A 144M-parameter diffusion language model trained from scratch
In about 50 hours over four weekend days, a developer used Claude Code to go from knowing nothing about diffusion LLMs to training one from scratch.
The full story
The poster went from knowing nothing about diffusion LLMs to training a 144M-parameter model from scratch in about 50 hours over four weekend days, using Claude Code.
They say Claude tended to hand-wave core concepts, flatten differences between papers and fabricate implementation details with confidence. They built custom Skills, such as a paper reader that traces formulas back to their assumptions, and read the full papers themselves while Claude wrote code, catching things it missed. They settled on roughly 80% AI and 20% human.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: A SELF-TAUGHT JOURNEY IN LEARNING DIFFUSION LLMs WITH CLAUDECODE
The maker's post on r/ClaudeCode.
The subreddit's moderators removed it, but the original is still readable (it can take a few seconds to load): archived copy of Reddit: A SELF-TAUGHT JOURNEY IN LEARNING DIFFUSION LLMs WITH CLAUDECODE (opens in a new tab)