wtf is xnn-research?

This repository is a research notebook from an independent group called XNN LLC that studies how large language models behave inside, not just what they output. The core finding is a strange one: you can ask a model what language a passage is written in, and separately nudge an internal signal so the model claims a different language, even though it keeps writing in the original language without changing a word. In other words, what the model says about text and what it actually does with that text can be pulled apart. The team compares this to a similar idea published by Anthropic in 2026, but adds an important check: they also test a scrambled version of the same nudge, matched in size but meaningless in direction, and that scrambled version barely changes the model's answer at all. This tells them the effect comes from the specific signal they added, not just from poking the model in general. You do not need an account or private access to try this yourself. The repo includes a short, commented Python script that runs on a small public model using standard tools like PyTorch and the transformers library, and it prints out the actual result on a real run so you can compare. The authors are unusually upfront about what has not worked. They also searched for specific stored weights inside models responsible for certain behaviors, and so far found none that held up against their scrambled control, including one costly negative result they still chose to publish. Two earlier phases of their research were closed out as failures and written up honestly rather than dropped quietly. The repository also includes a set of files meant to timestamp and prove when the team reached certain findings, using cryptographic hashes of private documents, without revealing those documents. This lets them claim priority on an idea without publishing everything behind it yet. The project takes no position on questions like machine consciousness. It is written and maintained by Gleb Stepanov of XNN LLC, an independent research company rather than a university lab.