If you talk about AI, you should run it yourself. My answer: a 13-inch laptop with Nvidia RTX and 8GB of GPU memory.
The slightly longer version: I’d had my eye on the device since early 2025. And I hesitated. And hesitated. Until it sold out at the end of 2025 — and there is no comparable device. Classic.
Then in February 2026, a stroke of luck: a good refurbished unit. Sometimes the universe does reward you after all.
Why all this? I advise on topics where AI is now starting to play a real role — for example: what does an information architecture in the PLM space need so that AI works sensibly there? Talking about such questions without ever having gotten hands-on with them myself feels wrong.
Experimenting locally also has a character of its own — no subscription, no API key, no privacy question. Just: do it.
If you’re planning something similar: sure, 8GB of VRAM should really be more. But for my purposes — learning and first steps — it’s entirely sufficient.
By now Ollama is running under WSL2, VS Code with Cline (via the Claude API) is set up — and a first small automation project is taking shape. Not production-ready, but instructive. And yes — it’s fun, too. Allowed to admit that.
If you advise on AI professionally, you should get it running yourself. Local experiments on your own hardware — without a subscription, API key or privacy question — build an understanding that merely listening can’t replace.