Personalized AI startup River AI raises $1.1B from consortium backed by Nvidia, AMD

Personalized AI startup River AI has successfully raised $1.1 billion in early-stage funding from a consortium of investors. The capital was received in two rounds, a seed and a Series A, with General Catalyst and AMP PBC as the lead investors. Notable participants also included Nvidia Corp., AMD Ventures, Y Combinator, and Temasek.

River AI's CEO, Igor Babuschkin, previously co-founded xAI Corp. and worked as a researcher at DeepMind. Babuschkin played a significant role in developing Alphabet Inc.'s AlphaCode system, the first coding AI that achieved competitive performance in a programming contest.

River AI's primary product, the River API, is a cloud service that allows developers to customize open-source large language models (LLMs) to suit their specific needs. The service supports LLMs with parameters ranging from 35 billion to 1 trillion. To achieve this, River AI employs a method called LoRA, or low-rank adaptation, which involves extending the model being customized with a small number of additional artificial neurons.

These extra neurons enable the LLM to develop capabilities it does not possess by default. LoRA is favored for its cost efficiency, as it requires only the training of these additional neurons, as opposed to the resource-intensive approach of retraining the entire model from scratch.

The River API boasts the ability to customize a model in 15 to 20 minutes and automates time-consuming prerequisites such as configuring the infrastructure for training. Models customized using River AI's service can be up to four times more cost-efficient than proprietary alternatives. The company aims to develop an expansive AI product suite, with the next addition reportedly being features designed for "personalization and continual learning for agents."

CEO Babuschkin has expressed the long-term goal of creating personal AI systems that can adapt to user preferences, emphasizing that these systems will be owned by the users, providing them with genuine control over their AI. The company is not solely focused on software development, as a job posting suggests plans to create a custom system-on-chip with an onboard machine learning accelerator.

This chip will be produced using advanced foundry nodes, and River AI intends to provide a compiler that automatically converts customer LLMs built using PyTorch into a form optimized for running on this silicon.