Project Matroska
In 2019, Autonomio proposed that a researcher should be able to describe a scientific problem in ordinary speech and have a machine determine how to investigate it. At the time, training a specialist model in TensorFlow or Torch required the researcher to devise the method and express it in code. The Automated Research roadmap set out 25 steps in five phases, and the accompanying essay placed the automated researcher in 2035.
In 2026, a researcher can already direct such an investigation from an ordinary laptop in conversation with an agent.
The terms of competition in using frontier AI to drive research forward remain unfair. In every scientific field, the teams that can afford more compute or more tokens can pursue investigations beyond the means of their peers. Even a team that can afford to comprehensively use frontier models throughout a study must work with decisions made by the lab that trained it, including which data the model learned from, how it was trained, and what kinds of answers post-training limited it to.
Project Matroska is Autonomio’s direct challenge to a system in which only a few frontier AI labs can afford to train models with capabilities matching those of human researchers. Its purpose is to give researchers in every scientific field the means to train and use models for their own investigations on the computers they already have. It follows the established practice of machine learning, in which the researcher selects the training data, determines how to train the model for the problem under study and decides how to use it for inference.
Matroska applies this method first to transformer-type models. The programme began in 2026 with a measurement rig on an ordinary laptop. Its measurements guide reductions in the computation required for data preparation, training and inference, with each reduction assessed over the whole experiment. The programme concludes by 2030 with the release of self-assembling binary logic circuitry that researchers can train and use for inference on conventional hardware.
Phase 1 of 5
Measurement on ordinary hardware
The first phase consists of building the measurement rig and establishing its precision. Repeated runs of a reference workload (training a 1B parameter language model) on the same laptop establish the normal variation in execution time, data movement and energy consumption. Controlled changes to the workload then show how small a difference the rig can reliably detect in each quantity.
Phase 2 of 5
Computational overhead
Without changing the mathematics, the second phase reduces the work of preparing inputs, scheduling operations, moving intermediate data and collecting outputs. Decomposing the workflow reveals where the same computation is repeated or the same data are moved unnecessarily. Factoring out the work common to several operations lets the program perform it once and reuse the result.
Keeping the hot path hot means arranging the operations a program performs most often so that they can reuse data while it is still in fast memory. The value of locality was well understood by 1970. In 2014, Autonomio described how layers of software can obscure the work a program asks the hardware to perform.
Phase 3 of 5
Matrix multiplication
The third phase examines the algebra of matrix multiplication to find faster ways to compute the products required for training and inference. For identical input matrices, each method must produce a matrix numerically equivalent to the product computed by the reference implementation. The rig measures the time required for the multiplication and the effect on the cost of the whole experiment.
Phase 4 of 5
Data preparation, training and inference
The fourth phase rebuilds the software used to prepare data, train transformer-type models and run inference. The new implementations eliminate the redundant computation and data transfers identified in phase two and use the faster methods of matrix multiplication validated in phase three. Each component must serve as a drop-in replacement for its predecessor so that researchers can repeat an established experiment using the same research code.
Phase 5 of 5
Self-assembling logic
In the final phase, the system trained on the researcher’s data is a circuit of binary logic gates that assembles for the task under study. In the proposed non-von Neumann architecture, the assembled circuit holds state and performs operations on it within the same structure. The programme delivers this architecture in software by 2030, so that researchers can train the circuits and use them for inference on conventional hardware.