Unlocking AI self-improvement at production scale

Effective AI systems need the ability to learn from experience:
that's the missing piece in deploying truly productive agents.

Osmosis enables AI self-improvement through real-time reinforcement learning.

About

Osmosis was founded in 2024 by Kasey and Andy, and is based in San Francisco. Kasey is a repeat founder who sold his first company in undergrad, while Andy was the youngest tech lead at TikTok.

We are backed by Y Combinator, and recently raised an oversubscribed seed round led by tier-one institutional investors and angels (like Paul Graham).

We are currently hiring engineers. If you're interested in learning more, please reach out here.

Early Access

Writing

May 20, 2025

LoRA's Limitations: Head-to-Head with Full RL

Low-Rank Adaptation (LoRA) has recently gained popularity due to its significantly lower computational cost compared to full-model post-training, while still achieving comparable performance on supervised fine tuning workloads. However, our experiments indicate that LoRA is less effective in reinfor

Release

May 8, 2025

Applying RL: Open Source SLM Trained for MCP

MCP is quickly becoming the open standard for AI agents, and for good reason! It’s well designed and easy to use — however, MCP implementations are limited since: 1. The best models are large and closed-source (3.7 Sonnet, Gemini 2.5 Pro) 2. It introduces tool sprawl (numerous MCP clients, MCP serve

View all posts