Exploring the Emergence of Open Language Models: The Story of Open Router

Exploring the dawn of multimodal AI, where language models like Open Router are revolutionizing software development.

  • 1. The speaker started Open Router in early 2023 to explore the potential of the AI market.
  • 2. They wanted to answer the question: "Will this market be winner-take-all?"
  • 3. Open AI was the leading model, with a few others close behind.
  • 4. The speaker built prototypes and investigated open source options.
  • 5. In January, they noticed users wanting different types of models for moderation.
  • 6. Next month, the open-source race began with models like Bloom 176b and OPT by Facebook.
  • 7. Llama 1, launched in February by Meta, outperformed GPT3 on most benchmarks.
  • 8. In March, Alpaca was successfully distilled from Llama 1, making it possible to transfer style and knowledge from large models to small ones for less than $600.
  • 9. Open Router started as a place to collect these language models but evolved into a marketplace over time.
  • 10. In April, the speaker launched Window AI, an open-source Chrome extension that lets users choose their model and web apps to "suck it in."
  • 11. Open Router now has 400+ models from over 60 active providers, with crypto as a payment method.
  • 12. Initially, there were only one or two providers for each open-source model, but this quickly changed.
  • 13. The ecosystem of language models became "out-of-control heterogeneous," leading to the creation of Open Router as a marketplace.
  • 14. Aggregating providers helped developers increase uptime and access real-world data on latency and throughput.
  • 15. Open Router helps users compare models, use their own prompts, have fine-grain privacy controls, and see usage across all models.
  • 16. The speaker believes the AI market will not be winner-take-all; instead, it will be multimodel.
  • 17. Many customers use different models for various purposes, realizing significant gains by doing so.
  • 18. Inference is becoming a dominant operating expense, and selecting and routing will be crucial.
  • 19. Open Router aims to make the "wild ecosystem" more homogeneous and easier to work with for developers.
  • 20. The speaker shared their idea for middleware within Open Router, enabling new features and abilities like web searching and PDF parsing for all models.
  • 21. Middleware is optimized for inference and works similarly to the way it does in web development frameworks.
  • 22. Plugins can call MCps (Model Cards for Model Reporting) from inside a plugin and augment results on the way back to the user.
  • 23. Open Router focuses on low latency, with custom cache work bringing latency down to 30 milliseconds.
  • 24. The company plans to add more modalities, like generating images through transfusion models, and improve routing capabilities.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!