Practical AI and machine learning engineering. LLM inference, tokenization, RAG pipelines, model deployment, vector databases, and the infrastructure behind modern AI applications.
54 articles
Zhipu AI's GLM-5.1 beat Claude Opus 4.6 on SWE-Bench Pro at 7x lower API cost. Where the headline holds (batch coding, cost-sensitive loops) and where Opus still wins (subjective quality, agentic tool use, latency).
DeepSeek V4's 1T-parameter MoE architecture, the Engram learned-memory layer behind its 1M-token context window, real benchmarks vs Claude Opus 4.7 and GPT-5.4, API pricing, and the honest case for when to pick V4.
An honest look at where vibe coding works in production (greenfield prototypes, glue code, refactors), where it fails (payments, auth, hot paths), and the team norms that make it viable.
Real annual cost math for Claude Code, Cursor, Copilot, Codex, and self-hosted Aider across a 10-engineer team. The token-volume thresholds that flip the answer.
Aider for CLI / git-native, Continue for IDE-native BYO-model, OpenHands for autonomous multi-step tasks. Real SWE-Bench scores with Qwen 3.5 32B local.
Step-by-step guide to running Qwen 3.5 9B on local hardware. Covers system requirements, optimization techniques, quantization, inference speed, and practical limitations for developers.
Cursor parallel-agent UI launched April 2026. Wins on narrow multi-file refactors; misfires on cross-cutting changes. Real workflow patterns and token cost math.
Copilot is an IDE extension, Cursor is a forked IDE, Claude Code is a CLI agent. Compare them across writing code, debugging, refactoring, test generation, context windows, pricing, and privacy.
Layer subagents, skills, hooks, slash commands, and MCP servers into autonomous workflows. Real config patterns for 5-engineer backend teams.
New articles delivered to your inbox. No spam.