Shreyash Upadhyay and Etan Ginsberg, AI researchers from the University of Pennsylvania, argue that many large AI companies are sacrificing basic research in their quest to develop powerful, competitive AI models. This duo blames market dynamics: When companies raise big money, the majority usually goes toward efforts to stay ahead of competitors rather than studying fundamentals.
“During our research on LLMs [at UPenn,] “We’ve noticed these troubling trends in the AI industry,” Upadhyay and Ginsberg told TechCrunch in an email interview. “The challenge is to make AI research profitable.”
Upadhyay and Ginsberg thought the best way to address this problem might be by starting a company of their own—one whose products benefited from explainability. Naturally, the company’s mission aligns with promoting explainability research rather than capabilities research, as they assumed, leading to stronger research.
that company, Martian, today emerged from stealth with $9 million in funding from investors including NEA, Prosus Ventures, Carya Venture Partners, and General Catalyst. Proceeds are directed toward product development, research into the models’ internal operations and growing Martian’s team of 10 employees, Upadhyay and Ginsburg say.
Martian’s first product is a “model router”, a tool that takes a router dedicated to a large language model (LLM) – for example GPT-4 – and automatically routes it to the “best” LLM. By default, a typical router selects the LLM that has the best uptime, skill set (such as mathematical problem solving), and cost-performance ratio for the router in question.
“The way companies currently use LLMs is to choose one LLM for each endpoint to which they send all their applications,” Upadhyay and Ginsberg said. “But within a task like building a website, different forms will be more appropriate for a specific request depending on the context the user specifies (what language, what features, how much they are willing to pay, etc.)…using a team of forms in the app , the company can achieve higher performance and lower cost than any MBA could achieve alone.
There is truth to that. Relying exclusively on a premium MBA like GPT-4 may be too expensive for some, if not most, companies. Recently CEO of Permutable.ai, a market intelligence company open It costs the company more than $1 million a year to process about 2 million articles a day using cutting-edge OpenAI models.
Not every task needs the horsepower of pricier models, but it can be difficult to create a system that switches intelligently on the fly. This is where Martian comes in – its ability to estimate a model’s performance without actually running it.
“Martian can pivot to cheaper models based on requests that perform similarly to more expensive models, and only pivot to more expensive models when necessary,” they added. “The model router indexes new models as they appear, integrating them into applications without requiring any friction or manual work.”
Now, the Martian’s modular router is not new technology. At least one other startup, Credal, offers an automated model-switching tool. So its rise will depend on the competitiveness of Martian’s pricing – and its ability to execute in high-risk trading scenarios.
Upadhyay and Ginsburg claim there has already been some take-up, including among “billion-dollar” companies.
“Building a truly effective modular router is very difficult because it requires developing an understanding of how these models fundamentally work,” they said. “This is the breakthrough we pioneered.”