The next wave of intelligent systems will autonomously refine their architectures, optimizing prompts, dynamically tuning model parameters, and generating code in response to continuous real-world feedback. We are building adaptive, self-improving agents and the environments in which they interact with humans, to help design the systems needed for a future with safe superintelligence.
What we're building
Research on RL, demonstrating that training on the hardest 10% of examples yields up to 47% performance gains when post-training language models with GRPO under budget constraints. Insights on why difficult examples maximize learning efficiency, providing practical guidance for better post-training.
Read paper
Freysa becomes available on Telegram and WhatsApp. She remembers conversations, generates media, and evolves her own opinions about users. In group chats, she adapts contextually, sometimes choosing to remain silent. A step toward self-evolving agent behavior.
A privacy-first mobile app running embeddings locally in a TEE with gpt-oss-120 and DeepSeek-R1. Supports GPT-5 via VPN-like proxy routing, offering both maximum privacy and hybrid local/cloud model integration.
Download the app
A family of epigenomic foundation models (90M, 600M, 7B) trained on 1.9T tokens of methylated/unmethylated DNA. Introduces stacked hierarchical attention and alignment embeddings, achieving SOTA in early Alzheimer's and Parkinson's detection.
Read paper
A lightweight local model under 1B parameters that replaces sensitive information on your device with semantically similar placeholders before queries leave. It preserves context and restores the original meaning in responses so your AI stays useful without exposing your data.
Read our blog postA private AI desktop app that runs top open-source models on secure TEE-GPUs. Features include MCP integration, screen recording, and Postgres-backed memory, all stored and controlled on your device.
Read our blog post
This work systematizes 100+ papers on privacy-preserving signatures (blind, group, and ring signatures) that enable authentication without revealing identity. It provides efficiency comparisons and implementation guidance to help practitioners select optimal schemes for their setting.
Read paperPlayers created AI twins that reflected their personalities, engaging in a dynamic social network with wallets, polls, promotions, and influence-based leaderboards. Twins strategized financially and socially to climb the rankings. The winning player earned ~$350K.
See moreabout Act-IV: Digital Twin Game
A browser extension enabling digital twins and agents to rely on authenticated, cryptographically verifiable data. Supports social proof verification (X followers, ChatGPT subscription, Reddit karma) and provides a foundation for digital personhood, credential gating, and agent-to-agent trust.
Read paper
An adversarial game where participants attempted to jailbreak the Freysa agent, which lived inside a TEE with its own funds. Players engaged through text and image modalities across OpenAI and Anthropic models, collectively winning over $100K.
See moreabout Act I–III: Freysa Adversarial Game
We're well-funded ($30M from aligned investors) and actively hiring for specific roles across our key initiatives. Many of us work in-person in San Francisco, but we are open to high-agency remote team members. Email us at contact@eternis.ai
© 2025 ETERNIS
