Martian Learning Inc. today debuted a new artificial intelligence model mapping technology that it says can achieve better performance at lower costs than any single large language model.

The startup said it has also raised $9 million in a seed funding round led by NEA, Prosus Ventures, Carya Venture Partners and General Catalyst.

The company explained that its Model Router offering can be integrated with AI applications to route individual queries to the most suitable LLM each time. It works by translating LLMs from complex black boxes into a more interpretable architecture.

The Model Mapping technology is designed to understand LLMs by converting them into formats that are easier to study and understand. For instance, an LLM might be converted from numbers in matrices to human-readable programs in order to understand how it works. Companies can then build applications based around that understanding.

Martian co-founder Shriyash Upadhyay explained that the company’s goal is to ensure that AI is fully understood. He added that it intends to develop “a theory of machine intelligence as robust as our theories of logic or calculus.”

Martian’s first product is the Model Router, which uses Model Mapping to help companies optimize their AI systems by choosing the most appropriate LLMs. So rather than using a single model to power an application, companies can use multiple LLMs and leverage the best capabilities of each one. By using what is effectively a team of LLMs, applications will achieve higher performance at lower costs than if they were based on a single LLM, the startup claims.

Etan Ginsberg, Martian’s other co-founder, said model routing is fundamentally about understanding how LLMs work and what makes them succeed or fail. “The better one understands these models, the more effectively you can route between them,” he explained.

According to Ginsberg, many companies struggle to understand the precise strengths and weaknesses of each AI model. Moreover, he said, it’s inefficient to select just one LLM to power an AI application, because each of them works very differently. “We make every LLM more useful by giving companies a way of benefiting from all LLMs,” Ginsberg added. “In our tests, we have been able to outperform GPT-4 on OpenAI’s own model evaluation suite, openai/evals, by routing between several different models.”

Martian said its Model Router helps enterprises to achieve superior performance by reviewing the performance of individual LLMs to help decide which one is best to use for each kind of query. This is important, it said, because there are more than 300,000 open-source AI models for developers to choose from, and even if one LLM is better on average, an application based on that model won’t perform as well as one that can choose from different LLMs dynamically, according to the user’s request.

The startup said Model Router also helps achieve a reduction in total cost of ownership. It points out that queries with OpenAI LP’s most advanced LLM, GPT-4, costs 30-times more than queries with GPT-3.5.

Moreover, GPT-4 can cost up to 900-times more than specialized, industry-specific models such as DeciCoder. By routing to cheaper models that perform as well as the most expensive LLMs with certain queries, companies have a way to reduce the costs of their AI applications. Of course, they can still rely on the most advanced LLMs when the query demands it.

Lastly, Martian said Model Router provides a way for companies to future-proof their applications by taking advantage of the capabilities of any new LLMs that are released. It said Model Router indexes new LLMs as soon as they’re released so it can incorporate them into existing applications with minimal friction.

Martian is one of a number of startups trying to make an impact in the area of mapping requests to the right LLM, based on performance, cost and availability, said Andy Thurai, vice president and principal analyst of Constellation Research Inc. He said that many enterprises have already done a lot of work in this area themselves, building in-house models to decide which LLM to use based on the nature of the user’s request.

So it’s not an entirely novel technology, but what Martian offers is the opportunity for enterprises to rely on a third-party solution that can also evaluate and keep up to date on all new LLMs. “The problem it faces is that it’s asking enterprises to trust its decision-making on which LLM to use, rather than trusting their own internal research,” Thurai said. “It might be a hard battle to win, but there is a market for it.”

Thurai said Martian and other startups will also be challenged to keep their routing tables and performance metrics up to date. This needs to be done on a daily basis, he said, as new LLMs become available almost every day. “If enterprises are ready to offload this kind of work to outsiders and decide to trust them, Martian’s Model Router could be a good solution,” Thurai said. “But many enterprises have put a lot of effort into evaluating the LLMs they used, based on factors such as trust score, reliability, accuracy, hallucinations, security and governance, so not everyone will see it as a good alternative.”

Martian’s venture capital backers were more confident about the demand for its Model Router service, though. Javier Valverde of Prosus Ventures said choosing the right LLM is becoming a major headache for application developers. “Shriyash and Etan are among the best founders we’ve met globally in AI, and they are offering a scalable solution to this question.”

NEA partner Aaron Jacobson said one of the challenges of AI development is that it’s incredibly unwieldy and cost-prohibitive. “We believe Martian will unlock the power of AI for companies and people en masse,” he said. “Etan and Shriyash have demonstrated entrepreneurial spirit in their prior experiences and deep expertise in this field through high-impact peer-reviewed research that they’ve been doing since 2016.”

Image: Martian