Standard Intelligence

Standard Intelligence is an American artificial intelligence lab founded in 2024 by Galen Mead and Devansh Pandey, headquartered in San Francisco. The company develops FDM-1, a foundation model trained on raw video of human computer use, intended to produce general computer-use agents that learn behaviors directly from pixel-space observation rather than relying on language-model tool-calling scaffolds. As of 2026, Standard Intelligence had raised approximately $75 million across its early rounds, with Sequoia Capital and Spark Capital as named lead investors and a reported valuation of approximately $500 million.

At a glance

• Founded: 2024 in San Francisco.
• Status: Private. Public benefit corporation (registered as si-pbc on Hugging Face).
• Funding: Approximately $75 million cumulative across early rounds. Named investors include Sequoia Capital and Spark Capital.
• Co-CEOs: Galen Mead (co-founder, age 21) and Devansh Pandey (co-founder, age 20).
• Other notable team: Neel Redkar, Yudhister Kumar, Ulisse, and Ryan, named in the company's FDM-1 launch post.
• Open weights: Some research artifacts published via Hugging Face (si-pbc).
• Flagship products: FDM-1, described as the first fully general computer-action model.

Origins

Standard Intelligence was founded in 2024 by Galen Mead and Devansh Pandey, who met as teenagers through the 2022 Atlas Fellowship, a selective program for high-school students focused on AI alignment and AGI. Mead was 21 and Pandey was 20 at the time of the company's public launch. Both left their undergraduate programs (Mead from the University of Toronto) to pursue the company full time.

The founding thesis was that general-purpose computer-use agents could be built more effectively by training a foundation model directly on raw video of humans using computers, rather than wrapping large-language models in tool-calling scaffolds or fine-tuning vision models on annotated screenshots. The approach was characterized publicly by Sequoia as "deeply contrarian and deeply bitter-lesson-pilled," referencing the canonical Rich Sutton essay on AI scaling. The framing positioned Standard Intelligence as applying Tesla's self-driving methodology, in which raw sensor data scales to capability without hand-engineered intermediate representations, to knowledge work performed on personal computers.

The company assembled an 11-million-hour dataset of screen recordings, then developed a three-stage training recipe: first training an inverse dynamics model on contractor-labeled data to associate screen video with the keyboard and mouse actions that produced it, then using the inverse dynamics model to automatically label the broader 11-million-hour corpus, and finally training the forward dynamics model FDM-1 on next-action prediction. The breakthrough involves a video encoder that compresses approximately two hours of computer-use video into the same number of tokens that competing approaches use for one minute, achieving 50 to 100 times better compression and enabling models to process hours of context at 30 frames per second.

The company is registered as a public benefit corporation (si-pbc), making Standard Intelligence one of a small group of AI labs (alongside Anthropic and the post-restructuring OpenAI) that have adopted PBC structure.

Mission and strategy

Standard Intelligence's stated mission is to develop "self-directing, competent computer use agents" as a step toward artificial general intelligence. The strategic premise is that the same scaling-and-raw-data approach that produced frontier capability in language and vision can be applied to computer use directly, bypassing the language-model-plus-tool-calling architectural pattern that dominates current computer-use agents from OpenAI, Anthropic, and adjacent labs.

The strategy combines three threads. First, foundation-model training on raw computer-use video at scale, with the 11-million-hour corpus and the 50-to-100x more efficient video encoder as the principal technical contributions. Second, a research-first commercial posture in which capability demonstrations are released gradually rather than packaged as immediate consumer or developer products. Third, an explicit alignment with the "bitter lesson" framing that scaling raw data outperforms hand-engineered structure in machine-learning research.

The competitive premise is that the language-model-plus-tool-calling architectural pattern used by current computer-use agents is structurally limited, because the language layer constrains the agent to behaviors expressible in language and the tool-calling layer requires hand-engineered tools for each new task. A pixel-space foundation model trained on human computer use can in principle learn any behavior demonstrated by humans, including behaviors that have not been encoded as explicit tools.

Models and products

• FDM-1. The company's first foundation model, trained on the 11-million-hour computer-use video corpus. Demonstrated capabilities include extruding CAD components in Blender, autonomous driving control in San Francisco after one hour of fine-tuning data using a web-based steering interface built on openpilot, and software-bug discovery through autonomous GUI exploration (a form of "fuzzing").
• Inverse dynamics model (IDM). The labeling model that automatically annotates raw screen video with the keyboard and mouse actions that produced it. Released as part of the FDM-1 training recipe.
• No public API or productized agent. The company has not released a developer API or end-user product as of April 2026. Some research artifacts are published via Hugging Face (si-pbc).

The commercial distribution strategy beyond research output has not been publicly stated. Whether Standard Intelligence intends to ship its own consumer or enterprise computer-use agent product, license FDM-derived capabilities to other labs, or pursue some hybrid posture has not been disclosed.

Benchmarks and standing

Standard Intelligence has not been independently evaluated on the standardized capability leaderboards as of April 2026. The standardized leaderboards measure language and reasoning capability rather than computer-use agent performance, and may not be the appropriate evaluation framework for FDM-1 even after a public model release.

The company's standing rests on the demonstrations published with the FDM-1 launch post, the Sequoia and Spark lead-investor signals, and the publicly disclosed compression ratio of the video encoder (50 to 100 times more efficient than competing approaches). Industry coverage has characterized the approach as a meaningful architectural alternative to the language-model-plus-tool-calling pattern dominant in computer-use agent research as of April 2026.

Leadership

As of April 2026, Standard Intelligence's named leadership and team includes:

• Galen Mead, co-founder and co-chief-executive (age 21). Left the University of Toronto to pursue Standard Intelligence full time. Public face for the company on technical claims and product framing.
• Devansh Pandey, co-founder and co-chief-executive (age 20). Met Mead at the 2022 Atlas Fellowship.
• Neel Redkar, Yudhister Kumar, Ulisse, and Ryan. Named in the FDM-1 launch post as members of the founding team. Roles and prior backgrounds not comprehensively disclosed.

The company's broader research and engineering team has not been publicly disclosed by name. Public reporting describes the team as small, with approximately six people involved at the time of the FDM-1 launch.

Funding and backers

Standard Intelligence's funding history through April 2026 includes approximately $75 million cumulative across early rounds, with Sequoia Capital and Spark Capital as named lead investors. Specific round structures, dates, and post-money valuations have not been comprehensively disclosed in publicly available coverage. The reported valuation is approximately $500 million.

The Sequoia lead is consistent with the firm's established pattern of investing in research-first AI Insurgents (alongside its lead investments in Flapping Airplanes and other 2025-vintage labs). The Spark Capital participation places Standard Intelligence within the senior-tier venture-capital syndicate for mid-size pre-product AI labs.

The company has stated publicly that it operates a 30-petabyte storage cluster in San Francisco for the screen-recording dataset and adjacent training infrastructure, indicating substantial compute and storage commitments relative to the team size.

Industry position

Standard Intelligence occupies a structurally distinctive position within the 2024-vintage Insurgent cohort. The combination of an unusually young founding pair (ages 20 and 21 at public launch), the pixel-space computer-use thesis, the 11-million-hour video corpus, the 50-to-100x more efficient video encoder, and the public-benefit-corporation structure produces a profile not directly mirrored at any other lab.

The closest peer comparators are research-first Insurgents pursuing computer-use agents through novel architectures. Magic pursues frontier-model coding agents through novel architectures. Cognition AI and Cursor pursue language-model-plus-tool-calling computer-use products. Inception Labs and Liquid AI pursue alternative architectures for general-purpose AI. None of these peers carry the youth-of-team and pixel-space-foundation-model thesis that distinguishes Standard Intelligence.

The strategic risks are substantial. FDM-1 has not been independently evaluated, the pixel-space foundation-model thesis has not been validated against the dominant language-model-plus-tool-calling pattern at scale, and the small team and young founders have limited prior frontier-engineering history. The commercial strategy beyond research output has not been publicly stated.

The strategic strengths are distinctive. The 11-million-hour dataset and the video-encoder compression ratio are unusual technical assets that are difficult for competitors to replicate quickly. The Sequoia-and-Spark lead-investor combination signals strategic-investor confidence. The bitter-lesson-aligned positioning differentiates the company from agent labs that rely on hand-engineered scaffolds.

Competitive landscape

Standard Intelligence competes with several Frontier and Insurgent labs and adjacent product-layer companies:

• OpenAI Operator and Anthropic Claude Code. The dominant computer-use agent products from frontier labs, both built on language-model-plus-tool-calling architectures. Standard Intelligence's pixel-space approach is positioned as an architectural alternative.
• Cognition AI and Devin. Leading autonomous coding-agent product. Competes on the same long-running computer-use task surface that FDM-1 demonstrations target.
• Cursor. Leading AI coding IDE. Competes on developer-facing computer-use augmentation, though through a different (LLM-plus-IDE-integration) approach.
• Adept and similar computer-use agent labs. Earlier attempts at general-purpose computer-use agents through alternative architectures.
• Tesla's autopilot and self-driving stack. Architectural reference for the bitter-lesson pixel-space training approach. Not a direct product competitor.
• Wayve and physical-world AI labs. Adjacent in the broader pixel-space-foundation-model research community, though focused on driving rather than general computer use.

Outlook

Several open questions affect Standard Intelligence's trajectory in 2026 and 2027:

• The first public release of FDM-1 weights, an API, or a productized agent.
• Independent capability evaluation against frontier computer-use agents from OpenAI, Anthropic, and adjacent labs.
• The commercial strategy, which has not been publicly stated.
• Whether the company accepts follow-on capital at a higher valuation, and on what timeline.
• The 11-million-hour dataset's continued growth and any disclosure of new capability domains beyond CAD, driving, and software fuzzing.
• Senior-talent recruitment as the team scales beyond its small early-2026 size.

Sources

• Sequoia Capital: Standard Intelligence: Training General Intelligence in Pixel Space. Lead-investor announcement and the bitter-lesson framing.
• Standard Intelligence: The First Fully General Computer Action Model. Official FDM-1 launch post with technical details and capability demonstrations.
• AI2Work: Standard Intelligence Raises $75M to Build AI That Uses Computers. Coverage of the funding round.
• Digital Applied: FDM-1: AI Trained on 11M Hours of Screen Footage. Technical analysis of FDM-1.
• Hugging Face: si-pbc (Standard Intelligence). Public-benefit corporation registration and research artifacts.
• Galen Mead personal site. Co-founder reference.