Martian Invents Model Router that Beats GPT-4 by Using Breakthrough “Model Mapping” Interpretability Technique; Receives $9 Million in Funding from Top VCs
Share This Article
Backed by NEA, Prosus Ventures, CVP, and General Catalyst; Martian, the first AI model-routing platform, emerges from stealth
SAN FRANCISCO, Calif., November 15, 2023-- Today Martian emerges from stealth with the Model Router, an orchestration layer solution that routes each individual query to the best LLM in real-time. Through routing, Martian achieves higher performance and lower cost than any individual provider, including GPT-4. The system is built on the company's unique Model Mapping technology that unpacks LLMs from complex black boxes into a more interpretable architecture, making it the first commercial application of mechanistic interpretability.
In addition, the company is announcing a $9 million seed round of funding from NEA, Prosus Ventures, Carya Venture Partners, and General Catalyst.
"All the effort being put into AI development is wasted if it's unwieldy, cost-prohibitive and uncharted for enterprise and everyday users," said Aaron Jacobson, partner, NEA. "We believe Martian will unlock the power of AI for companies and people en masse. 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."
"Our goal is to consistently deliver such breakthroughs until AI is fully understood and we have a theory of machine intelligence as robust as our theories of logic or calculus," Shriyash Upadhyay, co-founder, Martian, said.
Martian understands LLMs through a technique they pioneered called "Model Mapping." The software takes AI models (essentially complicated black boxes) and converts them into other forms that are simpler and easier to study. For example, AI models might be converted from numbers in matrices to human-readable programs. This lets Martian understand models more deeply than was previously possible, and allows for tools to be built based around that understanding.
Martian's Model Router
Martian's Model Router is the first tool built using Model Mapping.
AI development is moving at breakneck speed-there are 380,000 open-source AI models today (up from 80,000 less than a year ago), and countless closed-source models-creating a unique opportunity for companies to optimize their AI systems by using the right models.
Instead of using a single model, companies can think about these models like they would think about employees. A company wouldn't send junior work to a senior employee, or send scientific work to the finance department. By using a team of models in an application, a company can achieve a higher performance and lower cost than any single LLM could achieve alone.
"Routing between models is fundamentally about understanding how they work: what makes them succeed or fail. The better one understands these models, the more effectively you can route between them," Etan Ginsberg, Martian co-founder, said.
"For companies using AI, it's difficult to understand the precise strengths and weaknesses of every single AI model; you would have to hire an entire AI/ML engineering team working around the clock just to do that. And it's not enough to select just one model for each task: even within a single task (e.g. summarization or coding), different models will do better on different inputs," Ginsberg continued.
"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."
The Martian Model Router benefits enterprises through:
• Increased Performance - Existing evaluation tools review overall model performance so you can choose which single model to use. Even if a single model is best on average, it will still achieve worse performance than if different models were selected dynamically for each request. This enables Martian to outperform any single LLM on key benchmarks.
• Reduced Total Cost of Ownership - GPT-4 is 30 times more expensive than GPT-3.5 Turbo and 900 times more expensive than specialized industry-specific models like DeciCoder. Martian can route to cheaper models on requests that perform similarly to the most expensive models, and only route to expensive models when necessary. In addition, testing all the models available on the market means hiring an entire ML engineering team. Martian can automate the whole process, leading to savings in testing and in production.
• Futureproofing - Continually optimizing and dynamically updating models. With new models being introduced every week (e.g. GPT-4 Turbo last week) companies risk falling behind. The Model Router indexes new models as they come out, incorporating them into applications with zero friction or manual work needed.
"LLM choice is a growing question for developers building scalable AI applications. Shriyash and Etan are among the best founders we've met globally in AI. They are offering a scalable solution to this question," said Javier Valverde, investor, Prosus Ventures.
"Enterprises are in search of platforms that empower them to effectively leverage AI models. From choosing the right model to optimizing costs and performance, and governing the usage of new AI models - all of these components are key to getting AI models working smoothly in today's software. Businesses are racing to embrace AI, but will require specialized developer platforms," said Quentin Clark, Managing Director of General Catalyst. "GC is proud to support ambitious entrepreneurs like Etan and Shriyash that are working toward building just that."
Martian is a technology company on a mission to understand AI and build AI tools based on a central theory of machine intelligence. The company's Model Mapping technology unpacks blackbox LLMs to uncover the mechanisms that make models work and underlies Martian's first product for enterprises and startups: the Model Router API. The company has raised $9 million in a seed round of funding from NEA, Prosus Ventures, CVP, and General Catalyst. Martian co-founders Shriyash Upadhyay and Etan Ginsberg have sold their previous NLP companies and conducted LLM research at the University of Pennsylvania since 2020. For more information, visit www.withmartian.com.
Martian is hiring for senior ML Engineering and Senior Backend Engineering roles, which can be viewed here: work.withmartian.com. To partner with Martian, reach out to firstname.lastname@example.org.
Hot Tomato Marketing