muzzdang

Second Brain notes from YouTube & builds

Second Brain notes

AI Tools: I Don’t Trust Claude Anymore. Use This Instead.

Watch on YouTube: https://www.youtube.com/watch?v=RonPIktmLoQ

Channel: AI Edge

Why Local AI Is Becoming a Business Necessity

The conversation around AI privacy has shifted from theoretical to urgent. A recent video from AI Edge argues that relying solely on cloud-based models like Claude or ChatGPT means surrendering control over your most sensitive data — finances, contracts, health queries, and private thoughts. The creator’s thesis is blunt: “not your AI, not your intelligence.” If the substrate of your thinking lives on someone else’s server, you don’t own the output; you’re renting it.

What the Video Covers

The demonstration centers on a hybrid workflow that separates code generation from data execution. Cloud AI (Claude Code) writes the tools and scripts; a local runtime called QVAC runs those tools on-device, keeping raw data off the internet. Four concrete use cases are shown:

  • Bank statements and personal finances
  • NDA contracts and legal documents
  • Private voice notes — thinking out loud without a transcript leaving the device
  • Health questions answered by a specialized medical model

All of this works offline — on a laptop, a phone, even in airplane mode. The video also shows what it claims is the first on-device LoRA fine-tuning on a phone, letting a model adapt to the user’s writing style without any data leaving the handset.

Step‑by‑Step Walkthrough of the Hybrid Setup

  1. Install QVAC. The platform is a lightweight runtime (described as “like a game console”) that loads task‑specific open models. No special hardware is required for the smaller models; larger ones simply need more RAM.
  2. Load the specialized models. The video uses Qwen for general tasks, Tether’s medical model (claimed to outperform Google’s Med‑PaLM 2), and Transcribe Sai (built on NVIDIA’s Parakeet) for speech‑to‑text.
  3. Generate local tools with Claude Code. Ask Claude to write scripts that process your private files — categorizing transactions, summarizing contracts, transcribing voice memos. The code travels to the cloud; the data never does.
  4. Run the tools inside QVAC. Execute the generated scripts locally. Results stay on device, and the workflow survives internet outages or subpoenas.
  5. Optional: On‑phone LoRA fine‑tuning. Feed the phone model your notes or writing samples; it adapts locally, creating a personalized assistant that never uploads your style fingerprint.

A free setup prompt is provided in the video description to automate steps 1‑3 for Claude Code, Codex, or a plain terminal.

Tools, Models, and Numbers Mentioned

  • QVAC — Tether’s local AI runtime
  • Qwen — open general‑purpose model
  • Tether medical model — claimed to beat Google’s best medical LLM
  • Transcribe Sai — speech model based on NVIDIA Parakeet
  • Claude Code — used for writing the local tooling
  • On‑device LoRA fine‑tuning — first demonstrated on a phone
  • 300 M records leaked from 25 M users — cited as a real‑world breach affecting cloud AI chat logs

Caveats and Hype Check

The creator discloses a partnership with QVAC/Tether. Claims about the medical model surpassing Google’s best and the “first on‑device LoRA” are marketing‑forward, though live demos are shown. The hybrid Claude + local approach is practical and reproducible today. No revenue promises or affiliate links appear beyond the partnership disclosure.

Not your AI, not your intelligence — Paolo, Tether CEO

Your AI is now becoming the place where your thinking happens. And if that lives on someone else’s server, the most valuable thing you’re building — your own intelligence — it’s not really yours anymore

If this was ever leaked, would I want it public? If the answer is no, I’m trying to use local AI for that

A model trained for a single job that’s small enough for a phone — that’s sick

Concrete Next Actions

  • Download QVAC and run the free setup prompt from the video description to configure the runtime and models.
  • Audit your current Claude/ChatGPT usage: flag any finances, contracts, health data, or private notes that should move offline.
  • Build a local‑first workflow: use Claude for code generation, QVAC for execution on sensitive data.
  • Experiment with on‑phone LoRA fine‑tuning to capture your personal writing style without uploading samples.
  • Benchmark specialized small models (medical, transcription, categorization) against general LLMs for your specific tasks.

Portable, offline intelligence isn’t a futuristic concept — it’s a deployable stack today. The trade‑off is a modest investment in setup time for a durable gain in data sovereignty.