From my experience, I believe there are several limitations. Firstly, it requires a lot of data and resources to function correctly, and if that data contains errors or biases, it adopts them. I think it works best on specific tasks, but it struggles when something new or different arises. We've all experienced its ability to make simple mistakes that would be obvious to a human, and no matter how hard we try, it often still deceives us, albeit in a way that seems so realistic it sometimes even influences our approach. I think we trust AI too much, and we need to consider it for what it is: a support tool.
Even though AI delivers impressive results today, I see clear limits from my point of view. First, these systems lack a real understanding. They work with patterns and probabilities, not with meaning in the human sense. Second, they are highly dependent on their training data. If these are incomplete or skewed, this is directly reflected in the results. Third, AI finds it difficult to transfer knowledge flexibly – it usually remains trapped in its area of application. Another problem is the susceptibility to errors: systems can make convincingly false statements without recognizing it themselves. In addition, many models are difficult to understand; you often don't know exactly why a decision is made. From a didactic point of view, this is particularly problematic because transparency is actually important. Finally, one should not underestimate the high technical effort – both in terms of computing power and energy consumption. Overall, I would say that AI is a powerful tool, but it is not a substitute for human thinking and judgment.
You can look at it this way: Technical: expensive, data-hungry, limited stability Functional: impressive, but error-prone and contextual Conceptually: no real thinking, no awareness, no understanding The key point: Today's AI is an extremely powerful tool for pattern processing – but not a thinking being.
There are a bunch of problems born from the fact that we have approximated Moore’s law’s end, and several theories are being tested right now in an attempt to bend physics. Namely, fit more transistors in a smaller space (we want to print 1 nanometer transistors, that is, 5 silicon atoms long), use light instead of electricity to transmit information between them or harness quantum theory to shift the computing paradigms. We also have the energy problems associated with running this infrastructure, and we are seeing that some data centers are building their own energy plants for running, which speaks of how inefficient we are today at computing. Conceptually, I don’t think we are so stale as new and improved ways of using AI are coming on a regular basis, but the concepts fall flat when we're so bound by the current tech limitations.
Learning datasets are flawed witch often makes unreliable answers, the constant limitations of what it can and can't say whether it's for the "woke" people or limiting dangerous information really hurts performance in my opinion, In fact many people have said that chat gpt used to be better and faster in the early days because it hardly had any limits on what it could say but it got neutered because people complained
Hallucinations, dependency of answer on access to the right data e.g. used in the training set or accessible via search engines and internal knowledge bases if the model. Model’s outputs are very long and sound beautiful, but often lack substance.
Main limitations are that AI is not well integrated with everyday used stuff like Excel, PowerPoint.
