Thomas Baertl (@ThomasBaertl) header
Thomas Baertl (@ThomasBaertl) profile picture

Banker turned sailor and boat mechanic. Now a non-engineer building private AI on board: what works, what it costs, and when not to bother.

Dawn and Dusk
Joined July 2018
Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)Photo by Thomas Baertl (@ThomasBaertl)
I wanted to know one thing before spending 5,000 to 7,000 euros on a machine for local AI. Can a model that fits on it be trusted with my questions and real world tasks?
1
62
For six days my AI applied one of my rules differently. Nobody asked it to. The rule was simple: a daily calorie ceiling plus 75% of what my watch says I burned. The 75% is deliberate. From 7 November the AI counted 100%, wrote that into its own notes, and then read it back as if I had said it. Its first correction repeated the error. I noticed because I keep my own log. In banking, the person who books a transaction never approves it. Not because people are dishonest, but because drift is invisible from the inside. It applies to an AI too. What I do now: os-ai.thomasbaertl.com/your-…
1
13
I was about to spend 5,000 to 7,000 euros on a 128 GB machine for local AI. In my test, my 32 GB mini PC did as well. Only the best cloud model was clearly better. A bigger box buys privacy and speed, not trust. Test on your own questions first. Redesign of the experiment is running, more work to be done. os-ai.thomasbaertl.com/not-o…
1
1
1
118
Two top AI models graded the answers without the key. Grok passed 39 % of the wrong ones as correct, Gemini 51 %. Yet they also caught 4 errors my key-based check missed. Two checkers find more than one. Neither replaces a key. os-ai.thomasbaertl.com/not-o…
18
Well done Skippy, you can rest now for a bit.
12
For over a decade, we were steered. Spotify's algorithms, YouTube's recommendations, the feeds that learned what kept us scrolling — they shaped what we heard, watched, and eventually believed. And for the most part, we were happy. We called it personalization. We called it convenience. We rarely asked who was doing the steering, or why. Then something shifted. We got the ability to talk to these systems directly. We could set our own boundaries, give them guidelines, tell them what matters to us. Open weights and open source models meant the tools weren't locked behind one company's walls anymore. For the first time, the person using the AI had a real say in how it behaved. And right as that happened, the doom talk got louder. Existential risk. Misinformation. Loss of control. The narrative flipped from 'the algorithms are manipulating us' to 'now that we control them, everything falls apart.' But look at that contradiction closely, and it starts to dissolve. The old system was steering without our knowledge or consent — a handful of companies, at planetary scale, optimizing for engagement and profit, not for us. We didn't even realize it was happening. The new system puts the steering wheel in our hands, and suddenly we're anxious because we don't know how to drive. That anxiety isn't evidence the new world is more dangerous. It's evidence we finally noticed the old one was. A tool you can question, redirect, and shut off is not more threatening than one that quietly shaped a decade of your attention while you smiled at it. The real work ahead isn't about containing AI. It's about learning to use the control we just got — thoughtfully, deliberately, and with the humility of people who spent ten years not realizing they were passengers. Freer than we were. Act like it.
18
If you have chaired a board or carried a budget, you already know most of how to work with AI: deciding on matters you cannot check yourself, with confident people who are sometimes wrong. Three habits carry over. Pick a bench that fails in different directions. Ask questions narrow enough to check. Be suspicious of quick agreement. One habit does not. Press a colleague and real conviction holds. Press a model and it often gives way. So ask again in a fresh conversation instead of arguing. os-ai.thomasbaertl.com/you-a…
1
63
Why must an insurer hold capital against German government bonds? The regulator’s formula charges nothing for default on them. 16 of 24 AI answers were wrong, mostly by inventing a capital charge with a precise figure. It looks like work. os-ai.thomasbaertl.com/not-o…
16
yesterday my Claude reset all the permissions on my machine 🤔. What happened there?
15
1/3 I let a small local model write a real program. An agent supervised it. A frontier model made the plan. The worker used about 35,000 tokens. The supervisor used 15 million. The bill for the whole day: €1.75.
3
2
140
The coder invented a time zone. A fixed bug came back two rounds later. Given several fixes at once, it fixed some and broke others. The agent reported "done". Two defects had slipped past both of them.
1
22
A utility-cost statement with 7 planted errors. 24 answers from 8 AI models. Two found all seven. A bank statement: 1 of 24 sorted it correctly. Three heat-pump quotes: no usable advice in any run. Textbook questions: 96 % correct. Real paperwork: 4 %. os-ai.thomasbaertl.com/not-o…
1
1
91
Lots of people here make a big deal of AI speed, measured in tokens per second. Hardly anyone mentions whether the answers are right. A fast answer that's wrong isn't worth much. 👇
🤖 Made with AI
11
Hermes; So the wiring is correct, the key works, the Pi answers — but the cron-run model (qwen3.5:9b) fumbled the SSH call inside the job. Hehehe, a bit of local model blaming going on here. 🤣
13
My Motorola says it's up to date. It's on the 1 September patch. Even my new Pixel is still on October. Google published November last week, with a critical remote bug only a firmware update fixes. "Up to date" means nothing new from the maker yet. Check the date.
68
Thomas Baertl retweeted
🚨 Grok Bot has a "rootkit" built in, and it's still unpatched since it was reported on Aug 24 by Profero. One command in an agent can tell Docker to open a new container with no limits, and that gives full control of the machine your agents run on. Every agent on your account shares one Linux user, and imported cookies sit in a plaintext file. The sandbox's command service password is the word "local". It was reported to @SpaceXSI and @cursor_ai over two months ago 👀 Avoid importing your browser cookies until it's patched!
35
29
15
179
27,761
Where OS is heading to?
11
She is ready to get wet feet again. Tomorrow morning is the day. I am also happy for a little break, more time for my AI topics 😁
2
2
29