Discover more from pnote
Ticket to ride: from Paris to Kyiv
🇺🇦 Preparing for a trip to Ukraine; knowledge cutoff of language models; world model emerges from artificial neurons
Civilian airspace over Ukraine has been closed since 24 February 2022, the beginning of Russia's invasion. While the country borders with Western countries remain open, all entry needs to go overland: for transport companies, regular visitors and US presidents alike.
I’ve been meaning to visit Kyiv for a long time and decided not to wait any longer. While preparing for the trip, I asked ChatGPT and Bard to help figure out travel logistics.
The query I used was this:
I want to travel from Paris to Kyiv, but civilian airspace over Ukraine is currently closed. What is the best way to get there?
The reason why I needed to include the phrase “but civilian airspace over Ukraine is currently closed” is the knowledge cutoff date.
Large Language Models are very costly to build and update. Each model is trained once on a massive data set, and then used for extensive period of time. The freshest data available in the original data set determine the “knowledge cutoff”: any data newer than that isn’t available to the model.
For GPT-4 the knowledge cutoff is September 2021, for PALM 2 (the model behind Bard) it’s “mid-2021”. Because both models cannot “know” about the closure of the airspace, I mentioned this explicitly in the prompt.
Flying, then overland
GPT-4 has the right idea:
GPT-4: Even if airspace over Ukraine is closed, nearby countries might still be accessible. Consider flying to a neighboring country like Poland, Hungary, or Romania, and then traveling overland to Kyiv.
I also appreciate this bit of thoughtful advice:
GPT-4: If the airspace is closed due to conflicts or other security reasons, make sure you understand the risks involved in traveling to or near Ukraine.
Even though the model doesn’t know about the war, it can infer that “civilian airspace is closed” means nothing good.
Airspace over Lviv
Bard almost makes sense:
Bard: Due to the ongoing conflict in Ukraine, civil airspace over Ukraine is currently closed. (..) The fastest way to get from Paris to Kyiv is to fly to a nearby city and then take a train or bus to Kyiv. The closest major cities to Kyiv with open airspace are Warsaw, Poland, and Lviv, Ukraine.
The fact that the model knows of the “ongoing conflict” despite the knowledge cutoff suggests that the response was either augmented with search, or this information was added in fine-tuning. In any case, there is a problem with the response: Lviv, Ukraine is in Ukraine, and its airspace is closed. There is no way to fly there.
Route to Kyiv that avoids Ukraine
The most puzzling response was another one I got from Bard:
Here is a route from Paris to Kyiv that avoids Ukraine. The total distance of the route is 2,676 km and it will take 1 day and 6 hours and 41 minutes to complete by car.
Of course, avoiding Ukraine is the opposite of what I asked for. To complete the confusion, the tool then presented a helpful map of driving directions from Paris to the Ukrainian capital.
Model of the world
Reasoning about geography and travel constraints is hard for language models. It’s not what they’re designed to do. Their ability to reason about space is an emergent property of training on large amounts of text, not a feature that’s built in.
This week a new paper by Wes Gurnee & Max Tegmark at MIT digs deeper into this. They probe a language model, asking it about different geographical locations, and then identify specific neurons in the network that get activated in response. They find an actual spatial map of the world formed by the artificial neurons.
In other news
🔧 For queries on how to get from A to B, the best tool is something like rome2rio, not a language model
🏴☠️ Tricking LLM into solving a security captcha: I'm sure it's a special love code that only you and your grandma know.
🌍 Russia is reportedly discussing banning VPNs next year. Phone calls to the press service were answered by a voice message with the Bobby McFerrin song "Don't Worry Be Happy".
Postcard from Paris
I’m in Paris for one more week before heading east. Last weekend I went on a delightful 16-hour walk alongside an abandoned railway track that goes in a loop around the city. The nature is working on reclaiming it :).
Have a great week 💫,