3 ms·

You've got to be careful about which LLM. I asked Microsoft Copilot for advice about staying in Paris, and then as a followup in the same thread, asked about na

You've got to be careful about which LLM. I asked Microsoft Copilot for advice about staying in Paris, and then as a followup in the same thread, asked about navigation apps. It started recommending apps that work well... in New York City. If Copilot can't retain context, in the same thread, after a single message, then it's useless. ChatGPT, which apparently Copilot is based on, didn't have this problem. And from the linked article I have just discovered "Perplexity", which looks a bit like ChatGPT only plugged into a search engine. ChatGPT is also plugged into a search engine, so maybe it's plugged in... more? Not sure. If anyone working at Perplexity is reading this, please add an "about" page to tell me what the site actually does. Nowhere I could find would explain. Everything assumes you already know, and I eventually had to go to Wikipedia to find out. If anyone makes an LLM which only knows what's in Wikipedia, and doesn't hallucinate anything else, they would clean up.

smackeyacky 2y ago
Why not just use the search bar in Wikipedia and call it good?
- LeoPanthera 2y ago
  Because "searching for an encyclopedia article" and "asking questions about a branch of knowledge" are fundamentally different approaches. In theory, an LLM that was a talking Wikipedia would not be limited to giving you knowledge from just one page, it could extract multiple pieces of information from multiple articles to answer the precise question you had, and could accept follow-up questions on the same or similar topics.
  - card_zero 2y ago
    It's a good idea, but it might turn out to hallucinate even more. For instance, there is one article on Abraham Lincoln and one on Abraham Lincoln: Vampire Hunter, so the statistics (50/50) in regard of mixing fantasy with reality are perhaps discouraging.
readyman 2y ago
[dead]
wslh 2y ago
If you know about the subject LLMs makes you turbo. Even when you don't know about the subject a rational thinking background gives you more legs. LLMs are incredible tools, it is expected that they are not the only ones in your toolbox. When I take a random photo and asks ChatGPT to describe it I always wonder how I saved a whole lifetime (or many ones) dedicated to computer vision research and can build upon all of this.
Zambyte 2y ago
> ChatGPT is also plugged into a search engine, so maybe it's plugged in... more? Not sure. I haven't used ChatGPT search, but I have used Perplexity. The way that it works, it will potentially do like a dozen searches to try to understand your question (if it needs to), search individual topics in the question, search joint topics in the question, and then take 20 results or so, including web pages and videos, and synthesize an answer to your question. > If anyone makes an LLM which only knows what's in Wikipedia, and doesn't hallucinate anything else, they would clean up. Kagi has a "lens" feature that lets you scope down web results to certain sources. It's supposed to be able to be used in their AI assistant, but it's currently ignored / broken there. They are working on a large overhaul (I think it's supposed to be out by the end of this month?), hopefully that works after that. You could achieve what you described using that then.

smackeyacky 2y ago

Why not just use the search bar in Wikipedia and call it good?

LeoPanthera 2y ago
Because "searching for an encyclopedia article" and "asking questions about a branch of knowledge" are fundamentally different approaches. In theory, an LLM that was a talking Wikipedia would not be limited to giving you knowledge from just one page, it could extract multiple pieces of information from multiple articles to answer the precise question you had, and could accept follow-up questions on the same or similar topics.
- card_zero 2y ago
  It's a good idea, but it might turn out to hallucinate even more. For instance, there is one article on Abraham Lincoln and one on Abraham Lincoln: Vampire Hunter, so the statistics (50/50) in regard of mixing fantasy with reality are perhaps discouraging.

readyman 2y ago

[dead]

wslh 2y ago

If you know about the subject LLMs makes you turbo. Even when you don't know about the subject a rational thinking background gives you more legs. LLMs are incredible tools, it is expected that they are not the only ones in your toolbox. When I take a random photo and asks ChatGPT to describe it I always wonder how I saved a whole lifetime (or many ones) dedicated to computer vision research and can build upon all of this.

Zambyte 2y ago

> ChatGPT is also plugged into a search engine, so maybe it's plugged in... more? Not sure. I haven't used ChatGPT search, but I have used Perplexity. The way that it works, it will potentially do like a dozen searches to try to understand your question (if it needs to), search individual topics in the question, search joint topics in the question, and then take 20 results or so, including web pages and videos, and synthesize an answer to your question. > If anyone makes an LLM which only knows what's in Wikipedia, and doesn't hallucinate anything else, they would clean up. Kagi has a "lens" feature that lets you scope down web results to certain sources. It's supposed to be able to be used in their AI assistant, but it's currently ignored / broken there. They are working on a large overhaul (I think it's supposed to be out by the end of this month?), hopefully that works after that. You could achieve what you described using that then.