Key Takeaways
- • Standard Intelligence has developed FDM-1, described as the first fully general computer action model, trained on an 11-million-hour video dataset to perform complex tasks across CAD, web navigation, and real-world driving at 30 FPS.
- • The model breaks from traditional approaches by training directly on video rather than contractor-annotated screenshots, enabling long-context understanding and scaling capabilities that prior vision-language model approaches could not achieve.
- • FDM-1 uses an inverse dynamics model to label actions (keystrokes, mouse movements) from video and employs a video encoder capable of compressing nearly 2 hours of 30 FPS footage into only 1 million tokens, enabling internet-scale training on unlabeled content.
- • The architecture represents a fundamental shift in agent training methodology: previous systems required expensive contractor-labeled datasets (largest open dataset under 20 hours) and were limited to few seconds of context; FDM-1 trains unsupervised on millions of hours of publicly available internet video.
- • Sequoia Capital's involvement signals institutional venture capital conviction in the computer action model category, positioning Standard Intelligence as a foundational AI infrastructure play targeting professional workflows in CAD, finance, engineering, and ML research.
Summary
Standard Intelligence unveiled FDM-1, positioning it as the first fully general computer action model capable of performing tasks spanning website navigation, multi-step CAD modeling sequences, and real-world vehicle control—all at 30 frames per second. The model represents a departure from the prevailing methodology for building computer use agents, which had relied on finetuning vision-language models on contractor-labeled screenshots combined with task-specific reinforcement learning environments. That approach fundamentally limited scalability and contextual understanding.
The technical foundation of FDM-1 centers on training at internet scale using an 11-million-hour video dataset. Rather than requiring expensive manual annotation of screenshots, Standard Intelligence developed an inverse dynamics model to automatically label actions—keystrokes, mouse movements, and other inputs—directly from raw video. A specialized video encoder compresses approximately two hours of 30 FPS video into just 1 million tokens, enabling the model to maintain long-context awareness while processing high-framerate input. This architectural choice allows unsupervised learning across the entirety of public internet video, including film editing recordings, programming livestreams, and video game playthroughs accumulated over two decades.
The scale differential between FDM-1 and prior systems underscores the competitive advantage. Existing open-source computer action datasets contain fewer than 20 hours of 30 FPS video; FDM-1 operates across orders of magnitude more training material. Traditional contractor-annotated approaches proved prohibitively expensive at scale and incapable of handling tasks requiring sustained context beyond a few seconds. Standard Intelligence's unsupervised approach eliminates these bottlenecks, allowing the model to train on unlabeled internet video while consistently improving performance with additional scale—a pattern analogous to the scaling laws observed in large language models like GPT-3.
The platform targets professional domains including computer-aided design, financial services, engineering, and machine learning research, positioning FDM-1 as potential infrastructure for knowledge worker automation. Sequoia Capital's backing of the initiative signals institutional conviction that general-purpose computer action models represent a defensible and valuable category within the broader artificial intelligence landscape. The transition from task-specific agents to a general computer action model with long-horizon reasoning capabilities addresses a fundamental limitation in previous generalist AI systems and establishes a new benchmark for agent capability in structured digital environments.
The technical foundation of FDM-1 centers on training at internet scale using an 11-million-hour video dataset. Rather than requiring expensive manual annotation of screenshots, Standard Intelligence developed an inverse dynamics model to automatically label actions—keystrokes, mouse movements, and other inputs—directly from raw video. A specialized video encoder compresses approximately two hours of 30 FPS video into just 1 million tokens, enabling the model to maintain long-context awareness while processing high-framerate input. This architectural choice allows unsupervised learning across the entirety of public internet video, including film editing recordings, programming livestreams, and video game playthroughs accumulated over two decades.
The scale differential between FDM-1 and prior systems underscores the competitive advantage. Existing open-source computer action datasets contain fewer than 20 hours of 30 FPS video; FDM-1 operates across orders of magnitude more training material. Traditional contractor-annotated approaches proved prohibitively expensive at scale and incapable of handling tasks requiring sustained context beyond a few seconds. Standard Intelligence's unsupervised approach eliminates these bottlenecks, allowing the model to train on unlabeled internet video while consistently improving performance with additional scale—a pattern analogous to the scaling laws observed in large language models like GPT-3.
The platform targets professional domains including computer-aided design, financial services, engineering, and machine learning research, positioning FDM-1 as potential infrastructure for knowledge worker automation. Sequoia Capital's backing of the initiative signals institutional conviction that general-purpose computer action models represent a defensible and valuable category within the broader artificial intelligence landscape. The transition from task-specific agents to a general computer action model with long-horizon reasoning capabilities addresses a fundamental limitation in previous generalist AI systems and establishes a new benchmark for agent capability in structured digital environments.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.