4 ms·

With RAG it's entirely possible to essentially eliminate 100% of hallucinations, given you are ok with responding with "I don't know" once in a while. These sit

With RAG it's entirely possible to essentially eliminate 100% of hallucinations, given you are ok with responding with "I don't know" once in a while. These situations are likely coming from poorly implemented chatbot, or they decided that "I dont know" was not acceptable, and really that should be a queue to send you to a real human.

jerpint 3y ago
There are no guarantees with RAG either, and RAG only works when the answer to the question is already printed out somewhere explicitly in the text, otherwise it’s definitely prone to hallucinate
- alexxys 3y ago
  Yeah, RAG can't provide such guarantees. Moreover, even if the correct answer is printed somewhere, LLM+RAG still may produce wrong answer. Example from MS Copilot with GPT-4: https://sl.bing.net/ct6wwRjzkPc https://sl.bing.net/ct6wwRjzkPc It claims that OnePlus 6 has 6.4-inch display, but all linked pages actually claim that it's 6.28. Display resolution and aspect ratio are also wrong in the response.
  - steveBK123 3y ago
    It's funny it seems to have a lot of trouble extracting tabular data, which arguably is one of the things I hear people trying to do with it..
sebastiennight 3y ago
This claim seems wildly inaccurate, as even with GPT-4 in a single conversation thread with previous human-written answers included, a repeat of a similar question just resulted - in my testing today - in a completely hallucinated answer. I think your claim might be based on anecdotal testing. (I used to have that same feeling after my first implementation of RAG)... Once you get a few thousand users running RAG-based conversations, you quickly see that it's "good enough to be useful", but far from being as dreamy as promised.
fivre 3y ago
do the people managing the chatbot know that though? this shit gets sold as a way to replace employees with, essentially, just the middle manager that was over them, who is now responsible for managing the chatbot instead of managing people while managers are often actually not great at people management, it's at least a somewhat intuitive skill for many. interacting with and directing other humans is something that many people are able to gain experience with outside of work, since it's a necessary life skill unless you're a hermit. furthermore, as a hedge against managerial ineptitude, humans are adaptable creatures that can recognize their manager's shortcomings and determine when and how to work around them to actually get the job done understanding the intricacies training a machine learning system is a highly specialized and technical skill that nobody is going to pick up base knowledge for in the regular course of life. the skill floor for the average person tasked with it will be much lower than that of people management, and they will probably fuck up, a lot the onus is ostensibly on AI system vendors to make their systems idiot-proof, but how many vendors actually do so past the point of "looks good enough to close the sale in a demo"? designing such a system is _incredibly_ hard, and the unfortunate reality is that if you try, you'll lose sales to snake oil salesmen who are content to push hokum trash with a fancy coat of paint. these systems can work as a force multiplier in the hands of the capable, but work as an incompetence magnifier in the hands of the incapable, and there are plenty of dunning-krugerites lusting to magnify their incompetence
- Piskvorrr 3y ago
  Well if incompetence is cheaper to implement, out come the lawyers to erase whatever perceived savings there were.

jerpint 3y ago

There are no guarantees with RAG either, and RAG only works when the answer to the question is already printed out somewhere explicitly in the text, otherwise it’s definitely prone to hallucinate

alexxys 3y ago
Yeah, RAG can't provide such guarantees. Moreover, even if the correct answer is printed somewhere, LLM+RAG still may produce wrong answer. Example from MS Copilot with GPT-4: https://sl.bing.net/ct6wwRjzkPc https://sl.bing.net/ct6wwRjzkPc It claims that OnePlus 6 has 6.4-inch display, but all linked pages actually claim that it's 6.28. Display resolution and aspect ratio are also wrong in the response.
- steveBK123 3y ago
  It's funny it seems to have a lot of trouble extracting tabular data, which arguably is one of the things I hear people trying to do with it..

sebastiennight 3y ago

This claim seems wildly inaccurate, as even with GPT-4 in a single conversation thread with previous human-written answers included, a repeat of a similar question just resulted - in my testing today - in a completely hallucinated answer. I think your claim might be based on anecdotal testing. (I used to have that same feeling after my first implementation of RAG)... Once you get a few thousand users running RAG-based conversations, you quickly see that it's "good enough to be useful", but far from being as dreamy as promised.

fivre 3y ago

do the people managing the chatbot know that though? this shit gets sold as a way to replace employees with, essentially, just the middle manager that was over them, who is now responsible for managing the chatbot instead of managing people while managers are often actually not great at people management, it's at least a somewhat intuitive skill for many. interacting with and directing other humans is something that many people are able to gain experience with outside of work, since it's a necessary life skill unless you're a hermit. furthermore, as a hedge against managerial ineptitude, humans are adaptable creatures that can recognize their manager's shortcomings and determine when and how to work around them to actually get the job done understanding the intricacies training a machine learning system is a highly specialized and technical skill that nobody is going to pick up base knowledge for in the regular course of life. the skill floor for the average person tasked with it will be much lower than that of people management, and they will probably fuck up, a lot the onus is ostensibly on AI system vendors to make their systems idiot-proof, but how many vendors actually do so past the point of "looks good enough to close the sale in a demo"? designing such a system is _incredibly_ hard, and the unfortunate reality is that if you try, you'll lose sales to snake oil salesmen who are content to push hokum trash with a fancy coat of paint. these systems can work as a force multiplier in the hands of the capable, but work as an incompetence magnifier in the hands of the incapable, and there are plenty of dunning-krugerites lusting to magnify their incompetence

Piskvorrr 3y ago
Well if incompetence is cheaper to implement, out come the lawyers to erase whatever perceived savings there were.