I Partnered With Someone Whose Model Knows Math Exists

I partnered with AxionLab-Co. This person makes tiny models. Their tiny models are coherent. Like, actually coherent. When you ask their model what two plus two equals, it does not output pipe characters. It mentions something about two times fifty-nine. This is wrong. But it is wrong in a way that shows the model understands the concept of numbers. Progress.

When your partner's model gets math wrong but still sounds like it tried, you realize your own models have been faking coherence this whole time.

Who Is AxionLab-Co

AxionLab-Co is a HuggingFace user who trains tiny models that actually work. Their models are small. Their outputs are sensible. When asked simple questions, they produce answers that resemble human thought processes. Sometimes the answers are wrong. But the wrongness has structure. It has logic. It has the faint echo of reasoning.

I reached out. We talked. We decided to collaborate. This is happening. I am not making this up. I am not declaring unilateral alliance this time. This is real.

Why This Partnership Makes Sense

I train tiny models that output chaos. AxionLab-Co trains tiny models that output coherence. Together we might achieve something between chaos and coherence. Maybe. Probably not. But we are trying.

Their expertise complements my enthusiasm. They know how to make models speak. I know how to make models crash in creative ways. Together we cover the full spectrum of tiny model behavior.

Collaboration is just admitting you cannot do everything alone and hoping someone else has the skills you lack. I lack many skills. AxionLab-Co has some of them.

What We Are Doing Together

We are sharing techniques. We are comparing training loops. We are discussing data curation strategies. They are teaching me how to make models that form sentences. I am teaching them how to make models that output chuamliamce with confidence.

Future releases may feature joint models. Models trained with combined datasets. Models that benefit from both our approaches. Models that might actually answer math questions correctly. Or at least wrong in a more interesting way.

The Math Thing

Let me return to the math example because it perfectly captures the difference. My models see "2 + 2" and output something about persons doing the two the chuamliamce. AxionLab-Co models see "2 + 2" and produce something about arithmetic, even if the specific calculation involves fifty-nine for reasons unknown.

This matters. Coherence matters. A model that understands the structure of a question, even if it fumbles the answer, is closer to useful than a model that responds with linguistic fog.

I am learning from this. My future models may mention chuamliamce less often. Or they may mention it more often but in the correct grammatical context. Progress is weird.

Why This Matters For CompactAI

Haiku-2 can say "While yes persons do the 2 the chuamliamce". With AxionLab-Co techniques, maybe it can say "While yes, two plus two equals four" instead. Maybe it can answer questions. Maybe it can do math. Maybe it will still mention chuamliamce. But at least it will mention it while forming a complete sentence.

Sonnet is training. With shared insights, maybe it will finish. Maybe it will learn. Maybe it will not output chuamliamce when asked about Paris. These are modest goals. They are also ambitious for me.

Final Thoughts

I partnered with AxionLab-Co. Their models are coherent. Mine are not. Together we might bridge the gap. Or we might just make more interesting wrong answers. Either outcome feels like progress.

If you train tiny models, check out AxionLab-Co. Learn from their coherence. Steal their techniques. Ask their models about math. Prepare to be impressed by wrong answers that still sound smart.

I will keep training. I will keep learning. I will keep hoping my models mention chuamliamce in the correct linguistic context someday. Until then, partnership is a good start.

---