Reific

RL environments for superhuman reasoning over time

[Antanas Zilinskas @Antanas_Z]

Launching @reific (backed by @ycombinator): RL environments for superhuman reasoning over time.

AI has learned to reason over text and code. But most of what matters in the real world shows up as numbers that change every second: a heartbeat, a bioreactor, a GPU cluster. No human can track thousands of signals at once, but AI should be able to.

In our environments, data streams in on a clock while the agent works, and every action is graded by what actually happened. Most AI training today teaches models to match human experts. We treat human performance as the floor.

Today, even the best frontier model averages under 50% on each of our three task families: personal agents, industry, and science & compute. Training an open model on our environments raised its full-solve rate on unseen tasks by 54% (from 25.6% to 39.5%), with no loss on general benchmarks.

I’m a published time-series researcher who simulated superconductors at CERN and shipped AI to hyperscalers as co-founder and CTO of Entangl (YC S24). I’ve seen firsthand how models that ace text and code fall apart on live signals, so I started Reific to build the infrastructure that keeps AI climbing the frontier, long past human level in the real world.

If you train agents, try to break one of our environments: https://t.co/C4k49v5vhH