Martian

Martian is an AI research lab focused on understanding machine intelligence, developing model routing and AI interpretability software tools. The company was founded by researchers who have worked on Google DeepMind's early scientific initiatives, developed Anthropic's interpretability efforts, and commercialized superintelligence at Meta. Martian's mission centers on the belief that understanding machine intelligence is the central scientific problem of our time and a prerequisite for building systems that can be trusted with increasing autonomy. The lab works across three areas: Measurement (building rigorous tools to study model behavior, described as telescopes for models), Explanation (developing theory through feature geometry, long-horizon interpretability, and the mathematics of what models learn), and Application (commercializing research through products that scale with global LLM usage). Martian's products and research artifacts include ARES (online reinforcement learning for coding agents), Code Review Bench (beyond static mechanistic interpretability), and K-Steering (interpretability challenge framework). The company provides a model gateway for accessing various language models, tools for evaluating model performance, and routing prompts to appropriate language models. The name Martian references both the alien nature of AI models and the Hungarian emigre scientists of the 20th century (Teller, Szilard, von Neumann, Erdos) who were called the Martians for their inexplicable intelligence. Founded in 2022 and headquartered in San Francisco, California, Martian has raised $9.0M in Series A funding from investors including Accenture, New Enterprise Associates, Prosus, Carya, General Catalyst, and PROSPER.

AI Infrastructure > Enabling Technologies > Software > ML Orchestration

College Wise > Wharton School, University of Pennsylvania