just a matter of seconds. Probably, the only area in which humans can compete with machines is creativity and wisdom, though how many will be able to leverage these at scale? Probably, machines have some characteristics that can be associated with these intrinsic human characteristics, though usually more likely the brute computational power will prevail.
At Microsoft Build, Satya Nadella mentioned that foundry encompasses already more than 1900 supported models. In theory, one can still evaluate and test such models adequately. What will happen when the scale increases with a few orders of magnitude? What will happen when for each person there are one or more personalized AI models? AI can help in many areas by generating and evaluating rapidly many plausible alternatives, though as soon the models deal with some kind of processing randomization, the chances for errors increase exponentially (at least in theory).
It's enough for one or more hallucinations or other unexpected behavior to lead to more unexpected behavior. No matter how well a model was tested, as long as there's no stable predictable mathematical model behind it, the chances for something to go wrong increase with the number of inputs, parameters, uses, or changes of context the model deals with. Unfortunately, all these aspects are seldom documented. It's not like using a formula and you know that given a set of inputs and operations, the result is the same. The evolving nature of such models makes them unpredictable in the long term. Therefore, there must always be a way to observe the changes occurring in models.
One of the important questions is how many errors can we afford in such models? How long it takes until errors impact each other to create effects comparable with a tornado. And what if the tornado increases in magnitude to the degree that it wrecks everything that crosses its path? What if multiple tornadoes join forces? How many tornadoes can destroy a field, a country or a continent? How many or big must be the tornadoes to trigger a warning?
Science-Fiction authors love to create apocalyptic scenarios, and all happens in just a few steps, respectively chapters. In nature, usually it takes many orders of magnitude to generate unpredictable behavior. But, as nature often reveals, unpredictable behavior does happen, probably more often than we expect and wish for. The more we are poking the bear, the higher the chances for something unexpected to happen! Do we really want this? What will be the price we must pay for progress?
