Standard Intelligence PBC

Market position

The company positions FDM-1 against the prevailing approach of fine-tuning vision-language models on contractor-annotated screenshots plus task-specific reinforcement learning environments, which it says limits agents to a few seconds of context; it cites OpenAI's Video PreTraining (VPT) work and VideoAgentTrek as prior IDM-based approaches with shorter context or no video context. For audio, it claims hertz-dev's real-world latency is roughly 2x lower than the previous state of the art and that hertz-codec outperforms SoundStream and EnCodec at 6 kbps while being on par with DAC at 8 kbps in subjective evaluations [1][3].

Claimed differentiators are training and inference directly on long-context video rather than screenshots, an unusually token-efficient video encoder, the ability to learn unsupervised from internet-scale video without contractor labeling (the largest cited open computer-action dataset is under 20 hours of 30 FPS video versus its 11-million-hour corpus), demonstrated transfer from screen tasks to real-world driving, and an openly released full-duplex conversational speech base model [1][3].

Technology

FDM-1 is trained with a three-stage recipe: an inverse dynamics model (IDM) is trained on 40,000 hours of contractor-labeled screen recordings; the IDM then labels an 11-million-hour screen-recording video corpus; and a forward dynamics model is autoregressively trained on next-action prediction, with an output token space of key presses and mouse-movement deltas. A central component is a video encoder that compresses nearly two hours of 30 FPS video into about 1M tokens, which the company claims is 50x more token-efficient than the prior state of the art and 100x more than OpenAI's encoder. The system uses OS checkpointing and forking VM infrastructure to enable test-time compute and state-space exploration. The self-driving demo forked openpilot's joystick mode and used under one hour of fine-tuning data. On the audio side, hertz-dev uses causal convolutions for streaming inference, 15-bit quantized phonetic latents via Finite Scalar Quantization, and duplex handling via two concatenated projection heads, achieving a stated 80ms theoretical average latency and 120ms real-world latency on a single RTX 4090 [1][3].

Go-to-market

Distribution to date is through public research publication and open-source release: hertz-dev weights and code are published on GitHub with downloadable checkpoints, and FDM-1 results are shared via demo videos and technical blog posts. The company solicits researcher and engineer applications and investor contact directly through its site and email [1][3].

The company frames FDM-1 as a model with long-context training needed to act as a coworker for CAD, finance, engineering, and eventually machine learning research, and highlights automated GUI testing/fuzzing as an application area; no named customers are disclosed [1].

Geography

Operations are based in San Francisco, United States, including the downtown San Francisco storage cluster and local self-driving demonstrations [0][1][2][4].

History

Public blog posts trace a progression from the hertz-dev conversational speech base model, released when the company described itself as a team of four in San Francisco, to FDM-1, a general computer action model trained on IDM-labeled video, followed by a $75M funding round from Sequoia and Spark Capital with the team at six people in San Francisco [2][3]. A third-party directory lists a founding year of 2023 [4].

Risks & controversies

No controversies are reported in the sources. Notable uncertainties: capability claims for FDM-1 and hertz-dev are self-published by the company and not independently verified in the material provided; the team is very small (four people at the hertz-dev release, six at the funding announcement); the company states current alignment techniques are insufficient for models with human-level learning capabilities; and third-party funding, revenue, and valuation figures conflict with or are explicitly estimated rather than reported (a directory cites $90M total raised and modeled revenue/valuation ranges versus the company's stated $75M round) [2][3][4].