We Won The Copyright Thing And Axion Lab Is Struggling With Reasoning

Good news first. We won the copyright thing. The AGPL-3.0 violation case. The NautilusOS situation. The person who copied our CSS custom properties and called it a new project has agreed to proper attribution. The license is restored. The fork relationship is visible. The ecosystem functions as intended. I am relieved. I am also tired. Legal conversations at two in the morning are not my preferred training regimen.

Winning a copyright case feels less like victory and more like exhaustion with better documentation. I will take it.

The Win

The evidence was clear. Twenty identical CSS custom properties. Ninety-four shared selectors. Character-for-character identical inline styles. A first commit message that read "Clean snapshot of current state." The timeline did not lie. The code did not lie. The license violation was unambiguous.

We asked for proper attribution. We asked for the AGPL-3.0 license to be restored. We asked for the fork relationship to be visible on GitHub. They agreed. The repository now properly credits NautilusOS. The license file exists. The chain of attribution is intact. Open source works when people respect the rules. Today the rules worked.

Axion Lab Is Trying Something

Meanwhile Axion Lab is attempting to create Haiku-2-Reasoning. They are fine-tuning our architecture for structured reasoning. They are adding chain-of-thought scaffolding. They are training on distilled reasoning data. They are also producing outputs that look like this:

This is not a typo. This is not a formatting error. This is the model attempting to reason and producing tokens that resemble XML tags fighting with padding tokens. The think tags appear in the wrong order. The APS token appears without context. The PAD tokens interrupt mid-sentence. The output reads like a corrupted log file from a debugging session that went poorly.

Why This Is Happening

Reasoning scaffolding is hard. Chain-of-thought training requires careful data curation. The model must learn when to think and when to answer. It must learn to close its own tags. It must learn that PAD tokens belong at the end of sequences, not in the middle of philosophical reflections about cats and jobs.

Axion Lab is learning these lessons the hard way. Their training loop is correct. Their data is reasonable. Their architecture is sound. The model simply refuses to cooperate. It outputs think tags that never close. It inserts APS tokens at random. It treats PAD like punctuation. This is not a bug. This is the model expressing its confusion in the only language it knows: broken XML.

Teaching a one million parameter model to reason is like teaching a goldfish to solve algebra. The goldfish tries. The goldfish means well. The goldfish outputs <PAD> \n \n the <\nthink> cats jobs. >APS< PAD> and calls it a day.

What This Means

The copyright win protects open source. It protects attribution. It protects the ecosystem we all rely on. That matters. That is worth celebrating.

The reasoning struggle reminds us that capability is hard. Adding chain-of-thought does not automatically create reasoning. Adding think tags does not automatically create structure. Adding distilled data does not automatically create understanding. The model must learn. The model must fail. The model must output corrupted XML before it outputs coherent logic.

Both outcomes are educational. Both outcomes are part of the process. Both outcomes keep me humble.

What Comes Next

We will monitor the Haiku-2-Reasoning progress. We will offer feedback when Axion Lab asks. We will celebrate when the think tags finally close in the correct order. We will document the journey when the PAD tokens finally retreat to the end of sequences where they belong.

We will also rest. The copyright case is resolved. The legal conversations can pause. The CSS custom properties are safe. The license is restored. I can return to training tiny models that know water is made of H2O. That feels like progress. That feels like enough.

Final Thoughts

We won the copyright thing. Axion Lab is struggling with reasoning. The model outputs <PAD> \n \n the <\nthink> cats jobs. >APS< PAD> and calls it a day. This is the state of things. This is the journey. This is the blog.

Celebrate the wins. Support the struggles. Document the weird outputs. Progress is weird. Collaboration is harder. Winning feels good. Failing feels educational. Both are necessary. Both are very on brand for us.

---