Observe the development.
Understand the behavior.
Artificial intelligence is developing capabilities faster than our ability to fully understand their emergence. We study how intelligent systems learn, adapt, reason, cooperate, fail, and behave through controlled experimentation and systematic evaluation.
Study intelligence
while it develops.
Much of AI safety evaluates systems after significant capabilities already exist. We examine intelligence as a developing system — observing learning trajectories, uncertainty, cooperation, adaptation, and failure before behaviors become deeply embedded.
Development
Study how capabilities and behaviors emerge over time.
Environment
Use controlled simulated environments to observe learning and adaptation.
Measurement
Identify behavioral patterns, uncertainty, and emerging capabilities.
Four programs,
one research agenda.
Our work connects behavioral evaluation, developmental learning, simulated environments, and curriculum design.
Evals
Measure what intelligent systems actually do.
View Evals Research ↗Develop
Study intelligence as a developmental process.
Explore Develop ↗Worlds
Controlled environments for studying intelligent behavior.
Enter Worlds ↗Curriculum
Structure the experiences through which intelligence develops.
View Curriculum ↗What should we measure before
intelligence becomes more capable?
When does a model begin to recognize uncertainty?
01 / 07Evidence before assumption.
Empirical Institute favors measurable evidence over speculation.
Observe
Identify behavior or capability worth understanding.
Isolate
Create controlled experimental conditions.
Test
Expose intelligent systems to structured environments and tasks.
Measure
Collect behavioral and performance evidence.
Understand
Develop hypotheses about why the observed behavior emerged.