Raw LLM Responses

​​​@kerwynpkokay, let me explain why these AI models have no intelligence but instead mimic intelligence. The most simple definition of intelligence is the ability to aquire and apply knowledge. First let's setup some basic general definitions. Most traditional software systems are almost entirely deterministic. In these systems, the code is written to ensure a predictable outcome based on the input. When building a deterministic program, the developer explicitly implements every feature because they understand exactly how the internal logic should function to achieve the desired result. You can think of a deterministic program like a recipe or a list of steps that the program must follow to complete its task. Artificial Intelligence models, however, are stochastic. In complex scenarios, while the programmer knows the desired outcome, it is unreasonable to program the steps deterministically. Consider building a program to classify images of cats: for the deterministic approach you would have to manually write code to handle every possible variation of a cat (lighting, angle, fur color). If the system encounters data outside the specific set you manually programmed, it fails. It cannot dynamically handle the unexpected. With the stochastic approach Instead of writing precise instructions, you collect a massive dataset of examples. You then train a model to mimic the behavior of a deterministic program. Stochastic programs utilize probability by definition. This allows the program to identify complex patterns in data that are not readily apparent to a human programmer. However, the cost of this flexibility is variance: unlike a deterministic script, there is always a statistical chance of an unpredictable or "hallucinated" output. With those definitions established, we can evaluate AI against the standard definition of intelligence. Based on the mechanics of stochastic models, one could argue that current AI fails to meet this definition in two key areas; Failure to "Acquire" Knowledge and Failure to "Apply" Knowledge. A stochastic AI only truly acquires knowledge during its primary training phase (and perhaps during post-training fine-tuning). Once the model is compiled, that knowledge is frozen. You cannot simply keep training the model indefinitely. Eventually, you hit diminishing returns where further training causes the model to degrade or produce wild, undesirable outputs (often called "catastrophic forgetting"). AI interactions often feel like the model is learning because of context windows. This is where the system feeds a summary of your recent chat logs back into the model along with your new query. This simulates short-term memory, but it is not knowledge acquisition. The system does not retain this information long-term. For the AI to actually "know" this data, a completely new model would have to be trained with that data included in the base set. Finally, regarding the application of knowledge, the AI falls short because it isn't engaging in reasoning. It is not taking learned concepts and applying them to new problems. Instead, it is performing a complex statistical calculation—using a set of probabilities to guess which piece of data (or token) is most likely to come next in the sequence. Now with all that being said, there are some methods for a continuous learning in ai, but they are not being used by the mainstream and most of them are still in their infancy as far as development goes. But in reality this is just one part of the definition of intelligence. The large companies that are throwing money at AI are betting that somehow these programs will eventually fill out that criteria by becoming so good at mimicking that they are indistinguishable from the real thing.