Our blog

Resources and insights

Scaling AI Interpretability

Anthropic and OpenAI recently released groundbreaking mechanistic interpretability work on frontier models, using Sparse AutoEncoders (SAEs) at scale. Martian's research has uncovered why these methods are effective, leveraging category theory to understand models without manual inspection. This approach not only validates SAEs but also opens the door to other scalable interpretability methods, which Martian is exploring further.
May 31, 2024
min read

AI Safety vs Capitalism

The recent exit of safety researchers from OpenAI underscores a troubling shift among AI giants from prioritizing safety to focusing on enhancing capabilities, potentially compromising AI's safe development. As AI companies, driven by the need to remain competitive, increasingly prioritize advancing model capabilities, they neglect crucial interpretability research, creating a misalignment of incentives that could pose risks. To counter this trend, companies like Martian are emerging with business strategies that emphasize understanding and safety in AI development, aiming to realign the industry towards a more secure and interpretable AI ecosystem.
Chaithanya Bandi
May 24, 2024
min read

The Sustainability Challenge of AI: Tackling the Energy Footprint of LLMs

The rising energy consumption of large language models (LLMs) threatens the sustainability and scalability of AI systems. Martian's "model routing" technique and industry collaborations like the Green Software Foundation offer promising solutions to reduce costs and emissions. The AI community must prioritize sustainability alongside performance to ensure the benefits of AI are realized while minimizing its carbon footprint.
Etan Ginsberg
May 17, 2024
5
min read

Model Mapping: The Key to AI Alignment and Beyond

Model mapping, a novel approach towards mechanistic interpretability, transforms opaque neural networks into transparent, verifiable programs. This approach has multiple benefits: a way to measure AI alignment, improve model efficiency, adaptability, and human-AI interaction.
Chaithanya Bandi
May 10, 2024
min read

Expanding Horizons: Embracing the Multi-Model Future with Martian

The future of AI is multi-model, and with Martian, you're ready for it.
Etan Ginsberg
May 2, 2024
min read

Introducing RouterBench

Chaithanya Bandi
March 20, 2024
min read

Introducing Martian - Better AI Tools Through Better Understanding

We invented the first LLM router. It achieves higher performance and lower cost than any LLM provider.
Shriyash Upadhyay
July 20, 2023
8
min read