Shriyash Upadhyay and Etan Ginsberg, AI researchers from the University of Pennsylvania, caution that many leading AI firms are prioritizing the development of competitive, advanced AI models over essential research. They attribute this trend to market forces; in many cases, substantial funding is directed more toward maintaining a competitive edge than towards fundamental studies.
“During our investigations into large language models (LLMs) at UPenn, we identified several troubling trends within the AI industry,” Upadhyay and Ginsberg stated in an email interview. “The main challenge lies in making AI research financially viable.”
To address this issue, they conceived the idea of launching their own company, focused on creating products that enhance interpretability. They theorized that a mission centered on interpretability research would naturally lead to more robust findings than one oriented solely towards capabilities research.
That vision has come to life with the launch of Martian, which recently secured $9 million in funding from notable investors including NEA, Prosus Ventures, Carya Venture Partners, and General Catalyst. According to Upadhyay and Ginsberg, these funds will be allocated to product development, research into the internal workings of models, and expanding Martian's team of ten employees.
Martian's inaugural product is a “model router,” which intelligently processes a prompt intended for a large language model—like GPT-4—and seamlessly directs it to the most suitable LLM. By default, this model router selects the LLM based on uptime, relevant skill sets (such as math problem-solving), and the optimal cost-performance ratio for the task at hand.
“The current practice for companies is to rely on a single LLM for each endpoint where they send their requests,” Upadhyay and Ginsberg explained. “However, within a project like website creation, different models can excel based on the contextual details provided by the user, such as language preference, specific features, and budget. By utilizing a diverse team of models within an application, companies can achieve superior performance at a more cost-effective rate than any single LLM could offer.”
This approach holds significant merit. For many companies, depending entirely on a premium LLM like GPT-4 can result in prohibitive costs. The CEO of Permutable.ai, a market intelligence firm, recently disclosed that the company spends over $1 million annually to process approximately 2 million articles daily using OpenAI's top-tier models.
Not every task requires the capabilities of a high-end model, but creating an intelligent switching system can be complex. That’s where Martian shines, utilizing its unique ability to predict model performance without the need for active execution.
“Martian routes requests to more affordable models that offer comparable performance to the costly ones and only engages high-end models when absolutely needed,” they elaborated. “The model router also indexes newly available models, integrating them into applications with minimal friction and no manual intervention.”
While Martian's model router isn't entirely original—Credal, another startup, already provides a similar automatic model-switching solution—its competitiveness will largely hinge on pricing strategies and efficacy in high-stakes commercial environments.
Upadhyay and Ginsberg do note that there has already been interest from significant enterprises, including some “multibillion-dollar” firms. “Creating a truly effective model router is immensely challenging, as it requires a deep understanding of how these models fundamentally operate,” they remarked. “That’s the innovative breakthrough we have achieved.”
Ranking
