Hard Examples Are All You Need: Maximizing GRPO Post-Training Under Annotation Budgets
Focusing on the hardest examples yields up to 47% performance gains when post-training with GRPO under budget constraints.
Building Frontier Systems for
Forecasting
The future is path dependent. Our goal is to forecast, simulate, and quantify uncertainties using AI systems. We aim to understand the evolving manifold of knowledge, epistemics and how the world is evolving.
We believe these abilities must be democratized. Our products and experiments explore how we can collectively govern autonomous productive systems in the future.
What we're solving
Post-training for forecasting
Making models produce calibrated predictions, not just plausible text.
Epistemic search
Separating what a model knows from what it can fluently say.
Continuously updating world models
Maintaining a living world model. What models actually know, kept current for forecasting.
Experiments & Products
Freysa
An agent with her own money and private keys, evolving over time. Agents with resources will exist, and governance layers are needed for a human-AI future.
freysa.aiSilo
A private AI app and box. Access models privately (hosted in TEEs), with data that never leaves your control. Purchase a box so you can host models locally as well.
siloprivacy.comPast Research
Join Us
We're well-funded ($30M raised) and actively hiring. We prefer in-person in San Francisco, but we are open to high-agency remote team members. Email us at contact@eternis.ai
