The next problem is verifying all of that information.

It will require multiple dozen auxillary Google searches and reading through countless articles and message boards.

GPT-3 is a great research tool as a brainstorming engine, but sadly cannot currently be formally trusted to output truthful information, even with conditioning.

It takes someone with prior knowledge or subsequent research in order to make use of the result. I'd say this isn't what the author was intending.

Even if you got the info straight from another human being I can't imagine anyone would take it at face value and just schedule their trip.

I wouldn't buy the tickets sight unseen, but I would trust you that it's actually a real place, actually in Croatia, and not being conflated with similar-sounding places.

Humans, unless they're disturbed or malicious, tend not to hallucinate things out of whole cloth.

So... we’ve got ML models to the point of being able to hallucinate?

Yup: prompt a large language model to tell you about itself and it will sometimes weave weird tales. For example, I just had the following interaction with GTP2large[0]:

Me: What is your name? Model: I'm Mark A. <continue> Dreyer, and I'm with the <continue> Federal Communications Commission (FCC)

Obviously, it's not actually a guy named Mark A. Dreyer and, as far as I can tell, there is no Mark. A. Dreyer in a prominent role at the FCC. That, as far as I'm concerned, is essentially a hallucination.

If you think about how they work, this isn't totally surprising. The model has been programmed to take the input, turn it into a series of tokens, and then produce another series of tokens that are especially likely to follow the input ones. That's it.

To the extent that this corresponds to reality, it's only because reality is overrepresented in the training data, and so token sequences corresponding to reality are often more likely under its model. This also explains how it goes off the rails: Mark is a very common name (in the top 25 for men for decades) and once you generate a stereotypical middle-aged white guy name, it's not a huge leap to him being a bureaucrat, etc. I forgot to delete the huggingface boilerplate, which talks about advances in technology (etc), which probably sent it down the FCC route too.

That’s totally wild. The visual ML output is very obviously like tripping. But what you’re describing is very similar to how someone once described being high on acid to me, where no sentence could be completed without it turning into another and another and another until it was all just gibberish. It’s an interesting feat to be at this stage simply though back propagating neural nets, but there’s a huge gulf between where we are and creating __understanding__ (he said, trying to sound smart)

Hallucination in context means something rather different.

Something like coming to accept some data-set local optima as a global truth and deriving beliefs without underlying basis.

"Yes"?

The whole point of these models is to make up something that seems "realistic", but not regurgitate the original contents. If that's not hallucination I don't know what is.

Google's answer box also can't be trusted to output truthful information.

Fun one today: https://www.google.com/search?q=is+fortigate+open+source

Prompt: "Is Fortigate open source?"

Article: "Based on the parameters we compared, it seems that pfSense is the more favorable solution because it is open source and also offers great features."

Google: "It is open source!"

Yes, these are forever spreading disinformation on various topics in my social group. I am forever telling people to actually click and read the article rather than trust the automatic tl;dr by Google.

I believe this is the point that @drivebycomment was making, that answering such a query is not possible right now. The query needs to be reframed by using keywords like the GPT-3 prompt did, and even further refined than that, and you still have to finish the last mile of research.

Google simply returns human written results, so delegating the trust problem. You can track author's record and judge if they're trustworthy yourself. Doing the same thing for GPT-3 is much harder. How are you going to debug the model? Probably there is only a handful amount of people who is capable of debugging the model even with all necessary access granted. You can think it as autonomous driving which already has reached to a good level for 99% of use cases (at least for Waymo and Cruise), but no one think it's going to mass deploy robotaxi anytime soon.

Could this be addressed by an ensemble of GPT-3s fine tuned with different data?

Beat me to it, I got a very similar response by telling GPT-3 to behave as a 'highly advanced search engine'. It also went for Hvar and Korcula.

query: what is the koh lanta of croatia?

response:

Koh Lanta is an island in Thailand that's known for its beautiful beaches and clear waters. It's a popular destination for tourists, and it's also become a popular spot for digital nomads in recent years.

Croatia doesn't have an island that's quite like Koh Lanta, but there are a few options that come close. The island of Hvar is a good option, as it has beautiful beaches and a laid-back vibe. The island of Korcula is another option, and it's known for its medieval villages and stunning scenery.

That proves GPT-3 is good at constructing sentences based on a prompt, not that it actually knows how to vibe-check tourist destinations.

I have been in HVar. It's not cheap, it's the opposite... It's where the super rich people hang out.

Anecdote: When I told a random person I got there by plane he automatically assumed I owned a plane."

GPT-3 and DALL-E scare me. This is too good.

I wonder how long it will take before we have a GPT-3 bot commenting here on HN

I get sent GPT-3 blogspam as "guest posts" from wannabe "contributors" every day, and I can say that the moment you expect GPT-3 to write anything remotely technical, it turns to shit. Simultaneously, anything people have written by GPT-3 is frequently full of algorithmically detectable plagiarism.

Needless to say, all of these people get told to fuck off and get a job.

GPT-3 authored content was already top of HN not long ago

https://www.theverge.com/2020/8/16/21371049/gpt3-hacker-news...

the author of that post was called out by the HN mod for engaging in vote manipulation in the past, and lying about several aspects of the GPT3 post [1]. I wouldn't put so much faith into that author

[1]: https://news.ycombinator.com/item?id=24164749

How do you know there's not one already?

Whistleblowing is usually a way to redirect attention.

>It has tourist amenities without really being touristy

All the results it gave are flooded with tourists, and some of them are among the most expensive you can get in Croatia. :shrug:

Unsure of what to think of the results, as it seems to have very little relevance to the input description other than being in Croatia and being tourist spots.

I'd be super interested to see if these other places actually fit the descriptions it gives, or if it essentially just generated a likely pattern of text to be seen after the first part saying there's no exact match and then filled in random known Croatian location. I honestly don't have any intuition of which of those is more likely!

Hvar is filled with 25 year old tourists who drink a lot and maybe hope to get into modeling as a side hustle. It does not fit the prompt at all. Korchula is a slow, family oriented island. Maybe a better fit? But it really just seems like GPT has picked up on a list of the most popular destinations in Croatia.

Istria is also very touristy.

Can anyone familiar with Croatia say whether GPT-3 is reasonable here? It's great at saying plausible things, but it will freely make things up if it doesn't know.

I think it really depends on what you're after and the micro location where you spend your day and night. Hvar and it's city is definitely not laid back and slow. It is just very expensive, busy and loud. Korcula is fine, but I would suggest something like Vis island or for someone appreciating very small islands: Susak or Ilovik. Being on all of those places many times in my life, I think it mostly depends who you're with and where your apartment is. Every island can be very diverse in itself as well. Let me explain: if you're placed in a town usually it will be very loud on weekends and nights due to open air concerts/performances, and sometimes even during the weekdays. So if you desire something where you'll get a nice sleep this can actually be hard to come by because every cafee is open during the whole night each day entertaining a few drunks, blasting music, and not letting hundreds of people sleep with open windows (nights are hot as well). Actually it was much more bearable with covid restrictions in place because the cafees weren't allowed to play music after IIRC 10PM. A lot of islands also have no police or very minimal presence, so there is a lot of elements of lawlessness and locals making their own "rules".

I could write pages of this stuff, but let me write more positive things. The best way to enjoy the islands in Croatia is actually by boat, rented or otherwise (they are not expensive). There are beautiful quiet coves where you can spend your days and nights, only going for resupply in towns. You anchor in a quiet cove, and in the evening you take your tender boat to the town if you want to walk around it. Next day you sail to another island and repeat. That experience is actually awesome.

If a boat is out of the question, being in some apartment is an option, I would just advise not to be in the town center or near it, or anywhere where some very local beach is. Mostly there will be a caffee with blaring music and nobody likes to sleep with earplugs. Also this is and advice for any island, and true, a lot of those islands are quite different than each other.

> While these places do have some tourist amenities, they are not as touristy as Koh Lanta

This shows, if you had any doubt about it, that it did not actually understand the question.

If search engines ever become like this there will be debates about copyright just like we have for copilot.