I think the issue is most people don’t understand what an LLM is doing. It’s not thinking about your question and finding the right answer. It is just doing a bunch of math to calculate the most probable response based on all it’s training plus or minus some minor random variation. If your question could be answered by a thorough Google search then an llm can probably give you a good answer. If it’s about something you’re not going to find on the internet then the LLM will just make up something that sounds convincing. And that’s the problem. It may sound convincing but it’s a con man.
I think the issue is most people don’t understand what an LLM is doing.
It’s engineered to respond conversationally, not diagetically. You’re not supposed to know how it is processing your inputs. A black box by design.
And I get the knee-jerk impulse to see this as a search engine “stealing” click-throughs to strangle a site. But I think people often forget why click-throughs are so vital for a site’s existence to begin with. Why do you need to send your browser through the front door for the data? What does the site get out of that interaction that it doesn’t get out of an LLM fetching the data for you?
If it’s about something you’re not going to find on the internet then the LLM will just make up something that sounds convincing. And that’s the problem. It may sound convincing but it’s a con man.
You can see this with known hiccups in the model. “How many 'r’s are in strawberry?” is a classic (that was eventually fixed) that hinges on a question you’re not going to find pre-answered in the publicly scrapped data.
But if you’re fetching data from an existing website then this isn’t the problem you’re going to run into. It’s going to fetch the data accurately (baring some other glitch) and return a summary of contents modeled into whatever output template the LLM prefers.
The con-man aspect really kicks in as part of the templating of the answer. You could engineer an LLM to respond in the negative when it failed to find useful data. Or you could engineer it to return data as spreadsheets with “Description” and “Source” (not unlike how Google normally returns links and short blurbs) that allow you to interrogate the source info easily. Instead, we get a definitive pronouncement made with links buried at the bottom of the extended blurb.
I think the issue is most people don’t understand what an LLM is doing. It’s not thinking about your question and finding the right answer. It is just doing a bunch of math to calculate the most probable response based on all it’s training plus or minus some minor random variation. If your question could be answered by a thorough Google search then an llm can probably give you a good answer. If it’s about something you’re not going to find on the internet then the LLM will just make up something that sounds convincing. And that’s the problem. It may sound convincing but it’s a con man.
It’s engineered to respond conversationally, not diagetically. You’re not supposed to know how it is processing your inputs. A black box by design.
And I get the knee-jerk impulse to see this as a search engine “stealing” click-throughs to strangle a site. But I think people often forget why click-throughs are so vital for a site’s existence to begin with. Why do you need to send your browser through the front door for the data? What does the site get out of that interaction that it doesn’t get out of an LLM fetching the data for you?
You can see this with known hiccups in the model. “How many 'r’s are in strawberry?” is a classic (that was eventually fixed) that hinges on a question you’re not going to find pre-answered in the publicly scrapped data.
But if you’re fetching data from an existing website then this isn’t the problem you’re going to run into. It’s going to fetch the data accurately (baring some other glitch) and return a summary of contents modeled into whatever output template the LLM prefers.
The con-man aspect really kicks in as part of the templating of the answer. You could engineer an LLM to respond in the negative when it failed to find useful data. Or you could engineer it to return data as spreadsheets with “Description” and “Source” (not unlike how Google normally returns links and short blurbs) that allow you to interrogate the source info easily. Instead, we get a definitive pronouncement made with links buried at the bottom of the extended blurb.