Standard Intelligence: Training General Intelligence in Pixel Space
Summary
- Standard Intelligence bets that scaling raw video data is the optimal path to building general computer agents rather than relying on language models or screenshots.
- The team's approach involves full video pre-training where models predict mouse movements, clicks, & keystrokes directly from screen pixels.
- The company has created an 11-million-hour computer action dataset & a video encoder that is 50× more token-efficient than competitors.
- Standard Intelligence operates a 30-petabyte storage cluster in San Francisco for under $500K, which is roughly 20× cheaper than hyperscaler alternatives.
- Their first foundation model, FDM-1, can perform tasks such as CAD extrusion, car driving after fine-tuning, & software bug exploration.
- Founders Galen Mead (21) & Devansh Pandey (20) are former Atlas Fellowship participants who left undergraduate programs to pursue AGI development.
- The six-person team turned down conventional career paths to focus on this mission & emphasizes safety & technical scrappiness.
- Spark Capital partners Miko & Yasmin joined the Series A funding round led by the current investors.
Standard Intelligence’s thesis is that the best way to build a general agent is through full video pre-training on computer use, because it is the only approach that can truly scale action data.