There is truth in statistics. The minor errors are irrelevant in the actual LLM. Problems like the bad reddit quotes by google have nothing to do with and actual LLM, that is a RAG (augmented retrieval) and just bad standard code. The model itself is learning statistical word associations across millions of instances of similar data. The minor errors are irrelevant in this context.

Generative tools posted online are trash in their controls and especially the depth of capabilities. If you play with an enthusiast level consumer machine, with ComfyUI, the full nodes manager (not just the comfy anonymous repo), and the hundreds of nodes, things change. I’ve spent the last week reading white papers, following code examples, and trying new techniques. The possibilities are getting exponentially complex in a short period of time. I think most people working on generative AI in the public space are turning inward at the moment because it is hard to grasp all the possibilities, or maybe I’m just not following the right people.

We are in a data grab phase where it is feasible to collect more data as opposed to refining what exists. I think the techniques are growing too fast to say what will be the most efficient way of refining data. Eventually a refinement phase is likely.

Hallucinations are not actually a thing. The reasons they happen are just too complex to explain to a consumer public or no one would use the tool. If you learn about alignment and you really start reading into the tokenizer code, you’ll learn that it is just a complex system where most errors are due to safety alignment. The rest are generalizations made for an average use case. The underlying capability is far more complex and nuanced than any publicly hosted stalkerware data mining operation might appear. These real capabilities of the LLM are the building blocks of change. There are many other systems than just the tensor tables and word relationship statistics.