Magic

Building frontier code models to automate software engineering and AI research.

Contact for Pricing

Description

Magic is an AI research and development company focused on creating frontier code models. Their primary goal is to automate complex tasks within software engineering and AI research, viewing this as a promising route to developing safe Artificial General Intelligence (AGI) more reliably. The company leverages cutting-edge techniques including frontier-scale pre-training, domain-specific reinforcement learning, ultra-long context capabilities (demonstrated up to 100 million tokens), and substantial inference-time compute resources.

Supported by significant funding and operating with extensive computing infrastructure, including 8,000 H100 GPUs, Magic concentrates on fundamental research problems in AI development and code generation. They emphasize responsible development through initiatives like their AGI Readiness Policy, which outlines procedures for evaluating and mitigating potential risks associated with powerful AI systems.

Key Features

Frontier Code Models: Development of advanced AI models specifically for code generation and understanding.
Software Engineering Automation: Aims to automate various tasks within the software development lifecycle.
AI Research Automation: Focuses on using AI to accelerate AI research itself, including model improvement and alignment.
Ultra-Long Context Windows: Capability demonstrated with 100 million token context windows for processing extensive information.
Domain-Specific Reinforcement Learning: Tailoring model training through reinforcement learning specific to coding and research domains.
AGI Safety Focus: Integrated approach to evaluating and mitigating risks associated with advanced AI capabilities (AGI Readiness Policy).