AI samo przygotowuje opis, tytuły i tagi z nagrań - offline na NVIDIA GeForce RTX 5080

https://www.nvidia.com/pl-pl/ai-on-rtx/ @NVIDIAPolska A local AI agent takes a folder with raw recordings and prepares materials for publication: transcription, description, title suggestions, and tags. Everything runs offline, locally on an NVIDIA GeForce RTX 5080—no uploading files to the cloud and no token payments required for each run. This material was created in collaboration with NVIDIA. This isn't just another chatbot in a browser. It's an agent that receives a goal, plans steps, uses tools, checks the results, and saves the finished files to the computer. In the video, I show the actual workflow: from the raw recording folder, through transcription on the GPU, to the final output ready for publication. ═════════════════════════ WHAT YOU'LL SEE ═════════════════════════════ • How a local AI agent works on real files, not just in a chat window • Why local AI makes sense: privacy, cost, control, and context • How to combine LM Studio, Qwen3, OpenClaw, WSL2, and faster-whisper • What pitfalls break such an agent? First-time workflow • How the RTX 5080 accelerates local LLMs, transcription, and GPU work • How Agentic AI differs from a regular chatbot • Why a graphics card can be a local AI accelerator, not just a gaming and rendering hardware ════════════════════════════ VIDEO STACK ═════════════════════════ • NVIDIA GeForce RTX 5080 • Blackwell architecture, Tensor Cores, FP4 • LM Studio + local Qwen3 8B model • Q4_K_M quantization + GPU offload • OpenClaw as an agent layer • faster-whisper on GPU • CUDA 12, cuBLAS, cuDNN • WSL2 + Python venv • LM Studio endpoint on Windows • exec / read / write as basic tools Agent ═══════════════════════════ CHAPTERS ═══════════════════════════ 00:00 Why a local AI agent? 01:24 Task: Offline agent takes the recordings folder 02:43 End result - what we're building 03:17 Architecture: LM Studio + Qwen3 + OpenClaw + Whisper 03:50 What accelerates RTX 5080 04:22 Setup LM Studio, GPU offload, server, and WSL 04:51 Trap 1: Model context length 05:41 Trap 2: Too many tools 06:14 Trap 3: Agent workspace 06:40 Trap 4: Noisy stderr 07:38 WSL2, faster-whisper, and CUDA libraries 08:13 Agent instructions and constraints 08:46 Running the agent live 10:02 Material analysis and description generation 10:21 Benchmark: VRAM, tokens/s, GPU 11:22 From raw data to finished output 11:37 Towards RTX AI PC 11:49 GeForce: AI, ray tracing, path tracing, DLSS, and AAA games 12:22 Tensor Cores, FP4, and the AI ​​tool ecosystem 13:31 Why local: Privacy, Cost, Control 14:57 The Simplest Launch Path 16:36 Summary: Goal → Plan → Tools → Result 17:00 Cooperation with NVIDIA and What's Next ═════════════════════════════════ LINKS ══════════════════════════ 📂 Code Repository: https://github.com/Kacpers/nvidia-fil... 🌐 Dokodu — Automation and AI Agents for Businesses: https://dokodu.it Hardware in the film: NVIDIA GeForce RTX 5080. #RTXON #NVIDIA #LocalAI #RTXAI