Comment on How come no true use for recent AI developments has been found yet?
krashmo@lemmy.world 3 weeks ago
Current gen AI is pretty mediocre. It’s not much more than the bastard child of a search engine and every voice assistant that has been around for the last ten years. It has the potential to be a stepping stone to fantastic future tech, but that’s been true of tons of different technologies for basically as long as we’ve been inventing things.
AI is not good enough to replace the majority of workers yet. It summarizes information pretty well and can be helpful with drafting any sort of document, but so was Clippy. When it doesn’t know something it can lie confidently. Lie isn’t really the right word but I’ll come back to that concept in a second. Incorrect information is frustrating in most cases but it can be deadly when presented by a source that is viewed as trustworthy, and what could be more trustworthy than an AI with access to the collective knowledge of mankind? Well, unfortunately for us AI as we know it isn’t really intelligent and the databases they’re trained on also contain the collective stupidity of mankind.
That brings us back to the concept of lying and what I view as the fundamental flaw of current AI; namely that any sort of data interpretation can only be as good as the data it describes. ChatGPT isn’t lying to you when it says you can put glue on your cheese pizza, it’s just pointing out that someone who said that got a lot of attention. Unfortunately it leaves out all the context which could have told you that pizza would not be fit to consume and presents the fact that it was a popular answer as if that is the only thing that defines the best answer. There’s so much more that needs to be taken into account, so much unconscious human experience being drawn from when an actual human looks at something and tries to categorize or describe it. All of that necessary context is really difficult to impart to a computer and right now we’re not very good at that essential piece of the puzzle.
If we could assume that all datasets analyzed by AI were free from human error, AI would be taking over the word right now. However, that’s not the world we live in. All data has errors. Some are easy to spot but many are not. AI firms are getting companies to salivate at the idea of easy manipulation of data in one form or another. They aren’t worried about the errors in the data because they view that as someone else’s problem and the companies all think their data is good enough that it won’t be an issue. Both are wrong. That’s exactly why you hear a lot of talk about AI right now and not all that much practical application beyond replacing customer service reps, especially in the business world. Companies are finding out that years of bad practices have left them with a dataset full of errors. Can they find a way to get AI to correct those errors? In some cases yes, in others no. In either case the missing piece preventing a full scale AI takeover is all that human background context necessary for relevant data interpretation. If we find a way to teach that to an AI then the world is going to look vastly different than it does today, but we’re not there yet.
j4k3@lemmy.world 3 weeks ago
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.