Best X Builds
← All interesting reads

Teaching nanochat to count r's in strawberry

A full guide to adding a capability to a tiny LLM: synthetic tasks, tokenization traps, and why a honeybee-sized model needs the data over-represented.

by @karpathy
AIEngineering

Related reads

ThreadAIAgents

AI-native thinking

@dotey

Baoyu's Tencent talk. Find demand on the model capability line, kill bad ideas with prototypes, then redesign so the agent verifies itself.

ThreadAIAgents

Worktrees and coding agents

@davidgomes

David Gomes on harness-managed git worktrees versus letting the agent own the tree. The second one tends to become abstract art.

ThreadEngineeringAI

What happens when a GPU writes memory

@finn_fergus

Fergus Finn follows an RTX 4090 STG.E from the warp through L1, L2, dirty lines, write-back, and DRAM. The store finishes in 6 cycles. The data gets to DRAM later.