Hacker News Reader

Comment by GolfPopper | Hacker News Reader

Every time I see one of these topics, I go ask chat GPT a question to which I know the answer on a topic where I would like to be able to get useful answers to similar questions that I do not know the answer to.

This time it was, "Did Paul Edwin Zimmer write a fourth Dark Border novel?" (Real answer: Yes, Ingulf the Mad. You can find the answer on his Wikipedia page.[1])

ChatGPT's[2] answer: "Yes, Paul Edwin Zimmer wrote a fourth novel in the Dark Border series titled "The Dark Border." This book was published after the original trilogy, which included "The Dark Border," "The Gilded Age," and "The Silver Sphere." If you're interested in the themes or plot, let me know!" (Note: these are not the titles of the 2nd and 3rd novels in the series. Also, it gave me the same name for the putative 1st and 4th books.)

Pure hallucination.

1. https://en.wikipedia.org/wiki/Paul_Edwin_Zimmer 2. https://chatgpt.com/

View on HN

ChatGPT 4o and 4o-mini with Search were able to answer this question without any issues. Maybe you didn't enable the search functionality?

---------------

4o:

Yes, Paul Edwin Zimmer wrote a fourth novel in his Dark Border series titled Ingulf the Mad, published in 1989. This installment focuses on the characters Ingulf Mac Fingold and Carrol Mac Lir, detailing their meeting and acquisition of mystical swords. Notably, Istvan Divega, the protagonist of the earlier books, does not appear in this novel.

---------------

4o-mini:

Yes, Paul Edwin Zimmer authored a fourth novel in his Dark Border series titled Ingulf the Mad. Published in 1989, this book shifts focus from the previous protagonist, Istvan DiVega, to explore the backstory of Ingulf Mac Fingold and Carrol Mac Lir, detailing their initial meeting and the acquisition of their mystical swords.

The complete Dark Border series consists of four novels:

1. The Lost Prince (1982)

2. King Chondos' Ride (1983)

3. A Gathering of Heroes (1987)

4. Ingulf the Mad (1989)

These works delve into a world where the Hasturs' power is crucial in containing dark creatures, and the narrative unfolds through various characters' perspectives.

Why is it that comments complaining about flaky AI responses never share the chat link, and then invariably someone replies with examples of the AI answering correctly but also fail to share the chat link?

I just forgot that this feature exists. At any rate, isn't it more convenient to share the output directly? What's the difference?

4o-mini: https://chatgpt.com/share/6725f6fb-3b84-800b-9b4d-3e82554fbf...

I can't figure out how to share the 4o one. The UI is weird.

404 Not Found

I used the "Delete All Chats" button not knowing that that would include the archived chats. At any rate, you could try this out yourself with an account on ChatGPT. Just prefix the question with `/search`.

A lot of these responses are so funny. "Um, when _I_ asked this question to ChatGPT it gave the right answer!" Yeah, because it's flaky! It doesn't give the same answer every time! That's the worst possible outcome!

Isn't it even a bit interesting that GP has tried it every time something new has come out but not once gotten the expected answer? Not only that but gets the wrong titles even though Search for everyone else is using the exact Wikipedia link given in the comment as the source?

LLMs are run with variable output, sure, but it's particularly odd if GP used the search product as it doesn't have to provide the facts from the model itself in that case. If GP had posted the link to the actual chat rather than provided a link to chatgpt.com (???) I'd be interested in seeing if Search was even used as that'd at least explain where such variance in output came from. Instead we're all talking about what could have happened or not.

Normally I would agree with you but for such a fuzzy topic involving search it would require all pages and discussions it can find to be the same and not outdated. I don't see why anyone would presume these systems to be omnipotent.

Funny, yes. But... a webforum is a good place for these back-and-forths.

I'm on the other side of this fence to you. I agree that the conclusion here is that "it is flaky." Disagree about what that means.

As LLMs progress, 10% accuracy becomes 50% accuracy. That becomes 80% accuracy and from there to usable accuracy... whatever that is per case. Not every "better than random" seed grows into high accuracy features, but many do. It's never clear where accuracy ceilings are, and high reliability applications may be distant but... sufficient accuracy is not necessarily very high for many applications.

Meanwhile, the "accurate-for-me" fix is usually to use the appropriate model, prompt or such. Well... these are exactly the kind of optimizations that can be implemented in a UI like "LLM search."

I'm expecting "LLMs eat search." They don't have to "solve truth." They just have to be better and faster than search, with fewer ads.

Several folks have already mentioned that ChatGPT with search returns the correct answer to your question (with a source to explore directly).

I really think this latest release is a game changer for ChatGPT since it seems much more likely to return genuine information than ChatGPT answering using its model alone. Of course it still hallucinates sometimes (I asked about searching tabs in Firefox Mobile and it told him the wrong place to find that ability while citing a bunch of Mozilla help docs), but it's much easier to verify that by clicking through to sources directly.

It feels like a very different experience using ChatGPT with search turned on and the "Citations" right side bar left open. I get answers from ChatGPT while also seeing a bunch of possibly relevant links populate. If I detect something's off I can click on a source and read the details directly. It's a huge improvement on relying on the model alone.

Chat gpt will not always return the correct answers, thats a fundamental limitation of how it works since its non deterministic. So just saying "it worked for me" means nothing.

Yes and having direct links to sources to validate its answers is a huge benefit given the fundamental limitations of LLMs. Now you can decide if you trust the source material directly.

My golden hallucination tests for chatgpt are identification of Simpsons season and episode based on description of some plot lines or general0 descriptions.

I had to refresh my knowledge base visiting fandom websites to review the episode selected as answer as chatgpt tendency to mix things up and provide entirely made up episodes make it hard for bar trivias (and makes me doubt myself too). The same with other tv series such as House MD and Scrubs.

My gut reaction here is that the hallucination is caused by how you [rightfully] formed the prompt. GPT has no way of reliably determining what the fourth book is, so it infers the answer based on the data provided from Wikipedia. I'll bet if you changed the prompt to "list all books by Paul Edwin Zimmer", it would be incredibly accurate and produce consistent results every time.

Its true.

I usually seed conversations with several fact-finding prompts before asking the real question I am after. It populates the chat history with the context and pre-established facts to build the real question from a much more refined position.

Here's what ChatGPT 4o mini (free) answered to me to the same question:

No, Paul Edwin Zimmer did not write a fourth novel in the Dark Border series. The trilogy consists of "The Dark Border," "The Dark Border: The Return," and "The Dark Border: The Reckoning." After these, he focused on other projects and did not continue the series.

Here's o1-preview after thinking for 7 seconds,

> As of my knowledge cutoff in October 2023, Paul Edwin Zimmer did not publish a fourth novel in the Dark Border series. The series comprises three books: > 1. The Lost Prince (1982) > 2. King Chondos' Ride (1982) > 3. A Gathering of Heroes (1987) > > Paul Edwin Zimmer had plans to continue the series, but he passed away in 1997 before any additional novels were completed or published. There have been no posthumous releases of a fourth Dark Border novel. If there have been developments after October 2023, I recommend checking recent publications or official announcements for the most up-to-date information.

ChatGPT with search:

Yes, Paul Edwin Zimmer wrote a fourth novel in his Dark Border series titled Ingulf the Mad, published in 1989. This installment focuses on the characters Ingulf Mac Fingold and Carrol Mac Lir, detailing their meeting and the acquisition of their mystical swords. Notably, Istvan Divega, the protagonist of the earlier books, does not appear in this novel.

love these HN comments where someone ignores the topic entirely. with search enabled it gives the correct answer and sources Wikipedia correctly, not that you tried it

If AI produces a correct answer for something you know, there's no guarantee that it produces the correct answer for something you don't know. I'd rather find a way to verify that answer, if that answer has any importance.

Perplexity adds citations to the end of each statement. Its nice to be able to go to the primary source and verify whether its mixing anything up.

This is just unrealistic, if I have to check the sources every time I'd rather Google it.

Same. This is why I don’t really use AI in my everyday workflow other than for brainstorming.

It is not just for AI. The same can be said of Wikipedia, if correctness matters, for example if you are writing an article/paper/essay, go for primary sources, not Wikipedia, and certainly not ChatGPT.

You can use them as a starting point though.

This is why I initially held off on using ChatGPT: as a computer scientist, I was focused on testing adversarial strategies and approaching it with a hacker mindset. My takeaway? This isn’t the best way to use LLMs. They’re actually fantastic tools for quickly understanding, and then verifying information. While they can reduce the need for extensive searches, they also help uncover concepts you might not know, especially when you engage in dialogue.

I also search on Google and find "hallucinations": pages with false content that rank better than the real information. In Google the current issue is about ranking and forgetting information that was previously available.

I am using DDG's assist function: "Paul Edwin Zimmer write a fourth Dark Border novel".

DuckAssist Result: The fourth novel in Paul Edwin Zimmer's Dark Border series is titled "Ingulf the Mad." This book focuses on characters Ingulf Mac Fingold and Carrol Mac Lir, detailing their meeting and the acquisition of their mystic swords, while the main character from the earlier novels, Istvan Divega, does not appear.

Apparently, you forgot to click the "don't hallucinate as much" checkbox. How silly of you.

This is the answer I get : Yes, Paul Edwin Zimmer wrote a fourth novel in his Dark Border series titled Ingulf the Mad, published in 1989. This book focuses on the character Ingulf Mac Fingold, detailing his adventures and the origins of his mystical sword

(With source as wikipedia)

> 2. https://chatgpt.com/

Sharing the actual chat link is more useful since AI responses are non-deterministic by nature so reproducing your interaction is very difficult.

Chatgpt and perplexity both answer that question perfectly fine. Embarrassingly enough, Google's Gemini Advanced (which I got for free with a Pixel Pro phone but would cost 200$+/y) failed to answer it.

Why has not a single person in this thread who has tried this just shared a link to the ChatGPT convo where they got these results?

If you don’t trust the answer I always tell it to look it up and it invariably gets it right

Invariably seems like a strong word for the output of an LLM. If we can prompt a language model to double-check it's work (and that means it always gets it correct), why is that not part of the base prompt?

It’s less double check and more a request to go check sources online, which it does, provides links to those sources and then summarizes.

how many times people clicked a spam ad instead of real website in google? if you want to compare, then do it to all of them.

hallucination is user error

Reply on news.ycombinator.com