Shriyash Upadhyay and Etan Ginsberg, AI researchers hailing from the University of Pennsylvania, assert that numerous prominent AI companies are forsaking foundational research in favor of creating competitive, robust AI models. The pair attributes this trend to market dynamics, where substantial funds raised by companies primarily go towards outpacing competitors rather than delving into fundamental research.
In an email interview with TechCrunch, Upadhyay and Ginsberg expressed their observations during research on Large Language Models (LLMs) at UPenn and highlighted the challenge of making AI research profitable. To address this, they conceived the idea of founding their own company, Martian, with a focus on interpretability research rather than capability research. Martian has now emerged from stealth with $9 million in funding from investors such as NEA, Prosus Ventures, Carya Venture Partners, and General Catalyst. The funds will be allocated to product development, internal operations research of models, and expanding Martian’s 10-employee team.
Martian’s inaugural product is a “model router,” a tool designed to receive prompts for large language models (LLMs), such as GPT-4, and automatically direct them to the “best” LLM. The model router defaults to selecting the LLM with optimal uptime, skill set (e.g., math problem solving), and cost-to-performance ratio for the specific prompt. Upadhyay and Ginsberg explained that the traditional approach involves choosing a single LLM for all requests, but Martian’s approach allows for a team of models to be employed based on the context specified by the user, leading to enhanced performance and reduced costs compared to relying on a single LLM.
Martian’s model router assesses how a model performs without actually executing it, enabling intelligent switching between cheaper and more expensive models based on performance requirements. The ability to incorporate new models seamlessly into applications without manual intervention is highlighted as a distinctive feature. While similar technologies exist, Martian’s effectiveness will depend on factors such as pricing competitiveness and its performance in critical commercial scenarios.
Upadhyay and Ginsberg noted some adoption of Martian’s model router, including among “multibillion-dollar” companies, asserting that their breakthrough lies in developing a deep understanding of how these models fundamentally operate.
