We just wrapped up a sprint on the ARC Africa challenge. Big thanks to the organizers for putting together a competition that brought me back to Zindi after a looooooong break. It was a refreshing opportunity to explore this problem in a structured yet fast-paced setting. Congratulations to the top scorers !!!! I'm looking forward to learning from your solutions.

My main goal was to see how far we could get using relatively low-cost, rule- or prompt-based approaches without relying on models specifically trained or fine-tuned on ARC-like data. I also wanted to secure a spot in the #Top10 by the end date, so it was really fun to see how it happened at the very last minute.

Thoughts

I couldn’t fully test my hypothesis due to time and compute constraints, but I remain convinced that solving ARC should not depend on prior exposure to similar problems. I also strongly believe that ARC-AGI isn’t really an ML problem in the traditional sense. Some may have noticed LLMs often perform worse than a simple input-copy baseline. In this context, EDA proved once again to be the most effective way to guide strategy: understanding the structure of the puzzles and evaluation quirks helped define meaningful (and cheap) heuristics early on. Experimenting with local LLM runs via MLX models was particulary exciting to me.

Perspectives

While I may not secure the flight to Kigali (I really need to find a way to sponsor that trip 😂), the challenge sparked important reflections: If we want to push AI research meaningfully from and for Africa, we need broader access to affordable, modern compute.

Full write-up and repo coming soon. Happy to connect with others who want to explore the ARC-AGI space further...especially in a setup that respects its original principles.