Autonomous Experimentation Engine for ML
Discovery Loop
The core bet is that experiment execution is becoming a software scaling problem before it becomes a science problem. Discovery Loop starts in machine learning because the whole loop already lives in code, where an agent can edit models, launch runs, score results, and branch into thousands of follow on tests without waiting on people or lab hardware. The founding team also suggests the product is being built from the perspective of large scale ML infrastructure first, then pushed outward into other R&D domains.
-
This is closer to an autonomous researcher for code based R&D than a copiloted notebook. The company describes automating the full loop of proposing, running, examining, and refining experiments, and external coverage says the goal is to run thousands of loops in parallel across areas like model development, chip design, and biology.
-
The strongest near term wedge is ML research because the tools are already standardized. Training jobs, benchmarks, GPUs, and evaluation pipelines can be orchestrated through software APIs. Moving into biology, chemistry, and materials is harder because those domains need physical instruments, sample handling, and lab data systems that are much less uniform.
-
That puts Discovery Loop in a different lane from companies like Lila Sciences and Insilico Medicine. Lila combines AI with autonomous labs across life, chemical, and materials science, while Insilico has built Pharma.AI around drug discovery. Discovery Loop looks more like a general experimentation engine that can prove itself in digital domains before connecting to physical ones.
If this works, the bottleneck in frontier R&D shifts from having ideas to deciding which search spaces are worth exploring with massive automated trial volume. The likely path is software first dominance in ML and simulation, then expansion into lab based science through partnerships, integrations, or owned execution infrastructure that closes the loop in the physical world.