“I only ask questions to LLMs if I already know the answer”
Not a developer here. I’ve been thinking about this because I hoped LLMs would be able to help me learn things at first, like a patient tutor I can ask all my stupid questions and it’ll never get annoyed with me. Since it can’t do that though, because it lies all the time, I don’t think I have a use case for it at all… About all it can do for me is rewrite or summarise English, and it doesn’t even do a particularly good job of that most of the time so I end up saving time by doing that work myself anyway. I suppose it’s pretty good at translating, but I haven’t tried it for that as I don’t have a lot of call to speak foreign languages.
I hoped LLMs would be able to help me learn things at first, like a patient tutor I can ask all my stupid questions and it’ll never get annoyed with me
It can do that for school level stuff because that material is present in it’s input dataset in a redundant manner. For anything niche or domain-specific, it will hallucinate or fail.
I believe that when the bubble bursts, education will be one genuine usecase for LLMs.
It can do that for school level stuff because that material is present in it’s input dataset in a redundant manner. For anything niche or domain-specific, it will hallucinate or fail.
I typically don’t have an issue getting a grasp on fundamentals, so most of the things I want to ask it about might be beyond school-level. My main way of learning is to ask questions to make sure I understand the material - which means more potential hallucination points, and maybe worse impact because I’ll think I get it, but I’ve just been confidently lied to that I understood.
For example, I’ve wondered for a while if patches of space with less gravitational curvature “age” faster than patches that are more heavily distorted by gravity wells, and what the implications of that might be. Makes sense, we know that gravity slows down subjective time. But I can’t get a productive answer out of an LLM because I can’t trust it, and it’s not worth bothering my physicist friends about.
What worked reasonably well for me is “deep research”.
Even when summarizing the content of the sources, it fucks up and says something factually incorrect that the source never stated. But it is pretty good at understanding what you are looking for and giving a list of relevant links to read. Somehow google is worse with that and gives many irrelevant results.
Today for example, I tried using google to find a lithium battery of specific size and capacity and the results were not what I was looking for. Since I didn’t know the technical terms I just explained it in colloquial english “regular smartphone battery, but a footprint of 5x5cm” and it converted it to good search terms “single cell 3.7V lipo 105050/125050/135050” (which I then also used in my own search) and also provided a bunch of links to webshops that had those exact batteries!
I would have never known that I need to search for the conventional numerals that describe it’s size or that there is a difference with the charging controller between single cell and multi cell. (It gave me a source to that claim and I used that to read up on the matter)
Just googling my query only returned random non 5x5cm batteries, some multicell, some videocorder batteries etc. without any link to the difference or explanation of the tech terms or standards.
Ah-ha, now there’s an interesting use case. I’ve had occasion in the past to work on PCB design - to be clear I didn’t do the design myself, I have no idea wtf I’m doing there, but I’m capable of reading spec sheets and soldering - and when I had to sub in components, I managed to find what I needed by filtering down for specs on DigiKey using their search. I think an LLM could have saved me a bit of time there if I could’ve just fed it the BOM and asked for alternatives, and over the course of all the subs I had to do for the particular project I’m thinking of (it was during Chinese New Year so it was tough getting answers from suppliers) that would have added up.
Not a developer here. I’ve been thinking about this because I hoped LLMs would be able to help me learn things at first, like a patient tutor I can ask all my stupid questions and it’ll never get annoyed with me. Since it can’t do that though, because it lies all the time, I don’t think I have a use case for it at all… About all it can do for me is rewrite or summarise English, and it doesn’t even do a particularly good job of that most of the time so I end up saving time by doing that work myself anyway. I suppose it’s pretty good at translating, but I haven’t tried it for that as I don’t have a lot of call to speak foreign languages.
It can do that for school level stuff because that material is present in it’s input dataset in a redundant manner. For anything niche or domain-specific, it will hallucinate or fail.
I believe that when the bubble bursts, education will be one genuine usecase for LLMs.
I typically don’t have an issue getting a grasp on fundamentals, so most of the things I want to ask it about might be beyond school-level. My main way of learning is to ask questions to make sure I understand the material - which means more potential hallucination points, and maybe worse impact because I’ll think I get it, but I’ve just been confidently lied to that I understood.
For example, I’ve wondered for a while if patches of space with less gravitational curvature “age” faster than patches that are more heavily distorted by gravity wells, and what the implications of that might be. Makes sense, we know that gravity slows down subjective time. But I can’t get a productive answer out of an LLM because I can’t trust it, and it’s not worth bothering my physicist friends about.
What worked reasonably well for me is “deep research”.
Even when summarizing the content of the sources, it fucks up and says something factually incorrect that the source never stated. But it is pretty good at understanding what you are looking for and giving a list of relevant links to read. Somehow google is worse with that and gives many irrelevant results.
Today for example, I tried using google to find a lithium battery of specific size and capacity and the results were not what I was looking for. Since I didn’t know the technical terms I just explained it in colloquial english “regular smartphone battery, but a footprint of 5x5cm” and it converted it to good search terms “single cell 3.7V lipo 105050/125050/135050” (which I then also used in my own search) and also provided a bunch of links to webshops that had those exact batteries!
I would have never known that I need to search for the conventional numerals that describe it’s size or that there is a difference with the charging controller between single cell and multi cell. (It gave me a source to that claim and I used that to read up on the matter)
Just googling my query only returned random non 5x5cm batteries, some multicell, some videocorder batteries etc. without any link to the difference or explanation of the tech terms or standards.
Ah-ha, now there’s an interesting use case. I’ve had occasion in the past to work on PCB design - to be clear I didn’t do the design myself, I have no idea wtf I’m doing there, but I’m capable of reading spec sheets and soldering - and when I had to sub in components, I managed to find what I needed by filtering down for specs on DigiKey using their search. I think an LLM could have saved me a bit of time there if I could’ve just fed it the BOM and asked for alternatives, and over the course of all the subs I had to do for the particular project I’m thinking of (it was during Chinese New Year so it was tough getting answers from suppliers) that would have added up.