Privacy question

[dxcore35]

In your videos, you strongly emphasize a security-first approach and repeatedly state that the TELOS file should never be uploaded anywhere. However, even running Claude Code locally on a machine that has access to the TELOS file, and it is ending up on Antropics server, this is 100% exposing private information, from the file.

I would like to ask whether you could provide more concrete and technical guidance on how to safely run private LLMs in conjunction with a TELOS file. Specifically, I am interested in architectures or workflows where the model can extract only high-level or constrained information from the TELOS file, without ever transferring the file itself or its raw contents to any external server. The goal is to ensure that the TELOS file remains strictly local and that no unintended data leakage occurs during inference or processing.

Questions:

1. What is the safest recommended architecture for running a private LLM that interacts with a TELOS file while guaranteeing the file never leaves the local machine?
2. How can one technically restrict or sandbox Claude Code so it only receives abstracted or limited information rather than direct access to the full TELOS file?
3. What common mistakes or overlooked attack vectors could still lead to TELOS data exposure, even when everything is supposedly running “locally”?

[the-ebdm]

You can use claude-code-rotuer with ollama to run claude code with an entirely local model. Unless you have spent several thousand pounds on GPUs or just a couple thousand on an SoC with enough unified memory to run large models though, you're unlikely to see the quality you desire.

And on your point about attack vectors, prompt injection is the main thing to worry about - don't give it unfettered access to the internet or roam unknown software repositories that are sketchy - any instructions that might prompt your agent to ex-filtrate your info might be followed. This is more about not compromising your system rather than your data though.

Realistically, none of this is particularly sensitive - this guys personal Telos makes me cringe a bit but I also have things in my personal journals that might make others cringe. I keep my personal knowledge base repos private for this and many other reasons. But power to Dan for not giving a shit and plowing on anyway

[dxcore35]

Thank you for your ideas. Uh can you explain what kind of model you're looking for when running Ollama? I'm on a MacBook M1 Pro with 64 gigabytes of RAM, so I have enough unified memory, but what is the best agent model? Uh, it's a Devstral2 30B? Q4? The Telos file should be cringe it is the point of it, to be honest with yourself and expose your weakness, so you can harvest the benefits of Claude code "mentor"

[the-ebdm]

So I've built a kubernetes based CRD/operator pattern to allow me to distribute my models and many more patterns across multiple machines where I custom fine-tune various smaller models with the jsonl output that claude code quite handily creates for you (I had several GBs of training data even before I had noticed it was there) - and because it's trained on my own usage it has learned my patterns. I use Qwen 3 Omni as a base model because of it's multi-modal capability, audio to audio is handy whilst it's modifying my code in the background.

I'm never getting any model or algorithm's advice on how I live my life that sounds incredibly dumb. I use it as a tool to develop software and automate the tedium of day to day life. That allows me to ponder those more interesting things myself. Automating away your own personality isn't going to give you some magical depth, you need to work that shit out for yourself. If anything makes you "cringe" about your own behaviour, it should be fairly obvious how to stop doing that shit. I'm not sure how much simpler a "mentor" can make it - it's just a bit critical thinking

[dxcore35]

1. Every single successful person has an external mentor, which is actually a person. You barely hear about some successful person without the mentor. Here we speak, we are gathering the intelligence of the word so to speak to be your mentor, not to be your advisor, how to live, but to keep you on the path you define by yourself (also TELOS file is written by you not by LLM).

2. Aha, so you're using NVIDIA hardware when I see Qwen3 Omni. You are lucky you have hardware for this!
   I don't think the purpose of the TELOS file is learn your patterns tho. I think it's just to keep you on the "path" as I was saying. It's not really your patterns because most of the time you can work against your patterns. The TELOS file is here to ensure that you are following your plan and not getting distracted even by your own "patterns".

[the-ebdm]

Yes and I have great real mentors but they should be real people who have actual experience to call upon and reason with. A single context window is not long enough to form a coherent worldview.

As a former software engineer and now DevOps engineer, I see first hand how the weight bias these models to the average output in my fields of specialty, not the best, it's only with the right selective fine-tuning have I ever found optimal solutions. But as the expert in my domain I need to correct it and steer it - I treat myself as the model's mentor not the other way around. That way the model gets smarter and I don't get dumber in the process. Only a fool sees mediocrity and thinks it's a genius.

Yes, I've invested a lot in a distributed cluster of many Nvidia GPUs - only way I can run the large assortment of models that I do and fine tune them as well - it pays for itself though with how much more work I can achieve.

Good luck to you my friend but our visions are on opposite ends

[the-ebdm]

I'd look into Low-Rank Adaptation (LoRA) which is a technique to adapt models for specific tasks without retraining the entire model - it's how I can iterate so quickly without enterprise resources - you might have the hardware - I haven't looking into it on an SoC though