ā±ļø 4 min read ā€¢ šŸ“ 729 words ā€¢ šŸ“… Updated Sep 21, 2026
šŸ·ļø tools šŸ·ļø productivity šŸ·ļø ai

:material-brain:{ .anim-bounce } AI

I have a love-hate relationship with AI. I hate how wasteful it is and how hypocritical the frontier companies are. I love how it redefines the entire human-computer interface into one based on human language.

At work

I heavily rely on the following:

  • Claude (but GLM via fireworks these days)
  • claude-code as the harness
  • A /lot/ of tool use for access to internal apps/tools/infrastructure

As an SRE, a lot of my LLM use goes into information gathering, data analysis, & production troubleshooting.

Opinions#

I don’t care about MCPs much, I prefer tool use (I’ll admit MCPs are useful for centralized access to common tools).

I don’t care about prompt engineering much, I like iteration, iteration, iteration… Models become so much better over time that any effort in over-engineering your prompt is fairly pointless.

Talking to AI is much easier than typing. Microsoft’s Voice Access is sub-par; newer tools like Wispr Flow are quite nice. It’s easier to tell the AI to do everything rather than to do it yourself. “Write a spec file” becomes “Write your own spec please”, “Fixing a bug” becomes “Identify why you missed this, fix it, and ensure you don’t break again”.

With enough effort, AI is probably almost there for technical writing as well… BUT with the following tweaks:

  • I default to publishing my (unmodified) prompts in a side document. I hate reading somebody’s else slop and would much rather see their prompts (i.e. biases)
  • I expect to see a human name attached to the doc, so I know who signed off on the slop (or lack of it)
  • I (sometimes) ask the LLM to run the doc through an STE-100 pass; this significantly simplifies the language and has the happy side effect of de-slopifying the text

Timeline#

As with most people it started with simple toy projects, but these days I’ll admit it is an indispensable part of my daily work/play:

  • In late 2024 I was using it for harmless things like the random quotes at the bottom of this (and all other) pages: these are generated by a simple bit of JavaScript that ChatGPT helped create. It knew how to integrate this into MkDocs and generated a bunch of quotes too :)
  • In early 2026, Claude CLI with Skills is such a game-changer that it blows everything else away. I have not had this much fun playing with something in the tech industry in a /long/ time. It has automated a large chunk of my work and goes way beyond just coding and the chat interface.
  • In late 2026, I kicked it into a /much/ more interesting workflow for this site. The workflow goes like this: I add an interesting feature or bug in a Project in my Todoist. I initiate a skill in Google’s antigravity, and that launches a bunch of agents that read from the Todoist backlog, pick a feature to implement, and act as PM, Dev, QA, Code Reviewer and work together to deliver the feature. My role is limited to visually inspecting the local server and providing additional feedback
    • Here is the skill to manage the entire workflow.
    • So far the LLM has implemented UX improvements, easter eggs, bug fixes, site restructuring and so on.
    • It literally ported this entire site from mkdocs to Hugo in the span of 2 PRs (!)

Cons#

As someone who has a lot of respect for the Free Software community and who also enjoys drawing, it pains me to see how LLMs have scraped and stolen terabytes of content in order to train themselves. The hypocrisy of these companies is appalling. They also have the gall to whine when other LLMs distil their models and provide equivalent products at much cheaper rates (lol, schadenfreude).

I also hate how AI slop has taken over the internet.

Here are some thoughtful links that go deeper:

  • Chomsky on AI - His stance is very clear: as a scientist it does not have any value for him as it does not model the human brain in any way. And he does not care about the engineering aspects.

  • The Hater’s Guide to the AI Bubble - Ed Zitron really hates how AI is being hyped and he raises some fair points on the economic unsustainability of it all. I’m not an expert here but I suppose time will tell if he is right.

šŸ”— Related Notes