Due to the nature of our work, my firm has had early access to most LLMs including Bard (now Gemini). I might be short on imagination but I honestly cannot see how LLM general search implementations can ever be fixed. There is too much garbage data for any system to be able to intelligently parse and the results of our tests were laughable. Now, if you offer LLM search that is restricted to curated datasets like “The Library of Congress” or peer reviewed scientific papers, I can see the value in that. You’ll probably still have to triple check your results, but at least it can get you 80% of the way there rather than sending you in the wrong direction.
EDIT: For context, our clientele are all enterprises with very large, mission critical systems. They are not the type to use some buggy trinket just because it’s new and cool.
yes, I find Gemini actually not bad when it comes to my specific use case of showing generic examples for R programming, so I can figure out the syntax for my actual code. I don’t try to have it generate actual code for me because my topic of marine biogeochemistry is far too specific for it to have any idea how to work with it. Unlike ChatGPT, which often makes up nonsense functions or hallucinates whole packages, Gemini seems to do ok. I also found it pretty good for generating images of natural subjects. It did the best job of generating a pic of a giant clam of any image generator I’ve tried. I would never trust factual information from Gemini. So like Google+, it’s a pretty good product that in no way should be shunted into search results, Google Docs and other places where its output is not relevant, yet that is exactly the trap Google is falling into again.
The article is too long for me. 2 of its main ideas are “Everyone using large-language models should be aware of ai hallucination and be careful when asking those models for facts.” and “Firms that develop large-language models shouldn’t downplay the hallucination and shouldn’t force ai in every corner of tech.”
There was already so much misinformation on the Web before Chatgpt 3.5. There’s still so much misinformation. No need for the hallucination to worsen the situation. We need a reliable source of facts. Optimistically, Google, Openai or Anthropic will find a way to reduce or eradicate the hallucination. The Google ceo said they were making progress. Maybe true. Or maybe generic pr lie so folks would stop following up re the hallucination.
You are not logged in. However you can subscribe from another Fediverse account, for example Lemmy or Mastodon. To do this, paste the following into the search field of your instance: !technology@lemmy.world
This is a most excellent place for technology news and articles.
deleted by creator
Due to the nature of our work, my firm has had early access to most LLMs including Bard (now Gemini). I might be short on imagination but I honestly cannot see how LLM general search implementations can ever be fixed. There is too much garbage data for any system to be able to intelligently parse and the results of our tests were laughable. Now, if you offer LLM search that is restricted to curated datasets like “The Library of Congress” or peer reviewed scientific papers, I can see the value in that. You’ll probably still have to triple check your results, but at least it can get you 80% of the way there rather than sending you in the wrong direction.
EDIT: For context, our clientele are all enterprises with very large, mission critical systems. They are not the type to use some buggy trinket just because it’s new and cool.
yes, I find Gemini actually not bad when it comes to my specific use case of showing generic examples for R programming, so I can figure out the syntax for my actual code. I don’t try to have it generate actual code for me because my topic of marine biogeochemistry is far too specific for it to have any idea how to work with it. Unlike ChatGPT, which often makes up nonsense functions or hallucinates whole packages, Gemini seems to do ok. I also found it pretty good for generating images of natural subjects. It did the best job of generating a pic of a giant clam of any image generator I’ve tried. I would never trust factual information from Gemini. So like Google+, it’s a pretty good product that in no way should be shunted into search results, Google Docs and other places where its output is not relevant, yet that is exactly the trap Google is falling into again.
Every time. It’d be funny if it didn’t mean people constantly being punished and losing their jobs for errors made at the executive level.
The article is too long for me. 2 of its main ideas are “Everyone using large-language models should be aware of ai hallucination and be careful when asking those models for facts.” and “Firms that develop large-language models shouldn’t downplay the hallucination and shouldn’t force ai in every corner of tech.”
There was already so much misinformation on the Web before Chatgpt 3.5. There’s still so much misinformation. No need for the hallucination to worsen the situation. We need a reliable source of facts. Optimistically, Google, Openai or Anthropic will find a way to reduce or eradicate the hallucination. The Google ceo said they were making progress. Maybe true. Or maybe generic pr lie so folks would stop following up re the hallucination.
removed by mod