Building Smarter Java Apps with AI: LangChain4J, Memory & Multilingual Games! скачать в хорошем качестве

Building Smarter Java Apps with AI: LangChain4J, Memory & Multilingual Games!

🧠 Part 6 – Thinking Modes, Qwen3 Experiments & Model Mayhem (ft. LangChain4J) What was supposed to be a “let’s build core game features” day turned into a deep dive into model behavior — especially around thinking mode. We swapped LLaMA 3.2 for Qwen3, poked at token limits, and learned just how picky some models can be about following instructions. This time, we: Tested Qwen3 (4B, 1.7B) instead of LLaMA 3.2 Discovered that enabling/disabling thinking is easy… if the model respects it (spoiler: Qwen3:4B didn’t) Saw that thinking mode boosts quality — but at the cost of much longer responses Played with limiting thinking tokens for stable response times (not so easy in LangChain4J — maybe a PR coming!) Got great prompt/configuration improvements thanks to thinking Tried gpt-oss… but, uh, no 16GB GPU here 🤓 Noted that Qwen3:1.7B did accept the No-thinking parameter Started planning a custom Ollama model (thanks ‪@cat-edelveis‬!) 💬 TL;DR: Less game dev, more AI model wrangling — and some solid lessons for anyone building with LangChain4J. 🎯 What you’ll see in Part 6: Live tests with multiple open-source models How “thinking mode” changes output and performance Token limit experiments (and why they’re tricky in Java) Model quirks & gotchas that will save you time later A few tangents, GPU envy, and the usual live coding chaos 🛠️ Stack Java 21+ · LangChain4J · Spring AI (sort of) · GitHub Copilot · IntelliJ AI Assistant · Vaadin · Ollama (Qwen3, GPT-OSS) 📦 GitHub repo: 👉 https://github.com/JohannesRabauer/ai... 🎥 Subscribe + 🔔 for real, messy AI experiments — live in Java, with plenty of surprises. 00:00 Intro 00:56 AI Models (gpt-oss, deepseek-r1 and qwen3) 03:21 Demo with qwen3 (thinking) 05:03 What i did this week 09:30 Evaluating the thoughts of qwen3 17:17 Deactivate thinking with Langchain4j 27:21 Deactivate thinking in ollama 33:00 Detour in system prompts 36:40 Stop thinking with a different model? 51:30 Bugfixing system prompts 55:00 Removing system prompts again ^^ 01:01:15 Not thinking or not visibly thinking? 01:06:00 Limit thinking via tokens (fail :( ) 01:14:55 Conclusion #Java #LangChain4J #LLM #LiveCoding #AIAssistants #Vaadin #OpenSource #Qwen3 #ThinkingMode #Ollama #LangChain #ModelTesting