My work on ARC-AGI
Public stuff
I built a tiny transformer that scores 44% on ARC-AGI, trained from scratch in just 67 cents
- This is the best non-LLM method in the world
- Also the cheapest, and trains the fastest by far (new pareto frontier)
- Beats every single non-thinking LLM and many thinking ones
- Least # of params at this performance
- Went viral on X, got the attention of top AI researchers (many thought the result to be impossible initially)
Posts on X: https://x.com/evilmathkid/highlights
Blogs on the work: 44% on ARC-AGI 1
A New Pareto Frontier on ARC-AGI
Private stuff
Since then, I have doubled the score again and saturated v1.
Now working on v2 of the benchmark, currently at 21% (beats GPT-5 Pro).
I expect to beat the WR (40%) in 2 weeks.
Winner of v2 gets $425K (spoiler: its gonna be me)
Context: ARC-AGI is a famous metalearning benchmark. Easy for humans, hard for AI models, including LLMs. The benchmark’s been open since 2019 and despite a heavy cash prize, no substantial progress was made till thinking LLMs in 2024. I showed that vanilla transformers were good enough to make the same progress at OOMs less effort
Achievements
- Invited to Team India selection & training camp (International Olympiad in Astronomy & Astrophysics)
- NIUS fellowship by Tata Institute of Fundamental research, 2020-21
- Qualified Indian National Astronomy Olympiad 2018 (top 25 in country)
- KVPY fellowship twice 2018, 2019
Work Ex
1. Founding team, GetCrux (YC W24)
Led the project on automated insights for performance marketers. Was a backend dev and product manager on other projects
2. Intern, New ventures arm @ Tata
Evaluated business models / industry research for Deeptech investments. Regularly presented to senior leadership of Tata Industries and other companies
3. Research Intern, University of Paris Saclay
Found a correlation b/w DNA content & light absorption while analysing UV images of viruses. Did a bunch of physics experiments on RNA/protein mixtures
Misc stuff
- 1500 rank in JEE Advanced exam, out of 0.9 million starting candidates (did this without studying much)
- Institute student rep - Academic Rehabilitation committee, IIT Bombay. Recommended by Institute Deans to lead a 350+ member team
- Top 300 all India, National Standard Exam in Astronomy
- Runner up, Bluemix teen hackathon (I was in 9th grade, everyone else in 12th grade)