Autoresearch Hackathon
Sign-up via either partiful or Luma is sufficient.
LLMs are now improving LLMs. Welcome to Autoresearch.
The Context
Andrej Karpathy just dropped the ultimate “early singularity” playground: "autoresearch"; promptly described by Musk as "singularity". He's using a single GPU and an agent that iteratively improves the train.py file: architecture, optimizer, attention, hyperparameters — anything. After training for 5 minutes (wall-clock, reproducible on any NVIDIA GPU), it evaluates on val_bpb (validation bits-per-byte). It keeps the wins, discards the flops, commits, and repeats a hundred times overnight. You wake up to a results.tsv log of real science — and often a noticeably better model. Now, it's your turn to push the frontier.
The Hackathon
At this one-day technical hackathon, we’re turning Karpathy’s minimal 630-line framework into the next wave of autonomous AI research tools. What discoveries can you make? Can you optimize open-source kernels? In what other domains can agents run iteratively improving experiments? Prizes: To Be Announced
Schedule
11:00 Arrive
11:30 Talks
12:30 Hacker Lightning Talks
12:45 Hacking Begins!
1:00 Lunch Served
5:00 Dinner
6:00 Demos & Presentations
7:00 Judge Adjudication + celebrate our winners & adjourn!
Sponsors
Thank you to Hyperbolic AI [https://www.hyperbolic.ai/] for providing the high-performance compute infrastructure and enabling scalable and efficient AI processing through their cloud platform! Ad Astra!