Tech Jupjup desk
Hands-on notes on AI tools, models, and open-source projects filtered for real work.
A current guide to why PageIndex says it can search long documents without vector databases or chunking, and how to read its pricing, MCP surface, FinanceBench claims, and real limits.
A current guide to what PlayMCP really is, how Kakao's toolbox and mcp-gateway work, and why the OpenClaw integration matters.
A practical NotebookLM guide covering free limits, Deep Research, Gemini Notebooks, file generation, privacy caveats, and when to upgrade.
A hands-on Recordly review covering auto zoom, captions, cursor polish, webcam controls, extension support, current bugs, and where it stands against Screen Studio and Loom.
A deep dive into the four bad AI coding habits Karpathy warned about, the CLAUDE.md repo's four operating principles, what evidence supports them, and where the limits still are.
A practical overview of EXAONE 4.5, HyperCLOVA X SEED, Kanana 2, A.X 4.0, Mi:dm 2.0, and Solar Pro 3, focused on licensing, deployment, and real Korean-language capability.
An honest one-month review of OpenCode + Oh My OpenAgent, covering ultrawork, agent roles, pricing, failure modes, and whether it is ready to replace Claude Code.
A practical review of Qwen 3.6 27B and 35B-A3B for local coding work, covering benchmarks, VRAM, token speed, Apache 2.0 licensing, and Claude alternatives.
A practical guide to harness engineering: context, tools, evals, observability, CI gates, and security for production AI apps.
What changed in GPT-5.5, where it leads on agent benchmarks, why pricing looks 2x higher, and how Codex users should think about upgrading as of April 2026.
A practical guide to HOP, the free desktop app for opening HWP and HWPX files without Hancom. Covers download links, installation, PDF export, and how it differs from RHWP.
How to run GLM 5.1 in Claude Code at one-fifth the cost, plus what Z.AI's official docs reveal about subscription-backed API access, Vision MCP, and GLM-Image.
Kimi K2.6 is an open-weight MoE model with SWE-Bench Pro 58.6, HLE 54.0, 300-agent swarms, 256K context, and API pricing about 88% below Claude Opus 4.7.
Make Claude Code read a knowledge graph instead of grepping every file. Install Graphify, compile your codebase + PDFs + docs, and shrink per-query tokens 10x — full walkthrough.
Claude Opus 4.7 benchmarks, pricing, GPT-5.4 and Gemini 3.1 Pro comparison, xhigh mode, and the real 1.35x tokenizer cost issue explained.
Hit RAG's limits? Andrej Karpathy's LLM Wiki pattern is a markdown-first alternative — 3-layer architecture, 95% token savings, exploding ecosystem in 2 weeks.
Honest Claude Code review covering installation, pricing, vs Cursor, pros and cons — everything in one post.
Gemma 4 model lineup, benchmarks, Llama 4 vs Qwen 3.5 comparison, multilingual performance, and Ollama local setup in one guide.
