Recklesscoding
Online
•
Upload & Sell: Off
|
 p.2 #18 · Feel like I’ve been living under a rock | |
Ripolini wrote:
Are you sure?
Online, including on YouTube, you can find just about anything: a lot of nonsense, narratives influenced by sponsors, and only a few pieces of accurate information. To distinguish clickbait nonsense from accurate information, you need knowledge—or at the very least, a well-developed critical sense. So, if intelligence—including that of a human—is formed on a statistical basis, meaning it weights everything it reads equally, the result is erratic, completely unreliable learning. Many falsehoods tend to be considered true simply because they’re repeated over and over; much of the information in the media—and especially online— is built on this principle.
So even humans, bombarded with misleading information, believe they’re learning when, in reality, they’re just filling their heads with nonsense....Show more →
Yes, in this particularly context in which the OP is using ChatGPT. There is usually some selection and data annotation done to the training data. So for something as simple as camera settings, eh, I would trust the output.
Now the slightly longer answer for the rest of your post:
You can find a plethora of context as you say; usually that context and if there are opposing views then they should balance themselves out. If not, then there is a trend towards one or the other view. This is where machine learning excels at BTW; classification tasks like this. LLMs do that in a way. On top of that, actually, those models also contain implicit biases. Those are more scary in my opinion as you don’t know that they are there. In general though, OpenAI and all ML deployers do model monitoring for drifting and incidents. Still, given the stochastic nature of those models, you need additional explainability to best understand and filter out. In my academic research, I worked a lot on this particular topic. I think, AI is not a replacement but complimentary of expertise. So with appropriate explanations, an expert can do fact checking and decide to trust or not the output. The extent to which this needed of course varies with the use case. Then again, I wouldn’t trust an LLM with any decision making beyond multimedia generation.
Human intelligence is more complicated than just statistical basis. We are not born as tabula rasa; we have hardwired behaviours. As we grow, we learn through A-B learning, i.e. trial and error, and quite often we consider knowledge acquired like that to be “common sense”. In addition, we have—as the late Daniel Dennett called it—a Gregorian mind, i.e. we import information from our society. This includes knowledge and opinions. We are, in fact, the only animal that does so to this extent and what sets us apart. This means that we, quite often, import those biased opinions. I recall my PhD supervisor worked on a project which demonstrated how even if people know a gossip is a lie, they will still listen to it and propagate it further. All in all, what I want to say is that humans are also —as you say— reliable sources of information.
One other big difference, between LLM—and machine learning at large—with us is that we have a semantic meaning behind each word. The word door means more than just a word most likely associated with the words wall or house. We developed mental models of concepts, whether physical or abstract (e.g. freedom), and those concepts contain meaning, associations, and help us identify them when encountered. This where another big part of AI, which is nowadays old fashioned, called symbolic AI comes into play. We are still far from proper large-scale neuro-symbolic approaches. What I am trying to say is: there is no single way to get an entity —biological or artificial— to act in an intelligent manner if all you want is action selection at the end of the day. Arguably, it doesn’t need to be a single way and you don’t need AI systems to behave exactly like humans.
|
so we humans still have some relevancy.