New AI Model Learns Software by Watching Videos
Standard Intelligence has debuted FDM-1, a new AI model that learns how to perform complex software tasks by watching video footage. The model was trained on 11 million hours of screen recordings, enabling it to master programs like Blender for 3D modeling and to debug code. This vision-based learning represents a significant leap in capability, allowing the AI to operate software it has only observed.
- The 11 million hours of screen recordings used to train FDM-1 is a dataset larger than any publicly available for computer-use modeling. To generate training signals at such a large scale, Standard Intelligence employed an automated technique that reconstructs user actions like keystrokes and cursor movements from on-screen visual changes. - FDM-1 is capable of processing nearly two hours of continuous video in a single session, a significant increase in visual context compared to previous models that were often limited to short sequences or static screenshots. - While Standard Intelligence is described as a boutique consultancy for AI and data strategy, there is a prominent San Francisco-based company named Standard AI (also known as Standard Cognition) that focuses on AI-powered autonomous checkout for retailers and is backed by major investors like SoftBank. - The broader AI software market is projected to grow significantly, with one report estimating it will reach $467 billion by 2030, driven by both generative AI and "traditional AI" like computer vision. Early adopters of AI are already seeing substantial returns, with productivity increases in software engineering reported to be between 20-45%. - For the San Francisco Bay Area, the rise of AI is having a notable economic impact. The region is home to the most AI talent in the U.S., with 76,079 professionals, and has attracted three-quarters of the nation's AI venture capital funding since 2019. This has led to a rebound in office leasing, with AI-related companies leasing 1.1 million square feet of office space in San Francisco in the first half of 2025. - The increasing adoption of AI in software development is creating new career paths. Roles such as computer vision engineer, AI software engineer, and machine learning engineer are in high demand, requiring skills in programming languages like Python and familiarity with deep learning frameworks. - Case studies on the impact of AI in software development show significant productivity gains. For example, developers using GitHub Copilot completed 26% more tasks on average, with junior developers seeing the largest productivity increases. Another study highlighted a 40% reduction in client revision cycles and a 50% decrease in bug-fixing time with the use of AI-assisted tools. - The development of models like FDM-1, which learn from experience (in this case, video), aligns with a broader trend in AI research. Some experts predict a shift away from learning solely from massive text datasets towards agents that refine their abilities through action and environmental feedback.
