Naming things in machine learning has become an act of quiet defiance against search engines, and it is starting to cause real friction. When a practitioner searches for "SOAP" to debug a training run, they get soap recipes. "Shampoo" leads to hair care. "FOAM" brings up mattresses. Even "Muon," which is slightly more distinctive, collides with a fundamental particle from physics. This is not a minor inconvenience; it is a barrier to discovery and collaboration that the field has tolerated for too long. We believe this naming chaos is a symptom of a deeper issue: the community treats novelty as a branding exercise rather than a communication duty, and it makes the work harder for everyone who follows.
The problem is not that these names are whimsical; it is that they are functionally useless for anyone trying to learn or build on the research. A newcomer exploring modern optimizers should be able to find the original paper, a readable explanation, and a working code example in seconds. Instead, they are buried under pages of commercial products and pop culture references. This is where the frustration of the original poster becomes a shared pain point. We have seen this pattern before in adjacent fields. Consider how Understanding AI's Drive: What Sutskever Meant by "The Models Want to Learn" reframes a core AI concept in plain language that invites inquiry rather than confusion. The contrast is instructive: clear, intentional naming opens doors; obscure naming locks them.
What makes this situation more frustrating is that the tools to solve it exist. The community has already demonstrated how to build bridges between complex technical ideas and practical application. For instance, Discover how frontier LLMs read and write neural network weights natively shows that even deeply technical work can be presented in a way that feels accessible and actionable. Similarly, PyTorch Workloads Now Run Natively on Google TPUs proves that infrastructure changes can be communicated clearly without sacrificing depth. These examples share a commitment to clarity that the optimizer naming convention lacks. We are not arguing for sterile, academic labels. We are arguing for names that survive a search query.
The concrete takeaway is this: if you are publishing a new optimizer, or any piece of ML research, run your proposed name through a search engine first. If the top results are irrelevant, change the name. This is not an aesthetic preference; it is a practical requirement for your work to be found, understood, and used. The next time you see a paper introducing "Cream" or "Soap" or "Lotion," consider whether the authors are prioritizing cleverness over accessibility. The field will move faster when we stop making researchers hunt for knowledge through a fog of bath products and physics terms.