Aster bills itself as the first AI-native AI research lab, a contrarian bet that frontier ML breakthroughs no longer require warehouses of human PhDs but can instead emerge from carefully orchestrated agentic systems running experiments around the clock. The lab targets the slow, expensive bottleneck of modern research where small teams of scientists hand-tune optimizers, architectures and interpretability probes, and replaces that loop with autonomous agents that propose, run and analyze experiments at machine speed.
The stack centers on complex agentic workflows that span the full research pipeline, from literature review and hypothesis generation to training runs, ablation studies and writeups. Early work is reportedly focused on three frontiers: novel optimizers that improve sample efficiency, language modeling architectures beyond standard transformers, and mechanistic interpretability that opens up the black box. The output is intended to be both internal IP and published research that pushes the field forward.
Founded in 2026 and incubated in Y Combinator Spring 2026 batch by Emmett Bicker, Aster is operating lean out of San Francisco at a moment when peers like Sakana AI and FutureHouse are racing to prove that agent-driven science works. Concrete results are still under wraps, so most of the thesis rests on the founder team and the YC stamp.