Throwing everything at the same AI.
I used to do this too. One chat window, asking it everything: write articles, analyze data, check news, schedule tasks. And the result? It did everything, but nothing well enough.
Then it clicked: it's the same as running a company — you wouldn't have one person doing marketing, finance, engineering, and customer service, right? Same logic applies to AI.
Before I explain why I chose cloud APIs, let me break down what these two mean:
Local Model: You download the entire AI model — the program and all its parameter files — to your own computer or server, and run it on your machine. Common tools include Ollama. Pros: no API fees, better privacy. Cons: requires serious hardware, and it's usually slower.
Cloud API: Instead of running a model on your machine, you use an API Key to call someone else's service (OpenAI, Anthropic, Google, etc.), let their servers do the computation, and they send the results back to you. Pros: fast, typically stronger models. Cons: pay-as-you-go pricing — you pay for what you use.
I tried. Ran DeepSeek and Qwen through Ollama.
The verdict: Way too slow.
Local model inference speed is just in a different league compared to cloud APIs. If your system needs real-time responses (like a PM coordinating multiple members, or a data analyst needing results before the market moves), local models simply can't keep up.
But local models aren't completely useless — I kept Qwen around for running knowledge base indexing (Embedding). Indexing doesn't need real-time responses. Running slow is fine, and the important thing is it's free. Save the expensive models for tasks that need real-time reasoning.
Here's my design logic:
Step one: Start from your pain points. What problems do I need to solve? I needed crypto intelligence gathering and analysis, content creation and publishing, plus pushing forward on various development projects.
Step two: Break down the workflow. What stages are involved? Gathering news, analyzing data, writing content, reviewing and publishing, technical maintenance, project development, overall coordination — each stage requires different capabilities.
Step three: Understand each AI's personality. Yes, I said "personality." Every model has things it's good at and things it's not. You have to spend time with them to figure out who's best suited for what.
Step four: Put the right AI in the right position.