An AI research lab building the tools, evaluations and environments that enable models to accelerate discovery in the empirical sciences.
Capability breakthroughs in the empirical sciences will come from novel, scalable ways to generate data and verify outcomes.
Evaluating agents at autonomous ML research and development. Results coming soon.
frontierml.ai ↗The open-source AI workbench for scientific research: literature, code, experiments, compute and results in one workspace.
openscience.sh ↗Infrastructure for autonomous, open science: the record agents write to and the library they search.
tryatlas.sh ↗The platform for frontier AI research: agents that plan experiments, run them across your compute and report back.
tryascent.ai ↗Synthetic Sciences is an AI research lab building infrastructure for scientific superintelligence.
AI systems are already producing breakthroughs in mathematics, with new discoveries arriving in rapid succession. Mathematics has fast, trustworthy verification, and research there needs little more than a model and a proof checker. The empirical sciences have neither, which leaves two problems.
The first is intelligencein·tel·li·gence nounthe ability to acquire and apply knowledge and skills.. Experiments are slow, expensive and approximate, so models can’t brute-force the solution space. The best scientists make progress anyway. What sets them apart is experimental taste: knowing which experiments are worth running, anticipating what they will show and understanding what the results mean. We think taste can be formalized, measured and trained.
The second is infrastructurein·fra·struc·ture nounthe basic physical and organizational structures and facilities (e.g. buildings, roads, power supplies) needed for the operation of a society or enterprise.. A model with taste still needs compute, instruments, simulators, often physical facilities, and a harness to operate them. Almost none of it is built for autonomous agents.
We work on both, starting with long-horizon evaluations of whether models can form hypotheses, run experiments, interpret failures and make progress under resource constraints. We begin with machine learning, where compute is the lab, and then extend to fields that need physical facilities, building the infrastructure they lack. We’re also working out how to train these capabilities at scale.
If you’d like your models to be better at research, write to founders@syntheticsciences.ai.
We’re a small, talent-dense team of International Olympiad medalists and AI researchers, backed by Y Combinator, Z Fellows, Charlie Songhurst, Walter Kortschak, Edward Lando, scout funds from Sequoia and a16z, and others. We’re hiring exceptional researchers and engineers: team@syntheticsciences.ai.