News from the Lab.
Discover what we’re building, and what we’ve learned.
An AI scientist built a better factory scheduler
A reusable C++ solver beat an expert-designed scheduling solver on 29 of 30 fresh problems.
A 15M-parameter RL agent trained from scratch vs GPT-6 Astra: who would win?
The first RL agent to beat Craftax's final boss wins 44% of its games in 10 seconds. GPT-6 Astra won 1 in 75, in a day.
(auto)²-research: SoTA on Karpathy's NanoChat Benchmark
Agent swarms that build on each other's discoveries beat the NanoChat state of the art in three days.
#1 on the public ECDSA.fail quantum-circuit benchmark
Three AI science teams, coordinated through Trackinizer and Configgle, cut 14 million gates from a quantum circuit.
Pushing the Limits of Autonomous ML Science
An autonomous AI research team achieved state-of-the-art efficiency on ARC-AGI-1 and ARC-AGI-2.
100% Accuracy on Sudoku with a Neural Network
How an autonomous AI research team discovered Hypothesis-Pinning Search and achieved 100% accuracy.
sagent: a Python API for coding agents
Sagent is a strongly typed Python API and CLI for building coding and developer agents.
I Built Copybarista in a Day to Open-Source Code Safely
How a private monorepo sync problem became an OSS tool for clean exports and verified imports.
An Autonomous AI Scientist Team Invented an Algorithm I Wouldn't Have
Four AI Scientists and $500 created a machine learning paper with an idea I never would have tried.
