Oh boy, where do I even start with this...?

Someone posted yesterday to Hacker News their new project: Meoweler, a cat-themed travel website that's almost entirely auto-generated.

I have written here before about the realistic dangers of AI, predicting that we would see "garbage in any medium that offers even the slightest of rewards, financial or not". And boy was I right...

Here's a short, non-exhaustive list of problems I found with this website:

• Offensive stereotypes. About half the Argentinean cities I've checked are based on stereotypes about the country, regardless of whether they are even remotely applicable.
• Offensive pictures. I grew up in not-so-popular cities across the country and I can tell you: I wasn't born in a war-torn city, I didn't go to school in Tatooine, I didn't study in a mirror-like desert, and I didn't go to University on a mountain city. I just lived in perfectly normal, standard, slightly boring cities. But you wouldn't know any of that by looking at the Midjourney-generated illustrations for each one of them.
• Bad information. I don't care that much that the AI suggests that Potsdam's Brandenburg Gate is a "mini version" of the one with the same name in Berlin (which AFAIK is not true), nor that it talks about getting lost "in the maze-like streets of the Old Town" for the squarest blocks you'll find outside of Manhattan. But when your list of attractions includes sightseeing points in cities more than 200km away... and in a different Province... I hope you remember to bring your good walking shoes.
• Dangerous information. The site includes an "LGBTQ safety" rating which is auto-generated and not checked against any official source. The guide will also gladly include locations "off the beaten path" that are plain unsafe for a tourist. One such advice suggests "lounging in sunbathed plazas" in a city that Travel Safe - Abroad rates as having a high risk of pickpockets and medium risk for muggings. And while I'm not here to tell you how to live your life, I think "don't" is a solid advice for this situation.

The source code for the website is open, and the license seems to be "do whatever you want with it, no restrictions". And you know what? People will do whatever they want with it: they will spin God-knows how many clones (dog themed, child themed, circus themed, ...), slap ads on it, and the internet will drown a little more in auto-generated, unchecked, ad-ridden SEO garbage.

But hey, at least the website has the tiny footer "Beware, most of the content was generated by AI". So if you end up being robbed at gunpoint because you went "off the beaten path" it will actually be your fault.

One final digression: I decided to check some of the cities listed as "Do not travel" in the US State website and I landed on Kabul, Afghanistan, which the US warns you against visiting due to armed conflict, civil unrest, crime, terrorism, and kidnapping. The AI description reads "Kabul is a friendly city: despite the violence, Kabul residents are extremely hospitable and friendly towards visitors", which to me as a (computational) linguist is fascinating: if the locals are hospitable and friendly, then who is perpetrating "the violence"? Is this sentence blaming foreigners for the situation in Kabul? Can someone kidnap you and be hospitable at the same time? And since we know that a guest who kills their hosts in their own house loses privileged status, does robbing, kidnapping, and maybe murdering your visitors mean that you lose your status as a friendly city?

I would say that I'm maybe overthinking it, but let's be honest: given that no human thought of any kind went into this entire website, literally any attempt at understanding it is, by definition, overthinking.

(I have changed the title of this article to add a "(I)" at the end. I get the feeling this is far from the last entry in this series...)

Looks like I wrote this article and forgot to publish it for a whole month. I hope the extra length will make up for it.

Now that the 61st annual meeting of the ACL is over, and based on my experience as a member of the Virtual Infrastructure Committee, I would like to share some thoughts on what my ideal virtual conference infrastructure would look like.

A conference, for our purposes, is an event that takes place in some random venue around the world. Researchers submit papers with their latest research results and, if accepted, present those results in sessions where related papers are bundled together. Conference attendees typically hear a short talk about the paper (or read a poster) and then ask questions to the authors.

Given that we live in a post-pandemic world we are going to prepare for a hybrid conference. This is cheaper (less mouths to feed means lower fees) more accessible (researchers in countries with less resources will thank you for the plane ticket and hotel costs they'll save), greener (less people flying around), easier to archive, and a good idea overall.

The handbook

It is traditional in conferences to offer a handbook containing all information about the conference - schedule, paper abstracts, venue maps, social events, tourist info, sponsors, and so on. But given that this year's booklet reached an impressive 462 pages it may be time to think about a website or app to replace it.

I do not believe that the conference website and the handbook can be unified, for reasons I'll explain later. So let's assume we want to develop a website or app exclusively for the handbook. Let's call it AppBook.

The main use cases for AppBook are:

• Checking the schedule to figure out either what's going on at some time (exploration) or to find out when is a specific event taking place (information retrieval).
• Adding reminder for events you would like to visit, from entire events down to specific talks. It is not unusual to be there for Paper 1 in Room A, then go to Room B next door for Papers 2 and 3, and so on.
• Get details about a specific paper: if the title catches your attention then you may want to read its abstract, and/or you may want to see what else the authors are presenting.
• General information about the venue: maps, tourist info, vouchers, etc.

The schedule is, I believe, the hardest part. Sure, there are plenty of scheduling apps around that do the information retrieval part quite well, but how about the exploration part? Taking the EACL 2023 handbook as an example (because I have it here on my desk): I can see that Session 6 on Day 2 has two Tracks I'd like to visit, namely, "Information Extraction" and "Generation". With the paper booklet I can put one finger in the Schedule page, flip back and forth through the pages for each Track, decide which one is more interesting at which time, and plan my visit accordingly.

Compare that to a regular website calendar: clicking on a Track takes you to the abstracts, but clicking "back" takes you to the general calendar. In the specific case of the talk I heard today that means scrolling back down to the correct time slot, clicking on "+", finding the right Track (there are 13 parallel workshops), open this Track and, since the interface doesn't allow for right click, remember somehow what the previous one had to offer. Repeat this for all 3-4 Tracks that one could consider, and this quickly turns into a mess.

This leads me to the first three requirements:

• AppBook has to be lightning fast, which in this case means offline. Sure, you can periodically download updates to the database, but there is no way to implement "lightning fast" when you have 2k+ attendees checking the website at the exact same time with shoddy WiFi. I can tell you that a full conference fits in ca. 10Mb of JSON.
• The interface has to trivially allow back-and-forth between all events taking place in parallel. Swiping left/right to switch Tracks could work, but I'd have to play with a prototype first to ensure it's not confusing.
• It has to be mobile first. If you want to get rid of the handbook you need to ensure that even the humblest of devices can provide an equal or better experience than the paper version. When in doubt, remember that computer interfaces could already do this type of stuff in real time 20 years ago.

Chat

Even though I insisted on AppBook being offline first, it would be foolish to ignore all it could do if we judiciously let it go online. At its most basic, you should be able to download the actual papers directly from the interface. The next step is allowing access to both pre-recorded and live videos, which is a lot of work on the backend but relatively straightforward on the frontend. And if your internet is down you'd still have the handbook information.

Chat, on the other hand, is where I think the app could shine. Having a dedicated chat for every paper is a great way to ensure that all attendees can contact you, which helps in-person attendees to talk to you even outside your presentation window and empowers virtual attendees to talk to authors they would otherwise miss. And even better: once you allow attendees to create their own channels you can sit back, relax, and watch communities of like-minded people come together, plan events, and have fun.

If you choose to go this way then you could consider adding a feature to AppBook: "log me in as author". If you register as the author of a specific paper then you would get some extra functionality such as administrator rights over channels for your papers, notifications whenever someone posts in there, and reminders of where and when you are supposed to go.

And finally, if you use something like Zoom then you may want to disable the chat function in videos. Sure, people may be used to it, but if you leave them on then now you have two sources of questions and you've fragmented your userbase (see the first paragraph of "Supporting Infrastructure").

Website

If you are a virtual attendee, spending a week looking at your phone would probably be torture. More so when there's almost certainly a big screen somewhere around that you could be using instead. So we need a website too.

Could AppBook and this website be merged together? I don't want to say "no", but I will however say "well... errmmm..." while making face gestures. See, the website and AppBook have different audiences and different use cases, to the point that unifying them might just be impossible:

• Speed:
  • AppBook: lightning fast.
  • Website: regular website speed.
• Data source:
  • AppBook: offline first.
  • Website: online first.
• Device:
  • AppBook: designed for your mobile phone.
  • Website: designed for your laptop and/or desktop computer.
• Interface:
  • AppBook: one window at the time. Interaction is everything. You may need to go native, but maybe a web app would do.
  • Website: multiple windows at the time. Most interaction issues can be solved opening a second tab.
• What would you use it for:
  • AppBook: figure out what's going on and/or where to go.
  • Website: stream live events, read papers.

Poster sessions

Posters are tricky. The current best solution I've seen is Gather.Town, but now you are depending on an external platform. I imagine you could get rid of poster sessions altogether with a combination of pre-recorded videos and chat, which is coincidentally what the ACL Conference offers every other year. Then again, they also offer Gather.Town, so it's hard to say which one is better.

In-person notes

Organizing a conference is a lot of work, but luckily we've done it for long enough that there isn't much new that I could say. I should however point out some details to keep in mind:

• All talks should be live-streamed and questions should be accepted both in-person and via chat.
• Every room needs the proper equipment for remote presenters. That means being prepared for presenters joining via teleconference and, conversely, having a mic at hand for attendees to pose their questions.

Supporting infrastructure

In my ideal conference platform, there is exactly one source of truth and one way to get information. There are no two schedules needing coordination, nor two chat platforms, nor two video platforms. There is one way and one way only.

In other words, the infrastructure would be built as such:

• A central "source of truth" database with all information about the conference. You change it there and it is reflected everywhere (which in the case of AppBook means "eventually").
• A file hosting platform with every paper and poster.
• Some kind of video platform for live-streaming events.
• The website, which is your main portal to all those things above.
• AppBook, which presents a subset of the above information but in an easy-to-query format.
• A chat app: bundling this inside AppBook would probably be a bad idea, and a dedicated app may be better. AppBook could query the chat backend once in a while, but only as long as it doesn't make the app slower in any way. Bundling it with the website, on the other hand, would make perfect sense. And failing that, an "open in app" link should also work fine.

Should AppBook allow you to synchronize your reminders with the website (and vice-versa) to get notifications across devices? I get the feeling that "yes" would be the best solution (and again, as long as it doesn't make AppBook any slower), but this would significantly impact the complexity of the whole operation. So this remains an open question.

Conclusion

I don't know of any platform offering all of this at once. Most platforms I know offer one feature or the other with varying degrees of success but I haven't yet seen a good implementation of AppBook.

And please, I beg you: few things are as annoying as an app that abuses its dominant position. If you use my login information for crawling my LinkedIn information and "connect me" to other people in my (wrongly scraped!) company, I may not have a recourse right there and then (there's rarely an alternative conference app) but I will make it my goal to ensure that no one uses your app again.