Today, we launch Project Bedrock. Authorized on October 17, 2026 and formally begun on October 18, the initiative will run through a focused three-month period. Launch preparations are complete. LANCER is now examining the feasibility of its learning and practical directions, working together with CIE and other collaborators. The purpose is clear: to go deeper into the mathematics of AI, building an understanding of deep learning and neural networks that supports independent judgment, sound practice and questions worth pursuing.
The capabilities of large models rest on choices about representation, optimization, learning and computation. Why does a familiar formula hold? What makes an architecture work? Under which conditions can an empirical finding carry over? These questions repeatedly lead back to fundamentals. Bedrock places deeper mathematical capability at the center of its work, studying the foundations through concrete problems and tracing the relationships among assumptions, derivations and conclusions. We want to explain how a method works and judge whether its reasoning still applies under new conditions.
Understanding also needs historical depth. The history of deep learning records how questions emerged, how methods were revised and how different ideas shaped one another. Through broad and careful reading, the initiative will connect foundational work with current large-model research: what difficulty did each contribution address, and what did it make possible for the work that followed? Reading will develop into connected notes with clear sources, placing new terminology back in the context of the problems it serves and tracing how technical directions took shape.
Reading will be followed by first-hand derivation. We will trace important formulas back to their definitions and assumptions, work through the intermediate steps, and use small implementations to examine their computational behavior. A stalled derivation or an experiment that differs from expectation can locate a missing connection in our understanding. Recording that process makes learning revisitable and gives readers a path through the reasoning: one they can check, question and help improve.
Open-source practice will bring that understanding into real engineering. We plan to study the implementation and design of existing projects, following how theory becomes a usable system through data, code, evaluation and maintenance. From this foundation, Bedrock aims to develop representative open-source projects of its own. Their direction will grow from difficulties encountered while learning and needs expressed by the community. Each contribution should begin with a clear problem and develop through implementation, comparison and feedback.
Sharing is part of understanding. The initiative plans to establish an X account and share its knowledge, progress and reflections alongside Xiaohongshu: how to read a paper closely, how to derive a formula, what an implementation teaches, and why an unresolved question deserves further work. We want these accounts to be clear and concrete, giving readers with different backgrounds an entry point into the discussion. They will continue beyond the three-month intensive period, allowing the record of learning and the conversation around it to grow.
Community dialogue will also shape the questions we pursue. A user's difficulty, a developer's feedback or a different reading of the same paper may reveal a need that has not received enough attention. Bedrock aims to examine these questions carefully, understand the conditions in which they arise and seek improvements that can be tested. We hope to respond with clearer explanations, more reliable implementations or methods worth investigating further, gradually turning individual learning into useful contributions to the AI community.
Understand what we know more thoroughly. Pursue what remains unclear more deeply.
The three months will follow a connected path: establish links between mathematics and the literature, deepen understanding through derivation and implementation, then shape useful work into shareable artifacts and learn from community feedback. Reading, practice and dialogue will continue throughout, with the pace adjusted as feasibility work and practical progress develop. Planned public outputs include study notes, derivations, paper readings, implementation records and open-source work, making progress visible through concrete contributions.
Project Bedrock forms part of a broader personal growth plan. It connects deeper study with practice and seeks to build a wider presence through sustained, useful sharing. Over these three months, we hope to understand familiar ideas more thoroughly, pursue what remains unclear more deeply, and leave work that others can use and develop. The initiative is now underway. Its next steps begin with the questions worth understanding—and the foundations needed to pursue them.