// 04
From the trenches. Crash reports.
I post what I'm building while I'm building it, on X mostly. The crashes get posted with the wins. If a tutorial stops where it gets hard, my thread starts there. Mostly things going wrong: OOM crashes, NCCL hangs, evals that failed in ways nobody warned me about. I post them while I'm still annoyed, so they're honest.
“I'm 20, training a custom MoE LLM from scratch. Not fine-tuning GPT. Not wrapping an API.”
X · the thread that started it · 7,800+ impressions →
“Most LLM tutorials stop too early. They end right where the real problems begin.”
LinkedIn · on the gap between toy checkpoints and real training runs →
“Hit a wall nobody talks about: the 500GB fp32 master partition lands on-GPU with params-only offload. Here’s what actually fixes it.”
X · distributed-training war stories, posted mid-crash →
“If you build loosely with AI tools: how do you maintain actual knowledge ownership of what you build?”
X · a question I keep asking, and answering with documentation →
follow along: @AaronChakr47884 · linkedin
