A project called Handsdiff published results from a two-week experiment training a model to predict what a knowledge worker types next, based entirely on their screen activity. The core claim is straightforward: if a model can predict your next write, it can do the grunt work of prompting agents on your behalf, and the economics work because prediction is far cheaper than human thought.

Why predict what humans type

The bottleneck in agent-heavy workflows is not model capability. It is the human's ability to ingest information, apply judgment, and compose the next prompt. That effort is a tractable prediction task. A model trained on one person's read and write history could, in theory, produce the same output for a thousandth of the cost and time. The result: the human spends less time giving context to agents and more time thinking about novel problems.

The author describes a progression. First, a user model that predicts your next write speeds up the collection of context for long-running agents. Second, the model gets good enough to infer your goals and suggest actions you would not have composed yourself, or would have found too frictional to write. Third, frontier models work directly with the user model to complete tasks before presenting them to you. Finally, a single combined agent both learns your judgment and applies it faster than you could alone. Your presence remains necessary to give the machine direction, but the execution shifts to the model.

The data pipeline

The team built an open-source pipeline that captures raw computer use data on MacBooks and converts it into structured supervised fine-tuning datasets. Each stage, from capture to cleanup to training-ready format, was reviewed manually and deemed necessary. The raw data consists of read and write events: what appeared on screen, what the user typed, and when.

The training recipe used token-level cross-entropy loss on the user's writes, a 32,000-token context window with a sliding window of prior events, and Qwen 3.6-35B-A3B running on Tinker. The model updated after every 50 examples rather than on a fixed schedule, mimicking a live deployment where the model trains continuously as new data arrives.

Evaluation followed the same sequential pattern. The model trained on the first 50 examples, then predicted the next 50. After each batch of 50, training resumed and prediction moved forward. Every example appeared exactly once, in chronological order, with no shuffling across time boundaries.

Accuracy numbers and cost

The results come from a single-user, two-week study. Astra xhigh achieved 17.1% semantic accuracy on substantive next-write predictions at $0.30 per call. The base Qwen 3.6-35B-A3B model managed 0.9% accuracy. The continually trained version of the same model reached 3.0% accuracy at $0.01 per call. The gap between the base and trained versions shows that learning occurs, and the author reports early data suggesting loss reductions scale with the amount of trained data, similar to local scaling laws observed in larger model training.

The low accuracy across all models reflects the difficulty of the task and the small dataset. Predicting what a human will write next, given only their screen history, requires modeling not just language but intent, context, and domain-specific judgment. The 17.1% figure from Astra xhigh represents the current ceiling for this approach, and the team expects it to improve with more data.

What remains unsolved

Two weeks of data from one person produces wide error bounds on how performance scales with additional data. The dataset is too small to draw firm conclusions about scaling behavior. It is also unclear what accuracy threshold makes the model genuinely useful in a workflow. That can only be determined by integrating it into real work and measuring whether it saves time.

The jump from learning from a suboptimal human demonstrator to inferring goals and performing at a superhuman level may require different training methods. The current setup learns to predict your writes, not to exceed your judgment. The path from prediction to goal inference to superior action is not straightforward.

Building a production version means compressing the entire pipeline, data collection through serving, into something that runs cheaply, continuously, and without intervention for multiple users. Privacy standards must be strict given the sensitivity of the data. The team also needs to account for how displaying real-time model predictions to the user changes the training signal, since the user's behavior may shift in response to seeing the model's suggestions.

The argument is that without this kind of personal modeling, future AI systems consolidate around the goals and preferences of their builders rather than the people who use them. Capturing and learning from individual judgment, continuously, is the mechanism that keeps the machine aligned with the person directing it.