# Martian Martian started out known for commercial model routers, and it now presents itself as a research organization working on understanding machine intelligence. The team includes researchers who came from Google DeepMind, [[Anthropic]] and Meta. Their commercial routing lineage continues under a separate initiative, Thesean AI. I keep the note because the framing is worth stealing, well beyond routing. ## Their argument They use a bridge analogy. You do not need to understand physics to build a bridge, and you do not need to understand neural networks to build one either. But you will never build a steam engine by randomly searching the space of bridges. Applied to AI, that means trial-and-error scaling eventually runs out of road. If you want systems you can actually trust to run autonomously, you need theory, not just better benchmarks. ## How they organize the work - **Measurement**: tools for studying model behavior rigorously (ARES, Code Review Bench) - **Explanation**: feature geometry and interpretability theory (K-Steering, interpretability challenges) - **Application**: commercializing the research through products rather than consulting That last point is the one I find most interesting as a business observation. Plenty of research shops drift into consulting because it pays immediately. Choosing products instead is a bet that the understanding compounds into something ownable. ## Why it connects back to routing Predicting which model will handle a given input well, without running every model first, is an interpretability problem wearing a cost-optimization costume. [[Not Diamond]] and every serious router are solving a version of the same question Martian decided to attack directly. ## References - Official website: https://withmartian.com/ - Model routing reading list (community): https://github.com/Not-Diamond/awesome-ai-model-routing ## Related - [[Model routing]] - [[Not Diamond]] - [[Anthropic]] - [[Large Language Models (LLMs)]] - [[AI Agents]]