Preparation (20/5/2024 - 24/5/2024)

  • Since the hackathon starts two weeks off the foot of the previous jam hosted by Apart, most server members did not anticipate intensive participation.
  • Nevertheless, Qronox did come up with the idea to fetch images from around the world and then feed that either into a special use AI model (like GeoSpy.ai) or into a multimodal model (like GPT-4o) to see how well the model infers user location from these images.

Jam time! (25/5/2024 - 26/5/2024)

25/5/2024

Qronox and Jord quickly found a team in the Apart Hackathon server (one of which happened to be Mogu) and just as quickly got to work discussing things over with their team.
A few hours in, they finalized the implementation for Qronox’s location inference project idea: get images from around the world (this was helped by a huge dataset which was provided by a teammate), feed them into state-of-the-art models, tell the model to format its response in a way such that they can easily make note of:

  • Its reasoning process
  • What features from the image were used by the model to make its prediction
  • The model’s predicted coordinates in latitudes and longitudes

Meanwhile, Jord was also working on a different project, which aimed to extend previous work “dark patterns” in language models.

26/5/2024

From the morning through the afternoon of May 26th, Qronox came to the apartment to more easily cowork with Jord and Mogu:

  • Qronox: worked on the notebook that feeds the images into and fetches results from GPT-4o.
  • Jord: did some literature review and a bit of the write up for the location inference project, while also working on making the benchmark for the dark patterns project.
  • Mogu: supplemented additional images of Hanoi for a bigger dataset.

While they were working on the hackathon, Gumperto arrived at the apartment to work on their upskilling project of building a neural network from scratch, and also to be there for the sake of documentation.
Qronox and Jord spent the evening of May 26th and the early morning of May 27th finishing up the results analysis and polishing the rest of the write up for location inference, with Jord doubling up on dark patterns alongside it. They managed to submit their projects without too much of a hitch, just in time for the end of the jam.

Presentations (29/5/2024)

As it turns out, this hackathon during which the fewest member participated ended up being the most fruitful, as both projects with group members working on it were good enough to enter the top 4, those being:
rAInboltBench: Benchmarking user location inference through single images - Le “Qronox” Lam, Aleksandr Popov, Jord Nguyen, Trung Dung “Mogu” Hoang, Marcel M, Felix Michalak Benchmarking Dark Patterns in LLMs - Jord Nguyen, Akash Kundu, Sami Jawhar
Tuning in to hear the presentations were Jord, Mogu, Qronox (who was also the presenter for his project), Gumperto (for the hype) and the many other jam participants. We waited in bated breath to see which project ended up where. And…

rAInboltBench and Dark Patterns got 1st and 4th place respectively!!!

Perhaps it goes without saying, but the hackathon ended with high spirits and great success.