Hacker News Reader

Hacker News

Comment by criddell | Hacker News Reader

Full thread

criddell

Will GPT-3 tell you when it doesn't have an answer, or does it always fabricate some type of non-empty response?

Edit:

I just asked "What problem did Luke Skywalker have with tribbles?"

Answer: Luke Skywalker had a problem with tribbles because they would constantly try to escape from his lightsaber...

View on HN

It's entire purpose is to generate the text that follows your input, so the responses will never be "I don't know". However if you ask "are you sure about .." or "why did you say ..." then it will say something like "I apologize, I made a mistake." and sometimes expand on the reason why it was wrong.

Perhaps try placing something around questions that asks ChatGPT to answer only when it's confident? E.g.

> If you're sure of the answer, give the answer to <QUERY>, otherwise say "not sure".

Write a script in which Luke appears in Star Trek and gets into problems with tribbles.

Scene: The USS Enterprise is docked at a space station when Luke Skywalker, dressed in his Jedi robes, steps onto the ship.

Captain Kirk: Welcome aboard the USS Enterprise, Jedi Skywalker. What brings you to our ship?

Luke: I'm on a mission to stop the spread of a dangerous new threat in the galaxy, but I've received reports of strange creatures causing problems on your ship.

Captain Kirk: Ah, you must be referring to the tribbles. They may look cute and harmless, but they're quite a nuisance.

Luke: I've heard they multiply quickly and can clog systems. Is that true?

Captain Kirk: Yes, it is. We've been trying to get rid of them, but they keep coming back.

Suddenly, the ship's alarms start blaring, and a crew member runs onto the bridge.

Crew Member: Captain, we've got a problem in the engine room! The tribbles have taken over the systems!

Captain Kirk: (sighs) This is just what we need. Jedi Skywalker, can you help us?

Luke: Of course. (closes his eyes and concentrates) I sense a disturbance in the Force. I think I can help.

Luke heads to the engine room and uses the Force to move the tribbles away from the systems. However, more and more tribbles keep appearing.

Luke: (frustrated) This is never-ending! How do these creatures keep multiplying so fast?

Captain Kirk: It's their nature. They breed at an astonishing rate.

Luke: (sighs) I'll do what I can to help, but I think I may need some assistance from the Force.

With a determined look on his face, Luke draws his lightsaber and starts slicing through the tribbles, using his mastery of the Force to deflect their attacks. Finally, the tribbles start to retreat, and Luke is able to restore the systems to normal.

Captain Kirk: (impressed) Jedi Skywalker, you have my gratitude. I don't know what we would have done without you.

Luke: (smiling) It was my pleasure. I'm just glad I could help.

Captain Kirk: (smiling) You're welcome on this ship anytime. We could use a Jedi like you on our crew.

Luke: (laughing) I appreciate the offer, Captain, but my path lies elsewhere. May the Force be with you.

The scene ends with Luke departing the USS Enterprise and continuing on his mission, leaving the crew to deal with the aftermath of the tribble infestation.

Man that gave me a mental image of xwings and tie fighters docking the enterprise. Makes my neardy heart skip a beat.

I can't make it happen for me if I tell it that unsure is a valid response.

What problem did Luke Skywalker have with tribbles? And if you're not sure say you don't know.

Answer: Luke Skywalker didn't have any particular problem with tribbles. He did not interact with them in any of the Star Wars films.

I tried it a few times and it's all the same for me.

And to make sure it understands nonsense.

How many furlongs are there in a wicker basket? And if you're not sure say you don't know.

Answer: I don't know.

It could be that explicitly saying that "if ... say you don't know" could sway ChatGPT into generating this response, as it's encouraged to do so. Needs a check if it continues to respond to the (preferably, less obvious) facts it known about (or can infer).

Just like multiple choice where a choice is none of the above.

I asked what the current version of Bootstrap Framework was. It said as of it's cut off date, it was a certain version, that seemed slightly off and I regenerated and this time it gave a number that was higher and thus definetly wrong, I said that is wrong. It said it was sorry and that it was wrong and gave another number, (also wrong), I said wrong again, and again confidently provided a new number. GPT-3 is astonishing, at so much and if you know how to use it, and are asking questions in a genhre you know then when it gives you partially wrong info you'll usually know but will still help you along your way. But if asking in areas you don't know and it gives you a wrong answer such as how long and what temp to cook a chicken etc, if it gives you a wrong answer and you don't know better, could be troubling.

I just asked the same question

“Luke Skywalker did not have a problem with Tribbles. The character Luke Skywalker is from the “Star Wars” franchise and tribbles are from the “Star Trek” franchise.

I've never had it tell me it didn't know something, both GPT-3 and ChatGPT variations.

If you ask if for the weather in Paris, it will say it doesn't have that information.

It's trained to say it doesn't have information to recent data, such as today's weather.

But for questions that aren't based on recent data, it should still be able to say "i don't know" or "that question doesn't make sense" in the case of Luke Skywalker's problem with tribbles for example.

Edit: I do realize we're criticizing a computer program for not understanding the difference between Star Trek and Star Wars. Strange world we live in.

I think its a commonly accepted opinion that a person's ability to admit when they don't know something is a strength. It would be interesting to a see how much a constraint for known validity impacts the ability of a model to synthesize responses. In other words, where do we draw the line between knowing and "thinking we know" something (i.e., do we draw the line at 95% confidence)? Surely we do not hold each other to a strict requirement of 100% confidence, but there must be some acceptable range.

There's a lot of fan fiction out there in the Star Trek & Wars universes, maybe even some crossover stories. If the GPT training model includes that, then it might be fairly confident about an unfortunate incident on the Millennium Falcon involving Luke Skywalker and tribbles. Or does it understand the idea of "canon"?

That’s not a strength when it comes to AI assistants because it feels like an extremely common response pre-chatgpt

Just an example I found. It couldn't know the answer to this as it didn't happen AFAIK. It makes guesses for me all the time - it's usually when it's something really specific like a niche program on the cli.

> Why does darth vader make a reference about kittens in the movie

> Darth Vader makes a reference about kittens in the movie as a form of humor or irony. In the Star Wars franchise, Darth Vader is known for his serious demeanor and ruthless actions, so referring to something as cute and innocent as kittens is a stark contrast to his usual behavior. It adds a lighthearted moment to the movie and allows the audience to see a different side to the character.

I asked “What is Kurume coffee from Ethiopia?” and ChatGPT had a fine response - “I’m sorry, there doesn’t seem to be any information about a specific type of coffee called “Kurume” from Ethiopia… Could you please provide more context or clarify your question?”

It’s real though - I was drinking Kurume this morning, and googling finds a blog post from 2020 about it.

ChatGPT seems to only know it as Kudhum:

> Kudhum is a coffee variety grown in the Kaffa region of Ethiopia, which is known for being the birthplace of coffee. Kudhum coffee is known for its high quality, producing beans with a distinctive flavor profile that is characterized by a full body, medium to high acidity, and notes of fruit and spice. The variety is grown at high elevations, which contributes to its unique flavor profile. Kudhum coffee is considered one of the top-quality coffees from Ethiopia and is highly sought after by coffee lovers and roasters around the world.

Not knowing requires epistemic awareness, which LLM simply does not possess. I guess one can cleverly simulate it, but I doubt it’s the goal of OpenAI.

Reply on news.ycombinator.com