2
$\begingroup$

For quite some time now, I have been evaluating ChatGPT's capability to deliver accurate and helpful responses. While its performance is undeniably impressive, the issue of hallucinations poses a significant drawback to this otherwise capable model.

As I increased the complexity of my inputs during several sessions, my initial enthusiasm and excitement about the advancements in AI quickly waned due to the responses becoming riddled with hallucinations. This has seriously raised my concerns about the reliability and usefulness of chatGPT.

I have two questions:

  1. Is there a declining interest among people in LLMs and artificial intelligence due to the problem of hallucinations?
  2. How are LLM companies addressing and combatting this concerning aspect of AI models in general?
$\endgroup$
5
  • 1
    $\begingroup$ The answer to question 1 is I think a clear no. You can get useful output from LLMs without hallucinations being a problem. Some people are sold in it even in spite of them. Question 2 is much more interesting. $\endgroup$ Jul 20, 2023 at 15:14
  • $\begingroup$ agreed, I can generate any isekai with any plot whatsoever that I enjoy because of (1) $\endgroup$
    – Sanyou
    Jul 20, 2023 at 15:43
  • $\begingroup$ I've expanded my comment into an answer though I expect and hope someone else provides a more informed one. $\endgroup$ Jul 20, 2023 at 15:55
  • $\begingroup$ You should know that hallucination has always been an issue with text generation. For example, machine translation or summarization models sometimes produce text that is fluent but irrelevant to the source text. And such models are still widely used despite this. $\endgroup$ Jul 21, 2023 at 10:50
  • $\begingroup$ No, to my best understanding we should understand them better only. Hallucinations, same as any other answers, simply cannot be anything else than something the AI concluded via it's model. $\endgroup$
    – harism
    Jul 21, 2023 at 21:41

1 Answer 1

1
$\begingroup$

Hallucinations will not lead to a declining interest

The answer to question 1 is I think a clear no. You can get useful output from LLMs without the hallucinations being a problem.

If you are not an expect on a subject the LLMs output can be educational

Also some people are using the output as the basis for their own work without using it verbatim.

There are examples of people using it to provide the outline of an email or even a story which the user then completes.

In the world of software it can be used as a better intellisense even though when asked to produce code it can produce code that is at best non-optimal and at times dangerously wrong.

Unfortunately of course there will be people who don't use models with sufficient dilligence and will try to use them to exploit opportunities such as mass marketing or phishing and some people will be sold even in spite of of glaring problems.

This and the flaws create a potential image problem for this kind of AI but the obvious benefits and conveniences outweigh the flaws.

Combining LLMs with other systems makes hallucinations less dangerous

Question 2 is much more interesting

My feeling is that the answer is to use LLMs as a component of a more intelligent system rather than as the main driving engine of it.

This is a view shared by many others for example (mostly random selection):

Whether companies choose to focus on the LLM side or pivot to work on combined system will depend on the company. I expect wealthly companies like openai to be more likely to use that wealth to do so. LLMs coming out of big companies like Google/Alphabet already have that machinery behind them.

There is active research into reducing hallucinations

That is not to say that hallucinations in LLMs are not a problem in and of themselves that is not worth addressing because it certainly is and there is of lot research going into that area which other people can probably describe much better than I.

$\endgroup$

Not the answer you're looking for? Browse other questions tagged .