When Emmett Bicker was 18, he moved to San Francisco to train language models at Magic. The work gradually pulled him toward a bigger question: could AI help do the research itself—not just answer questions researchers had already framed? He left to start Aster, a small SF team building an autonomous research system. The problem is partly speed, but also scope. A research team can spend hours or days exploring possible approaches, and fixed benchmarks capture only a narrow slice of scientific work. Aster’s system breaks a goal into research directions, sends AI agents to investigate them in parallel, and has planner agents coordinate the work. The agents can propose ideas, design experiments, run code, and return results; the company’s own writeups also describe manual intervention when experiments go off track. One striking example: Aster reports that about 1,000 concurrent model calls raised a ProteinGym score from 0.507 to 0.524 in 30 minutes. The winning change was not a new protein model, but a different way to score mutations using the model’s confidence. That’s a useful, concrete glimpse—not proof that open-ended science is solved. The harder question is whether systems like this can reliably find ideas worth testing beyond a scoreboard. If they can, researchers may spend less time waiting for searches to finish and more time deciding what the evidence actually means. asterlab.ai/