-
Bias in LLM (AI-) shampoo comparison
Using Claude 3.5 Sonnet to compare different shampoo brands by analyzing their ingredients, with prices ranging from €5 to €30 per bottle. When brand names were included in the analysis, expensive "professional" brands like Kerastase ranked highest, but when ingredients were analyzed blindly without brand names, more affordable (or less popular) options ranked better, suggesting potential bias in the LLM's initial assessment. The experiment highlights how LLMs might be influenced by brand recognition and marketing, rather than purely evaluating product ingredients and effectiveness.
-
Thunderbird, Guix and GPG
How to get Thunderbird to recognize a Yubikey GPG key on Guix
-
Backup and decrypt WhatsApp Android
How to create a backup of WhatsApp, that you can decrypt and open on your computer, without rooting Android.
-
All in on Sway with Guix
Notes on what I have installed, and how-to fix things.
-
Bitcoin's black market rate
Two rates; Trusted Bitcoin with clear trail to KYC & Untrusted Bitcoin from unknown sources, offshore exchanges and mixers
-
Sway with Wayland on PantherX OS
A 3-step guide on how-to try Sway Desktop in a VM on PantherX.
-
SEEED reTerminal display stm32 stuck in boot mode
Attempts to get stm32 unstuck
-
Install Bluez 5.66 on Debian Bullseye (Pi4, CM4)
Install newer Bluez version and update RPI bluetooth firmware
-
Thunderbird with Dovecot on Guix/PantherX
Configure Thunderbird with Dovecot and a Self-Signed Certificate on Guix
-
Mullvad Wireguard VPN on RUTX11
Configure Mullvad Wireguard VPN on Teltoinka RUTX11 (OpenWRT-like)
-
Hello Nostr
The Nostr protocol offers a decentralized way to build services and platforms like social networks, crowd-funding sites, or Uber alternatives. It uses a single key for identity across all services, ensuring more control over your account and data. However, it comes with challenges like irreversible account loss if you lose your key and some client-side performance issues.
-
Develop your own LLM like ChatGPT with Tensorflow and Keras
In this post, I will show you how to develop your own LLM like ChatGPT with Tensorflow and Keras.
-
AI prompt engineering: Teach your device or robot new tri...
Imagine AI devices or robots designed to learn and adapt to your unique needs, offering personalized experiences and moving beyond traditional apps and vendor restrictions. These versatile, connected solutions can be taught by users without the need for programming expertise.
-
AI will make us a lot dumber and less social
As access to information has become easier and cheaper, AI and chatbots are now changing how we communicate and learn. It's crucial to be aware of the potential consequences and ensure AI is used responsibly and ethically to avoid negative impacts on our lives.
-
Run llms like ChatGPT on your laptop's CPU
This post explains how to run large language models like ChatGPT on your laptop.
-
Accounting for busy people (with AI)
Beancount is a text-based, secure, and donation-based double-entry accounting software that has become more accessible with the help of AI language models like ChatGPT. By using ChatGPT, users can create Python scripts to automate importing transactions from various sources, making the software more convenient and user-friendly for those without programming experience.
-
All at once: I made a huge mistake
I ventured into developing PantherX OS, a cutting-edge open-source operating system, aiming to tackle the issues found in Windows and MacOS. Throughout the process, I encountered numerous obstacles and gained valuable insights from the experience.
-
Cheaters, liars, and firecrackers
Story about running a Kickstarter campaign for an open source Linux project, hiring a PR firm that created fake websites and articles instead of legitimate promotion.
-
Gitea behind nginx on Debian
Quick setup guide for Gitea 1.6.1 running with supervisor on Debian 9.x, using nginx reverse proxy and LetsEncrypt SSL.
-
GuixSD on DigitalOcean
Step-by-step guide to convert a DigitalOcean Debian 9.5 droplet into a GuixSD installation, requiring at least 2GB memory for successful compilation.
