What Perceptron AI does
Perceptron AI is building foundation models designed for real-time, multimodal intelligence across the physical world. Founded in November 2024 by former Facebook AI Research (FAIR) scientists Armen Aghajanyan and Akshat Shrivastava, the company's thesis is that after foundation models transformed the digital realm, the next frontier is physical AI: models that process real-world video and other sensory streams. Its approach treats video, audio, text, and structured data streams as equally important inputs while maintaining real-time performance across hardware.
Key capabilities
Perceptron's first model, Mk1, targets video analysis and has been reported to match frontier-lab performance while costing roughly 80-90% less than comparable offerings from Anthropic, OpenAI, and Google for video tasks. The models are aimed at low-latency understanding of continuous real-world streams rather than static text or single images, enabling applications that need to perceive and react in real time.
Who it's for
Perceptron AI targets builders in robotics, manufacturing, geospatial analysis, security, and content moderation who need cost-efficient, real-time multimodal perception. The company is backed by Bessemer Venture Partners, S32, and Foundation Capital.