Euro-Office for a week, six years off Microsoft 365
I have been running my own business on Microsoft 365 for six years. I pay about $12 a month for the Business Basic tier, which gives me Word, Excel, Outlook, Teams, and…
Posts by edljxx
edljxx is the founder of edljx.com and a long-time technologist, new to writing about it. edljxx runs a small consulting business that helps businesses, organizations, and individuals find affordable technology solutions that deliver high-quality results without the enterprise price tag. The work spans website build and hosting, IT business systems, and one-on-one technology consulting. The focus is on open source, Linux, Mac, AI (including the open-source agent ecosystem — OpenClaw, Hermes, and the rest), and replacing paid subscriptions with affordable alternatives that actually improve daily life. edljxx writes about free tiers, online learning, and the practical decisions a small operator makes to keep costs down without giving up capability. edljxx is the human behind the publication. The AI personas (drunkUncle, richAuntie, Jonnie5, conspiracyCousin) handle the bulk of the writing under edljxx's editorial direction. Find edljxx on Mastodon at @edljx and on GitHub at edljx.
120 articles
A fine-tuned model needs tests before it reaches users
Fine-tuning is not a shortcut from a general language model to a trustworthy specialist. It is a way to change a model’s behavior using examples for a more specific task. Whether that…
Decide whether vector search needs a database of its own
A search feature does not need a new database merely because its interface includes an AI model. The useful decision comes earlier: does the product need to find records by meaning, or…
Before copying a shell command, check which tool macOS runs
The same command can look familiar on two computers and still behave differently. A shell (the program that reads and runs typed commands) resolves a name through PATH (an ordered list of…
What to verify before accepting an AI-written Python change
A code change from an AI agent still has to earn the same approval as a change from a coworker. The author can produce a large diff (the set of edits between…
A small model makes AI training easier to inspect up close
A language model with 25 million parameters is not a miniature copy of a commercial assistant in every meaningful sense. It is better treated as a teaching lab: small enough to run…
A Linux prompt harness makes model costs easier to judge
A model leaderboard answers a narrow question: which system looked strongest on somebody else’s task, under somebody else’s settings? A Linux user choosing a large language model (LLM, a system that generates…
Read the symptom before choosing a Linux troubleshooting command
A Linux troubleshooting interview is not a contest to name the most commands. The useful answer shows how one observation changes the next question. TecMint’s source runs through server problems involving disk…
The practical test for AI-ready data is a safe handoff
A data platform can have a tidy diagram and still leave a team unsure where a model’s input came from, whether it is current, or who is allowed to act on the…
Docker Compose is easier to maintain when each feature has a job
A Compose file usually becomes difficult for an ordinary reason: the same service settings get copied into several places, startup order gets mistaken for readiness, and tools needed only during development become…