I suppose it's all a matter of what one is using an LLM for, no?
GPT is great at citing sources for most of my requests -- even if not always prompted to do so. So, in a way, I kind of use LLMs as a search engine/Wikipedia hybrid (used to follow links on Wiki a lot too). I ask it what I want, ask for sources if none are provided, and just follow the sources to verify information. I just prefer the natural language interface over search engines. Plus, results are not cluttered with SEO ads and clickbait rubbish.
Hmm I don't feel like this should be taken as a tenet of AI. I feel a more relevant kernel would be less black and white.
Also I think what you're saying is a direct contradiction of the parent. Below average people can now get average results; in other words: The LLM will boost your capabilities (at least if you're already 'less' capable than average). This is a huge benefit if you are in that camp.
But for other cases too, all you need to know is where your knowledge ends, and that you can't just blindly accept what the AI responds with.
In fact, I find LLMs are often most useful precisely when you don’t know the answer. When you’re trying to fill in conceptual gaps and explore an idea.
Even say during code generation, where you might not fully grasp what’s produced, you can treat the model like pair programming and ask it follow-up questions and dig into what each part does. They're very good at converting "nebulous concept description" into "legitimate standard keyword" so that you can go and find out about said concept that you're unfamiliar with.
Realistically the only time I feel I know more than the LLM is when I am working on something that I am explicitly an expert in, and in which case often find that LLMs provide nuance lacking suggestions that don’t always add much. It takes a lot more filling in context in these situations for it to be beneficial (but still can be).
Take a random example of nifty bit of engineering: The powerline ethernet adapter. A curious person might encounter these and wonder how they work. I don't believe an understanding of this technology is very obvious to a layman. Start asking questions and you very quickly come to understand how it embeds bits in the very same signal that transmits power through your house without any interference between the two "types" of signal. It adds data to high frequencies on one end, and filters out the regular power transmitting frequencies at the other end so that the signal can be converted back into bits for use in the ethernet cable (for a super brief summary). But if want to really drill into each and every engineering concept, all I need to do is continue the conversation.
I personally find this loop to be unlike anything I've experienced as far as getting immediate access to an understanding and supplementary material for the exact thing Im wondering about.
Above average people can also use it to get average results. Which can actually be useful. For many tasks and usecases, the good enough threshold can actually be quite low.
A broken clock is right twice a day. If you're spending more time writing instructions for the AI, and then rewriting those instructions several times before you get usable code, then you're absolutely wasting your time. At least this is my experience with "AI". Even the fancy autocomplete portion of "AI" gets in my way more than it helps. Of course, YMMV, but I seriously doubt you're doing any better with "AI" than anyone else actually is, and a lot of people using it just don't realize all the time it's wasting, if they paid attention and added it all up.
This is in line with another quip about AI: You need to know more than the LLM in order to gain any benefit from it.
I am not certain that is entirely true.
I suppose it's all a matter of what one is using an LLM for, no?
GPT is great at citing sources for most of my requests -- even if not always prompted to do so. So, in a way, I kind of use LLMs as a search engine/Wikipedia hybrid (used to follow links on Wiki a lot too). I ask it what I want, ask for sources if none are provided, and just follow the sources to verify information. I just prefer the natural language interface over search engines. Plus, results are not cluttered with SEO ads and clickbait rubbish.
Hmm I don't feel like this should be taken as a tenet of AI. I feel a more relevant kernel would be less black and white.
Also I think what you're saying is a direct contradiction of the parent. Below average people can now get average results; in other words: The LLM will boost your capabilities (at least if you're already 'less' capable than average). This is a huge benefit if you are in that camp.
But for other cases too, all you need to know is where your knowledge ends, and that you can't just blindly accept what the AI responds with. In fact, I find LLMs are often most useful precisely when you don’t know the answer. When you’re trying to fill in conceptual gaps and explore an idea.
Even say during code generation, where you might not fully grasp what’s produced, you can treat the model like pair programming and ask it follow-up questions and dig into what each part does. They're very good at converting "nebulous concept description" into "legitimate standard keyword" so that you can go and find out about said concept that you're unfamiliar with.
Realistically the only time I feel I know more than the LLM is when I am working on something that I am explicitly an expert in, and in which case often find that LLMs provide nuance lacking suggestions that don’t always add much. It takes a lot more filling in context in these situations for it to be beneficial (but still can be).
Take a random example of nifty bit of engineering: The powerline ethernet adapter. A curious person might encounter these and wonder how they work. I don't believe an understanding of this technology is very obvious to a layman. Start asking questions and you very quickly come to understand how it embeds bits in the very same signal that transmits power through your house without any interference between the two "types" of signal. It adds data to high frequencies on one end, and filters out the regular power transmitting frequencies at the other end so that the signal can be converted back into bits for use in the ethernet cable (for a super brief summary). But if want to really drill into each and every engineering concept, all I need to do is continue the conversation.
I personally find this loop to be unlike anything I've experienced as far as getting immediate access to an understanding and supplementary material for the exact thing Im wondering about.
Above average people can also use it to get average results. Which can actually be useful. For many tasks and usecases, the good enough threshold can actually be quite low.
That explains why people here are against it, because everyone is above average I guess.
I'm not against it. I wonder where in the distribution it puts me.
At the "Someone willing to waste their time with slop" end?
Seeing that it gives me code that I then then use, then it's not a waste of my time.
A broken clock is right twice a day. If you're spending more time writing instructions for the AI, and then rewriting those instructions several times before you get usable code, then you're absolutely wasting your time. At least this is my experience with "AI". Even the fancy autocomplete portion of "AI" gets in my way more than it helps. Of course, YMMV, but I seriously doubt you're doing any better with "AI" than anyone else actually is, and a lot of people using it just don't realize all the time it's wasting, if they paid attention and added it all up.
>Below average people can use AI to get average results.
But that would shift the average up.