Deep Travel: AI as a Travel Writer
When I was at Google Zurich for a while I was part of the crew that introduced new hires. As part of that we’d always ask them what their hobbies were. The most common reply was “photography, hiking and travel”. So it’s no wonder that people in tech often come to the conclusion that the best next thing to do would be a travel startup. The best advice for these people is “Don’t do it” — I should know, I’ve done two.
So it should come as no surprise that when GPT-3 came out, my first thought was: can we use this to help write an automatic travel guide? The value proposition of the last travel startup I was involved in, Triposo, was that the web has all the information to travel the world, all you need is some cleverness to package that information up into a handy travel guide that fits the phone you already have in the palm of your hand.
GPT-3 has all the information of the web already, so all we need to do is ask. The basics are not hard. GPT-3 can complete bits of text, so if we give it a list of cities, each followed by a snippet taken from say, Wikivoyage, we get some decent results:
By the way, I’m embedding an image here, since Medium doesn’t seem to know about tables. I put the tables in a Google Doc in case you want to copy the text for some reason.
It reads pretty convincingly, but not all that exciting. More factual than “I want to go to there’. If we feed the descriptions not from Wikivoyage but from the slightly more dramatic Lonely Planet into the AI, the computer starts writing something more inviting:
That’s better, maybe? Definitely more interesting and more likely to draw visitors to your travel site. Can we make it even more interesting? What if instead of Lonely Planet, we feed it examples that sound like the utterances of a long term traveler bragging in a local bar about all the amazing things they’ve done? It becomes a lot more personal, maybe something that could power an instagram travel feed? Maybe with better pictures.
I’m sure someday we’ll have the technology to produce a matching image for these text fragments, but for now I just got them from Google Image Search — sufficiently cropped to claim fair use I think.
There are some other interesting experiments you can do based on travel guides. You can reverse the generation where you describe a city the way that Wikivoyage would and you get back which city matches best, which makes it a bit like a travel recommendation system: “Decent beer and pretty canals” -> Amsterdam. Fun, but like a lot of these things a bit of a hit and miss. For skiing and beach it recommended Lanzarote. And even though Expedia offers a selection of ski hotels for that island, that’s just another piece of software gone wrong.
I wanted to take this one step further and generate more of a travel guide per destination. Remember, the way that GPT-3 works is that you give it a piece of text and then it will finish it for you in the best way it can. So if you show it a piece of text with the name of a city, followed by the description of a city, followed by the name of another, followed by the description of that city and we repeat that pattern for a bit, ending on the name of the city we want a description for, GPT-3 catches on to that and spits out a somewhat relevant description.
That works fine for the shorter pieces of text, but doesn’t really scale up to a one pager about a destination with a short introduction followed by a list of points of interest, each with its own description. We can take advantage of this structure and let GPT-3 not only suggest a description for a destination, but also the names of the points of interest. For each of those we ask GPT-3 to generate a description again.
The results are … mixed. Oftentimes it looks pretty convincing, especially for places you don’t know much about. If you look at a place you know well though, it quickly falls apart. For example, about Ghirardelli Square in San Francisco it says:
Sounds reasonable. The year is correct, I am not sure about the Paris reference, but otherwise, looks good. I’m ready to go and explore those cafes.
In Karlskrona, a town in Southern Sweden, it has this to say about the Hanseatenhaus:
The hanseatic league was a thing, but Karlskrona was founded in 1680, at which point the league had lost most of its power. Hanseatenhaus does sound like something where merchants could stay — there is one in Germany. Ulrik Frederik Gyldenløve was a Norwegian general involved in the wars that sealed the fate of the Hansa and was actually born in Germany. Hans Christian Andersen is from Odense, a town not too far away, so it seems quite plausible he would stay in a hotel in Karlskrona.
Somewhat more in between:
This is a famous theatre in Buenos Aires and seems totally worth a visit. It’s probably not the largest operatic stage in the world though. Callas and Domingo sang there, that’s all true. There is a tour of the interior, but I can’t find a reference to any green room. So it’s not exactly true, but neither is it all made up.
To me this is most fascinating. These large language models easily drift between fact and fiction, coming up with factual descriptions, but making up stuff where they lack knowledge without really being able to tell the difference. It’s very human! People say stuff all the time that isn’t true per se but sounds like it could be and when pressed they’d deny they made it up.
Have a look around!