April 2024 Summaries

Users can begin using Nebius AI by logging in with a Google or GitHub account, allowing them to manage cloud resources through a console, CLI, API, or Terraform provider. The platform provides documentation to assist users in creating an account, setting up billing, and managing taxes, which helps facilitate the integration of necessary services and third-party products. This setup aims to create an optimal environment for data preparation, model training, fine-tuning, and running inference. For further assistance, users have the option to directly contact technical support through the console or via email.

Large language models (LLMs) are powerful tools trained on extensive datasets, yet they face limitations such as being resource-intensive and lacking access to real-time or private data. The retrieval-augmented generation (RAG) technique is emerging as a solution to enhance LLMs for enterprise applications, particularly in regulated industries, by enabling them to access and integrate up-to-date and proprietary information at the time of prediction. This approach is beneficial in contexts where traditional LLMs fall short, such as in providing timely and specific knowledge not present in their training data. A practical example is a chatbot designed to assist users in navigating company dashboards, where RAG helps by retrieving relevant metadata and contextual information. The article outlines a proof-of-concept implementation using open-source tools like Flowise for application development, Pinecone for vector storage, and the OpenAI API for LLM functionalities. While this setup is suitable for experimentation, scaling such a solution would require a more robust production environment, a topic to be addressed in an upcoming webinar.

The latest release introduces managed databases and container registries to enhance machine learning (ML) workloads by offering secure, fault-tolerant, and ready-to-use environments, with cloud providers handling setup, maintenance, and scaling. This update frees users to focus on optimizing models while storing metadata for ML tools. A webinar by cloud solutions architect Panu Koskela explores Slurm and Kubernetes for ML, discussing their architectures and considerations for platform choice. Users can now monitor GPU usage on virtual machines via dashboards to manage resources and detect anomalies, and a new guide details setting up GlusterFS distributed storage for scalable model training. Strategies for efficient large model checkpointing are also shared. Nebius AI is at the forefront of adopting NVIDIA B200 Tensor Core GPUs, emphasizing energy efficiency. Additionally, a new blog category on AI research kicks off with discussions on alternatives to transformers in large language models.

Building an AI-centric cloud platform involves a balance of innovation and customer satisfaction, as demonstrated by the company's commitment to research, development, and client collaboration. Despite being a relatively new venture, the platform has shown stability, particularly in its GPU cloud services where clients engage in training, fine-tuning, and inferencing. This complex process is supported by a high-speed InfiniBand network and managed through Kubernetes, with additional tools available via a partner Marketplace. The company's growth and success are attributed to a dedicated team, enhanced by a bootcamp program that allows new employees to gain diverse experience across departments. The organization emphasizes not only technical achievements but also the importance of a supportive work environment, illustrated by their attention to staff well-being and practical metrics like the unique 'hare rate' KPI.