Standard Intelligence drops FDM-1 for computer use

The current standard for building a computer use agent is painfully slow: take a bunch of screenshots, pay human contractors to annotate where they clicked, and feed them into a vision-language model. It is wildly expensive, and the moment a task requires more than a few seconds of continuous context—like CAD modeling or video editing—the agent completely falls apart.

AI startup Standard Intelligence unveiled FDM-1, a computer action foundation model trained on video from an 11-million-hour screen recording corpus. Instead of treating tasks as a series of isolated screenshots, FDM-1 processes fluid video and predicts keyboard and mouse actions natively.

Compressing two hours into a single window

The fundamental barrier to video-native agents has always been token hunger. Typical VLMs burn through roughly one million tokens just to parse a single minute of 30 FPS screen footage. That level of inefficiency makes learning long-horizon workflows completely impractical.

Standard Intelligence bypassed this with a custom video encoder trained on an unsupervised masked compression objective. The company claims the encoder compresses nearly two hours of 30 FPS video into 1M tokens—making it 50x more token-efficient than previous state-of-the-art architectures and 100x more efficient than OpenAI's Minecraft encoder from the VPT paper. It also achieves around 100x faster convergence during training compared to standard ViT setups.

Under this architecture, a standard 32k context window holds about 3.5 minutes of live video, while a 1M token window extends continuous context to 1 hour and 40 minutes, all while retaining fine-grained text legibility.

Turning internet video into training data

While the web is packed with decades of coding streams and creative workflows, that video lacks explicit action logs. Current open-source computer action datasets sit at under 20 hours because human labeling is simply too expensive.

To bridge the gap, the team developed an inverse dynamics model (IDM) trained on 40,000 hours of contractor-annotated screens. Using a non-causal masked diffusion architecture, the IDM inspects screen state changes over minutes of context to reliably deduce what keys were hit and how the mouse moved. The IDM then labeled the massive 11-million-hour corpus automatically.

FDM-1 acts as the forward dynamics engine, autoregressively predicting the next action without chain-of-thought tokens, sub-word BPE, or external tool-use layers. Mouse deltas are split into X and Y components and categorized into 49 exponential bins to handle both microscopic nudges and sweeping screen pans. To stress-test this, the company deployed 80,000 forking Ubuntu VMs running over 1M rollouts per hour, with a single H100 GPU orchestrating 42 virtual desktops simultaneously.

Why it matters

We think FDM-1 represents the most sensible divergence from the mainstream agent playbook we have seen in a while.

Big tech players like Anthropic and OpenAI have treated computer use as an extension of chat: take a screenshot, run reasoning chains, generate an API call, and wait. That setup is too sluggish and brittle for continuous manipulation like dragging sliders in Blender or high-speed data entry. Standard Intelligence skips language priors entirely and treats computer operation as an embodied dynamics problem, which looks like the right technical bet for speed and fluency.

For enterprise users and developers, this could unlock practical digital coworkers for heavy software like CAD and financial modeling. However, the potential pitfall is generalization: operating without a reasoning backbone means FDM-1 relies purely on perceptual pattern matching. If a software update slightly changes a workflow or an ambiguous error pops up, we suspect it might struggle where reasoning-heavy models succeed. Still, its 50x video compression breakthrough will likely force frontier labs to rethink how they feed screen data to their own agents.