Notes on autonomous research, intelligence, and the systems that turn it into cumulative knowledge.
Paradigma's first model, built for high-throughput mathematical reasoning.
A 1-billion parameter transformer trained from scratch for mathematical reasoning, rivalling models tens of times larger on competition-level math. Released with its evaluations, a training-time value model, and a custom vLLM inference plugin.
How results could be shared in the new age of research, using the Riemann Hypothesis as a working example.
As models get better, everyone is tempted to spend their compute on the big open problems. We are still figuring out how to collect and share that work so others can build on it. Here is a prototype of what we envision, with a result on Gröchenig's Riemann hypothesis criterion shared as a Flywheel graph.
How the Via Panisperna Fellows set records across four frontier AI training challenges.
The ideas behind our work on the Modded-NanoGPT Main and Optimization Tracks, OpenAI's Parameter Golf, and the SlowRun Tiny Track—and what the process taught us about collaborating with coding agents.
Pointing intelligence at the edge of the unknown.
Every hard problem humanity has not solved yet is, at its core, a problem of insufficient intelligence applied for insufficient time. This essay argues that if intelligence becomes scalable, the infrastructure that organizes its output becomes load-bearing.