How cmart's Blog Uses (and Doesn't Use) AI

This started as a boring policy disclosure of how this blog uses (and doesn’t use) LLMs, in the spirit of Blink if you’re human.¹ It ended up also being a quasi-advice post, worth every penny you paid.

(I’m sorry that this is my fifth consecutive post about AI, and that it’s meta.)

In short, LLMs help me blog in several ways, but not for writing, revising, or proofreading at the level of prose. So, every sentence of this blog is my human authorial voice.

Most of my blogging-related AI use is via Kagi Assistant, for two uses which I often inter-mingle:

1. To find information on the web. If you have been wronged by Google’s AI-generated search result summaries, or by wildly hallucinated raw LLM output, Kagi Assistant is worth a try. It will do higher-than-Google-quality web searches, sometimes multiple rounds of them, and then answer your prompt only after putting the most relevant results into its context window, and citing the source material for its load-bearing claims. This isn’t always successful, but usually it anchors the LLM output to ground truth reality. LLMs still lack epistemic awareness! So, I usually go look at source material before substantially updating my understanding of reality, or writing from that understanding. I weigh the need to do this (and the credibility of each source) in a vibe-y but hopefully-well-calibrated way.

A digression about use #1: much of the benefit is taming the enshittified, hard-to-browse web that I otherwise experience using a public VPN and JavaScript disabled by default. Kagi Assistant avoids many CAPTCHAs, login-walls, and having to enable just enough scripts to load what should be static content. Maybe this browsing habit leads to slightly more enshittification, but I believe the existence of this blog far offsets that. If more people behaved as I do, the open web would have more discoverable writing and fewer shakedowns.

2. To explore concept-space rapidly. What should an interested person know about topics X and Y? What do people disagree on? For a first blush, this often feels as useful as talking to someone with expertise, and is much more convenient. I don’t need to schedule lunch with the LLM, treat its time as valuable, or even convince it to talk to me. Of course, the failure mode that I sometimes painfully notice in retrospect, after my questions have moved from broad to specific and nuanced, is the LLM moving from faithful representations of truth to plausible-sounding bullshit. There is no clear threshold between these states. (In the architecture of an LLM, they are the indistinguishable.) So, it’s more like talking to someone who has some expertise, but also a sneaky ulterior motive to always give confident answers to your questions even if they must lie, and they are very practiced. Using Kagi Assistant (with web search enabled) reduces this failure mode drastically compared to a bare LLM, but doesn’t eliminate it.

So, the LLM is a gopher for background learning and a hypnotic rubber duck. It’s a power tool for inductive work, but it should not drive the agenda or bear epistemic load. (Does that sound like Claude? Ugh.)

I sometimes use my coding agent to help make charts and tables, like this one. This is not vibe-charting with an image generation model, whose output is hard to control and mostly wrong. It’s more like: here is some data, please write a script that renders a chart for it. Here is the script for the above-linked chart, and I’m sorry, LLMs reach for matplotlib instead of plotly or ggplot2. Of course I check to ensure faithful representation of the data! I also make many nitpicky manual changes, because even vision models lack data viz taste. But this still saves a lot of time and de-bikesheds many silly decisions (like which chart library to use).

I will sometimes, not for every post, ask an LLM for a critical review of a draft, not for style, but substance: here is the confused writing of someone who is definitely not me, please spot the issues. (Disclaiming authorship points the LLM’s sycophancy in a less-counterproductive direction.) Several times, this has stopped me from making a poorly-substantiated or misleading claim!

If I cannot think of a word or phrase for something, I’ll ask an LLM. But I think it’s tacky when someone uses too much thesaurus, so I try to do this sparingly. (Thusly I discovered “thesauraus abuse”.) Using the most incisive, evocative phrase for every concept is the stylistic direction of LLM output, and it feels gross coming out of my fingers.

I do not do any spelling or grammar checking, with LLMs or any other tools. Ever since ~2023, when I see a typo in informal writing, I get the warm feeling that a real person probably wrote this! Of course someone could sprinkle typos in their LLM output, but this seems uncommon in practice.

Corollary: you don’t need to tell me about my blog’s artisanal typos. Unless they hurt readability, I won’t fix them.

Finally, if you go back a couple of years, you’ll find a few posts that begin with obviously-AI-generated images. This now feels cringe and I have stopped. The images remain, mostly because I think we should be gracious about people’s views and habits evolving over time, and sanitizing this part of my past feels grosser than letting it be. That said, I’m still proud of this one.

Reading what I just wrote, it seems possible for someone to infer that I believe everyone should write according to these standards (or not at all). I would not say this. Agreeing with Dynomight, I think that depending on the purpose of the writing, it’s reasonable to end up stricter or looser than I have, with two caveats.

1. I always prefer to read your own writing, no matter how bad, instead of the output of your LLM prompt. If you have something to say but you feel short on writing skill or executive function, I guess it’s possible that if you’re speaking through an LLM and exerting enough steering of the result (e.g. iterated output and requests for revision), then publishing the result could be net-positive for the world. But if I can smell the LLM, it’s difficult for me to avoid binning it as plausible-sounding noise, or to feel that you’re respecting the reader.

Having just written that, I am reminded that acoustic cyclists who sneer at e-bike riders (’not a real cyclist, not real exercise’) are missing the most important thing, which is it’s not a car. Maybe I should be more supportive of new ideas, syntheses, and analyses getting out into the world, however that has to happen.

2. I feel uncommonly suited to using LLMs responsibly, because I know I am dispositionally suspicious (perhaps to a fault, but here it’s the correct level). I worry that many people are not like this (hence, AI psychosis). I worry that many people who feel “10x’ed” by AI (for coding, writing, or anything else) are incrementally losing whatever carefully-built, independent view of reality they had previously.

And for the ‘creative’ writing that I’m most proud of, AI could never. And at risk of this turning into a clip show, I’ll stop now.