2026-09-01
The Shape of Thought now has a blog. The paper on this site froze a snapshot of the Mycelium project; this space is for what the ledger captures day to day — registered predictions, honest negatives, and the occasional result worth telling as a story.
A taste of the current campaign: the project recently taught its certification machinery to grade 1,570 wild math problems against independent answer keys. The alarm bucket — cases where the parser confidently passed a wrong graph — came back zero. The same run confirmed the parser cannot yet read most wild prose (it was raised on synthetic templates and a narrower dialect), so the next generation of the model is now training from scratch on a diet that includes the wild text it failed on. Understanding must be learned; verification is built in. That division of labor is the whole thesis.
More soon.