Most foundation models learn from text. Standard Intelligence is making a different bet: train a model that learns by watching software being used, pixels, clicks, scroll events, and interface interactions, and build generalization from there rather than from language.
The founding team is small enough and focused enough to stay close to the problem. Six people, $80M from Sequoia and Spark Capital, and Andrej Karpathy as an advisor. That is a very specific kind of conviction about a very specific problem. The FDM-1 release suggests they are moving faster than most people realize.
The Market
A pre-category bet in a market measured in trillions.
The honest answer about the market for computer-use foundation models is that it does not exist as a measured category yet. The AI software market broadly will reach $391B by 2030. The specific slice Standard Intelligence is after, models that understand software as interaction rather than description, has no clean analogue in any analyst report. That is the point. The interesting bets are always pre-category.
What makes the broader context relevant is the scale of what AI agents are being asked to do. Gartner projects that 40% of enterprise applications will embed AI agents by end of 2026, up from under 5% in 2025. Most of those agents are being built on top of text-trained models that were subsequently given eyes. Standard Intelligence is asking whether training from the ground up on what software actually looks like produces something different.
The Landscape
Where the main approaches differ.
The competitive frame here is not company versus company but approach versus approach. Most labs extend existing text models with vision capabilities. Standard Intelligence trains from the ground up on computer-use data. The difference in approach is the bet.
scratch on pixels
model
focused
RPA
● Primary focus◐ Partial / secondary○ Not the bet
The Engineering
The state-space problem.
Text has finite vocabulary and well-understood structure. A computer screen at any given moment is essentially unbounded in what it can contain. The challenge of generalizing across arbitrary software environments, UI patterns, and interaction styles, without overfitting to any one of them, is unsolved at scale. That is what makes training from pixels genuinely different from fine-tuning a language model.
Having worked on quantizing and running models like Mistral-7B on GPU for production inference at Movius, I have a reasonable sense of the gap between a compelling research idea and something that actually ships. Six people doing this at the foundation model layer is not the obvious move. That is what makes it interesting to watch.
Most foundation models learn from text. Standard Intelligence is betting the computer screen is a better teacher.
The Frontier
What the bet actually is.
The question at the center of Standard Intelligence is whether training from the ground up on computer-use data produces models that generalize across software environments better than fine-tuning an existing language model on the same data. The hypothesis is yes, and FDM-1 is the first test of it. The honest answer is that nobody knows yet at scale.
What makes this bet legible is the team. Sequoia and Spark Capital do not write $80M checks to six-person teams without conviction about the approach, and Karpathy does not advise teams without a technical opinion. The question I keep sitting with is not whether the approach is interesting (it clearly is) but whether a small team can maintain the research velocity to stay ahead as every major lab piles into computer use from the other direction.