Published on November 16, 2023, 3:31 pm
Shriyash Upadhyay and Etan Ginsberg, AI researchers from the University of Pennsylvania, have raised concerns about the prioritization of competitive AI models over fundamental research in many large AI companies. They believe that when these companies secure substantial funds, most of it is allocated towards staying ahead of rivals rather than exploring the core principles of AI.
During their research on Large Language Models (LLMs) at UPenn, Upadhyay and Ginsberg noticed these trends in the AI industry. They emphasized that the challenge lies in making AI research profitable. To address this issue, they decided to establish their own company, Martian, with a focus on interpretability. By aligning their mission with furthering interpretability research rather than capabilities research, they hope to advance stronger and more meaningful research in the field.
Today, Martian has emerged from stealth mode with $9 million in funding from investors such as NEA, Prosus Ventures, Carya Venture Partners, and General Catalyst. The funds will be used for product development, conducting internal operations research on models, and expanding Martian’s team which currently consists of 10 employees.
Martian’s flagship product is a “model router” designed to enhance language models’ performance by automatically routing prompts to the most suitable LLM. When faced with a prompt for a large language model like GPT-4, the model router selects the LLM based on factors such as uptime, skill set (e.g., math problem solving), and cost-to-performance ratio. This approach allows companies to achieve higher performance and lower costs compared to relying solely on a single LLM.
Using only high-end LLMs like GPT-4 can be expensive for many companies. For example, Permutable.ai revealed that it costs them over $1 million annually to process approximately 2 million articles per day using OpenAI’s top-tier models. Martian aims to provide an intelligent solution by estimating the performance of models without actually running them, enabling cost-effective model routing.
Although Martian’s model router is not a new concept, as another startup called Credal provides a similar automatic model-switching tool, the success of Martian will depend on its competitive pricing and ability to deliver in high-stakes commercial scenarios.
According to Upadhyay and Ginsberg, Martian has already gained traction among multibillion-dollar companies. They believe that building an effective model router is challenging due to the need for a deep understanding of how these models function. Their breakthrough lies in pioneering such an understanding and incorporating it into their product seamlessly.
In conclusion, Martian’s emergence from stealth mode brings hope for prioritizing interpretability research over competitive AI models. By providing an innovative model router tool, Martian aims to help companies achieve higher performance at lower costs by intelligently selecting the most suitable language models for each task. With continued advancements in this field, the future of AI research looks promising.