After 4 months of daily use, here is where I landed with my AI cluster, “BOB”. - my digital brain
BOB is made from three computers. Everything runs on LM Studio.
The idea: three independent entities that can multitask, share outputs, and work together when needed.
1) Sam: A 2019 Mac Mini for simple and fast questions where security is not an issue. Running on OpenClaw.
2) BOB: A Mac Studio with 512GB of RAM. The main powerhouse, running qwen3.5-397b-a17b-mlx, modified and checked by our BottleCap AI team for biases we found during testing. Running on Hermes.
3) Jensen: An Nvidia Spark DGX running smaller models, currently Qwen 3.5 24B. It handles faster and simpler reasoning tasks (also checked & tweaked).
What Bob actually does every day:
Bob has processed a huge amount of private and public information about me: All podcasts, articles, old emails, documents, business history, decisions, projects, and communication patterns.
The goal is not to create a chatbot.
The goal is to have something that reasons like me, understands reasoning behind my past decisions, and has context for what I am solving today.
It understands how I think, what I care about, which mistakes I made, which opportunities paid off, and which ones were distractions dressed as opportunities.
I would never put this context into a random cloud system. It is too personal, too sensitive, and too strategically useful.
The killer feature:
Bob is trained to disagree with me from a hardcore first-principles angle.
It attacks weak assumptions, points out when something is not grounded, and compares new ideas against old decisions mistakes, with historical context.
Examples:
Bob found useful patterns in old Beat Saber music pack release strategy when we were starting and had very limited resources. It suggested better strategy straight up. Good lesson.
Yesterday, Bob suggested dropping an investment-related opportunity because it found a similar pattern in my 2020 history in different field. Back then, I had done something similar, with basically no meaningful benefit without remembering it.
Bob helps also with emails, but I do not let it reply automatically. I experienced few moments when somebody asked me why I rejected something and I had no idea what they're talking about. Now I send everything myself.
Bottom line:
For my use case, local models are the only relatively secure way where I feel comfortable going down this path and I use them daily.
They are still not in frontier cloud territory, but they are improving extremely fast.
For private reasoning over your own data, they are already good enough to become a real operating system. They just need to be checked.
Recent Anthropic restrictions on frontier cloud models showed how easily you can get cut off from tools your products or infrastructure might depend on.
That does not happen with local models sitting in your server or desk.
I strongly believe local models are a big part of the future, and a big opportunity for Europe btw...