The Sealed Files — 7 Local AIs Ranked on Your Contracts, Health, Money and Hardest Emails, Every Receipt + the 3-Step Setup
The companion to the video. We ran 7 free AI models (Meta Llama 3.2 3B, Alibaba Qwen 3.5 9B, Google Gemma 4 12B, DeepSeek R1 14B, Microsoft Phi-4 14B, OpenAI gpt-oss 20B, Mistral Small 24B) entirely on one MacBook (M4 Pro, 24GB), Wi-Fi off, temperature 0, and gave them five private-life tests: a freelance contract with 4 planted red-flag clauses (plus an innocent control clause), a blood panel hiding 2 real problems in 14 numbers, a money question with a 401(k)-match booby trap, a manager-took-credit email, and a lie-detector question about a document where the right answer is 'nothing was decided.' INSIDE: (1) THE FULL RANKING — two perfect 22/22 scores and why only one (Gemma 4 12B, 27 tok/s in 7.6GB) is actually livable; the speed + RAM table from this machine. (2) EVERY TEST, EVERY MODEL — pass/partial/fail with verbatim receipts: the model that called an unlimited-revisions trap 'a common practice,' the one that computed a 122% return and argued against its own math, the 63,000-character think-loop that never answered, and the email that surrendered your credit. (3) THE FINDING — at temperature 0, direct questions about a document were safe 7/7; the hallucination hid in a SUMMARY instead. (4) INSTALL TONIGHT — the 3-step Ollama setup, the pick for 8/16/24GB Macs, and exactly where your data lives (one folder, no account, no server). All test documents are fictional, written for the battery; all quotes verbatim from the kept transcripts. Independent test from Hyperautomation Labs.
Free. No spam. Unsubscribe anytime.