A community publication
From kernels to clusters.
An open community for learning, writing, and tinkering on the infrastructure behind modern AI: inference engines, training systems, ML stacks, and everything in between.
A schematic inference step: model operations become scheduled kernels, the GPU executes them, and the resulting scores are used to sample the next token.
Latest article Agents
On Windows, Claude Code's Bash tool halves runs of backslashes before bash reads the command, except runs right before a double quote. Our coding agent kept breaking Python files it patched through shell heredocs, even with the trap written into its memory. This walkthrough turns that correction into a hook, then measures what changed across 102,298 shell calls, including what the block cost.
Read the articleArchitecture 10 min read
Training Systems 9 min read
Training Systems 12 min read
Trainers, optimizers, recipes, debugging large runs.
Transformers, MoE, SSMs, hybrids, and what's next.
Planning, tool use, multi-agent systems, memory, and orchestration.
More threads to explore
Explore a weight-bandwidth ceiling and a fixed KV-cache memory example across precisions, batch sizes, and GPUs.
Turn a model size and token budget into training FLOPs, GPU-hours, wall-clock days, and cost, using the 6ND rule and real GPU peak numbers.
Per-GPU VRAM breakdown for training and inference — params, gradients, optimizer state, activations, KV cache, with ZeRO/FSDP sharding.
Explore a synthetic attention map illustrating causal masking and recurring patterns.
More to try
Knowledge grows when it's shared.
A useful explanation. A hard-won lesson. A question worth exploring. If you're learning or building ML systems, there's a place for your perspective here.
One email per publication. No spam, unsubscribe anytime.