Here's what your digest looks like

Show HN: I built a real-time collaborative code editor

892points · pgreenway· comments

Commenters are impressed by the CRDT implementation but debate whether WebSocket or WebRTC is better for the sync layer. Several suggest adding Vim keybindings.

Why we moved from Kubernetes back to bare metal

734points · devops_mike· comments

Strong opinions on both sides. Some argue K8s overhead isn't worth it for small teams, while others point out the move ignores multi-region resilience.

The hidden costs of LLM inference at scale

651points · ml_sarah· comments

Discussion centers on GPU memory optimization and batching strategies. Multiple engineers share their cost breakdowns running Claude and GPT-4 in production.

Anthropic releases new Claude model with 1M context window

1203points · dang· comments

Excitement about the context window, with many sharing benchmark results. Some skepticism about whether longer context actually helps in practice.

Fine-tuning open source LLMs on custom datasets: a practical guide

445points · huggingfan· comments

Useful thread with practitioners sharing training configs and cost estimates. LoRA vs full fine-tuning debate is the main thread.