OpenAI has introduced a public beta of its Realtime API, allowing developers to create applications that facilitate natural, back-and-forth conversations with AI chatbots. The new API supports low-latency, multimodal features and enhances voice interactions, enabling seamless communication across languages. Based on its advanced voice mode this allows app developers to integrate humanlike voice interactions into their products. OpenAI plans to expand the API’s capabilities, including new modalities and increased rate limits, while it is currently being tested by applications like Healthify and Speak for personalized coaching and language learning experiences.
Liquid AI, a startup from MIT researchers, has launched its Liquid Foundation Models (LFMs), which are non-transformer AI models that outperform traditional transformer-based models in performance and memory efficiency. The LFMs come in three sizes and are designed for various applications, including financial services and biotechnology. They utilize a unique architecture based on dynamical systems and signal processing, allowing for efficient processing of sequential data. Liquid AI invites early adopters to test the models before a full launch event on October 23, 2024.
Why do we care?
IT service providers working with clients in customer-facing industries (e.g., retail, telemedicine, financial services) will want to explore the integration of AI-powered voice solutions. Years ago, I predicted voice would become a user interface of the future. I may have just been wrong about the timing.
Broadly, the introduction of these more efficient models could translate into lower infrastructure costs for clients running large-scale AI applications. That is a differentiator more than larger training sets. Current leading models are based on the transformer architecture, and this could be a sign of some new competition and innovation.