AI PC Setup

This is the fork in the road that determines everything else. Cloud AI tools (ChatGPT, Claude, Gemini, Perplexity): Run entirely on remote servers. Your PC just needs a browser. No GPU requirements. No model downloads. Works on a 10-year-old laptop. This is what 90% of people actually need. Local AI models (Ollama + Llama, LM Studio, Stable Diffusion, ComfyUI): The model runs on your hardware. Your data never leaves your machine. No subscription fees. But you need a capable GPU, significant storage, and some technical tolerance for setup. Why choose local? Privacy-sensitive work where you can't send data to third-party servers. Offline use cases. Research or development work on AI itself. Or you just want to run your own models without paying $20/mo forever. If you're not sure which you want: start with cloud. You can set up local tools later when you have a specific reason.

For NVIDIA GPUs (most common for AI): 1. Go to nvidia.com/drivers → search for your card → download the Game Ready or Studio driver 2. Run the installer -- choose "Custom" and select "Clean Install" if you've had driver issues 3. After install, open Command Prompt and type: nvidia-smi 4. You should see your GPU model, driver version, and CUDA version. If this command errors, your driver install failed. For AMD GPUs: 1. Download AMD Software: Adrenalin Edition from amd.com 2. After install, confirm ROCm support -- AMD's AI compute support is improving but is narrower than NVIDIA's Also check: Windows Update. AI frameworks often have dependencies on Visual C++ redistributables and .NET versions that ship through Windows Update. For cloud AI tools only: skip all of this. You don't need any of it.

Go to python.org → Downloads → get Python 3.11 (not 3.12+ yet -- some AI libraries lag behind latest releases). During installation: CHECK "Add Python to PATH." This is unchecked by default and it's the source of 80% of Python errors for new users. After install, open Command Prompt and verify: python --version → should show 3.11.x pip --version → should show pip 23.x or later If these commands error, Python is not in your PATH. Fix by: searching "environment variables" in Windows → Edit → Path → Add the Python install folder (usually C:\Users\YourName\AppData\Local\Programs\Python\Python311\). One more thing: install a virtual environment manager. The standard tool is venv, which ships with Python. Every AI project should have its own virtual environment so package versions don't conflict.

For machine learning / development work -- PyTorch: Go to pytorch.org/get-started/locally → select your OS, package manager (pip), Python version, and CUDA version → copy the install command → run it in Command Prompt. Example command: pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 Verify: python -c "import torch; print(torch.cuda.is_available())" → should print True if your GPU is set up correctly. For running chat models locally -- Ollama (easiest): Go to ollama.com → download the installer → run it. Then in Command Prompt: ollama pull llama3 → waits for download (4--7GB) → then: ollama run llama3 → you're chatting with a local model. For a GUI interface for local models -- LM Studio: lmstudio.ai → download → install → browse and download models from Hugging Face inside the app. No command line needed.

GPU VRAM is the bottleneck for local AI. The model must fit in VRAM to run at usable speed. If it doesn't fit, it offloads to system RAM or storage, which is 10--100x slower. VRAM to model size guide: - 4GB VRAM: 3B parameter models (Phi-3 mini, Gemma 2B) -- basic use, limited capability - 8GB VRAM: 7B models (Llama 3 8B, Mistral 7B) -- good general capability - 12GB VRAM: 13B models -- strong general capability - 24GB VRAM: 30--34B models -- approaching GPT-3.5 quality locally Storage: AI models are large files. Set up a dedicated folder on your SSD for models (not your system drive if it's small). Ollama stores models in C:\Users\Name\.ollama\models by default. You can move this by setting the OLLAMA_MODELS environment variable. RAM: Windows needs its baseline RAM for the OS plus whatever apps are running. For AI workloads, close Chrome (its RAM usage is embarrassing) and any unnecessary applications before running large models.

For cloud AI tools (ChatGPT, Claude, etc.), your browser is the interface. Browser performance matters. Enable hardware acceleration in Chrome: Settings → System → "Use hardware acceleration when available" → toggle on → relaunch. Install these extensions if you use AI tools heavily: - AIPRM for ChatGPT (prompt templates) - Superpower ChatGPT (conversation folders, quick access) - Sider (sidebar AI access from any webpage) Bookmark organization: create a dedicated "AI Tools" bookmark folder. Put ChatGPT, Claude, Gemini, Perplexity, and any other tools you use regularly there. Reduce the friction to open them. One more thing: if you use multiple AI tools, consider a password manager (Bitwarden is free, 1Password is excellent). You'll have accounts on 4--6 platforms and the login friction is real.

Cloud AI setup checklist: ✓ ChatGPT or Claude loads and responds to a message ✓ If using paid tier: verify billing is active and you're using the right model (GPT-4, not 3.5) ✓ Browser hardware acceleration is on ✓ Bookmarks are organized Local AI setup checklist: ✓ nvidia-smi shows your GPU correctly ✓ python --version shows 3.11.x ✓ torch.cuda.is_available() returns True (for PyTorch setups) ✓ Ollama: ollama run llama3 successfully loads and responds ✓ Model runs at acceptable speed (7B models should generate 20+ tokens/second on an RTX 3060) If speed is terrible: check that the model is fully loaded into VRAM (not RAM offloading). In Ollama, run ollama ps -- you'll see the model and how much VRAM it's using.