Cotangent makes the work after backpropagation faster: clipping gradients and updating weights. It batches tensors while preserving the tested reduction arithmetic. Across four configurations on one M5 Pro, training took 23.55–34.07% less time than matched compiled-model baselines. The paper includes ablations, numerical checks, and raw measurements; final quality equivalence remains unresolved for three configurations.
Open projectAn agent that explores unfamiliar grid games without instructions, works out the rules, and plans its next move. It builds a model of what its actions change as it plays. Seven official game campaigns have verified local wins. The latest unseen-game competition score was 0.18%, leaving a large gap between solving familiar environments and generalizing to new ones.
Open projectA strategy agent that predicts moving planets and fleets, coordinates attacks, and keeps enough ships back for defense. Every match is recorded for replay and comparison. In fresh local tests, it won 82% across a 50-game opponent panel, but only 36.7% in 30 games against its hardest opponent.
Open projectA rocket-landing agent that learns an internal simulator from experience, then practices inside imagined futures. A dashboard compares those imagined flights with what actually happens, and checkpoints let training resume. The first 80-step run on Apple Silicon completed the training loop with a 46,287-parameter model, but produced no successful landings. It establishes a working experiment, not strong control performance.
Open projectA system for finding faster code by testing candidate implementations, measuring their speed, and retaining the best correct version. A packaged Levenshtein run achieved a verified 5.18x speedup. Its fixed candidate library makes the experiment reproducible; discovering new algorithms is still an open extension.
Open projectA sequence model with a small memory bank that learns when to write, overwrite, and forget information. Controlled tasks test whether those choices help it retain changing facts. In a three-seed pilot against similarly sized models, learned forgetting showed no clear advantage; the local Transformer was roughly five times faster.
Open projectPredicts how a cell population responds to gene interventions it has not seen during training. Synthetic cell worlds make the underlying interactions known and testable. Across five worlds, prediction error was 0.0353 for new gene combinations and 0.0337 for unseen genes, beating a direct-shift baseline in both tests. Biological validity still needs real-data evaluation.
Open projectTests whether checking several GPT OSS answers improves accuracy over taking the first answer. It compares voting, verifier ranking, and adaptive sampling, recording accuracy alongside the extra generation cost.
Open projectTrains a rocket to land using PPO, with thrust, torque, fuel use, and landing constraints. An expert controller teaches the initial policy; longer training still has stability problems.
Open projectClassifies galaxy images as elliptical, spiral, or edge-on disks. ResNet18 reached 93.18% test accuracy on high-confidence labels; training adjusts batch sizes to available GPU memory.
Open projectA browser agent for student assignment workflows. It gathers requirements through chat, plans the task, and runs browser actions with live monitoring and account controls.
Open projectBuild and revise a website through chat, then publish it at a named URL. Helios generates responsive HTML, CSS, and JavaScript through a Django application.
Open projectA private subscription app with Free and Plus plans. Account dashboards track usage, search history, and activity, with billing infrastructure for managing paid access.
Private repo