Linting burns tokens but keeps internal patterns consistent

I want to talk about linting in my home-spun LLM Harness:

Whenever I have the LLM do any significant amount of work, a summary HTML page gets created that I call a 'Node.' Each node also gets a short description, mainly so when I ask my system "last week I think I asked about..." there is a place it can go look that doesn't involve putting all the HTML from the previous week in context.

A problem ever since I started doing this: the descriptions wouldn't stay short.

And even worse than descriptions occasionally being a little long, what seemed to happen is that over time the descriptions would creep and get longer, because once there was one longer description in the project that served as an example that my 'no long descriptions' instruction was not gospel-truth (kind of like the OpenAI Hugging Face message board shenanigans. The examples LLMs see of what previous sessions did are latent system prompts!).

The only lever I seemed to have over this was ALL CAPS and BOLD TEXT ALL CAPS in my system prompt. Until linting.

How linting works here is that my preference for short descriptions got coded up (by the LLM) into boring deterministic javascript that runs every time a new node gets created. If it fails? Fail message inserted into system prompt and it tries again.

Mcluhan-ism #1: we still call it 'wireless'!

"The word 'wireless,' still used for radio in Britain, manifest the negative "horseless-carriage" attitude towards a new form."
- Marshall Mcluhan, "Understanding Media" 1964

Understanding Media is weird, dense, and a very interesting read. The dated references to TV personalities and cultural events are confusing and I never know whether it's worth looking up whatever he's talking about. He also has this habit of making massive, massive claims without really any particular evidence. But all that said, there are just these absolute gems that make it all worth it. Like "wireless."

He's writing in 1964 about how wireless is a ridiculous term...and we're still using it 62 years later!

We know how this technology works, more or less, we should have a name that reflects this. I mean...it gets complicated fast with modulation, CDMA, the different spectrums and rules, all the conventions and protocols. But it's always waves, right?

Wave-ful, not wire-less. Or if we're so attached to wireless, we should add some other things that aren't in it too. Wire, dog, and parrot-less. Wire, carrot, and potato-less. Wireless, vegan, and made without any animal testing. Makes just as much sense. Just as there's no wire, there's no dogs or cats involved either.

Mcluhan-ism #2: photography -> painting, LLMs -> writers?

I won't quote in depth but Mcluhan makes the point I think others have made, too: when the photograph arrived, painting got weirder. Before photography, faithfully recording an image was one of the core things a painting could do. With photography around, painting adjusted and found new niches.

I value words less now. Both my own, and the words of other people.

I don't know what to do with that, but I've felt it for long enough now that I feel comfortable coming out and saying it. I value all words less. I want more from people, and I want to give more than tokens and clicks.

Mcluhan-ism #3: how many tokens should pass in front of a person's eyes in one day?

There’s nothing at all difficult about putting computers in the position where they will be able to conduct carefully orchestrated programing of the sensory life of whole populations. I know it sounds rather science-fictional, but if you understood cybernetics you’d realize we could do it today. The computer could program the media to determine the given messages a people should hear in terms of their over-all needs, creating a total media experience absorbed and patterned by all the senses. We could program five hours less of TV in Italy to promote the reading of newspapers during an election, or lay on an additional 25 hours of TV in Venezuela to cool down the tribal temperature raised by radio the preceding month. By such orchestrated interplay of all media, whole cultures could now be programed in order to improve and stabilize their emotional climate, just as we are beginning to learn how to maintain equilibrium among the world’s competing economies.
https://www.understandingnewmedia.com/mm1/class_materials/mcluhan-playboy.pdf


I don't want to reprogram society, but I am interested for myself: what is my ideal media diet? How many tokens do I want to try to process in a day? What ratio of human to LLM generated text am I comfortable with receiving?

Very roughly, I think there are tiers of tokens.

  • Tier 1: Human-to-human (me, specifically)

My favorite tokens are human generated tokens that specifically were generated for me. I don't get enough of those tokens in my life, and I'll always be wanting more of those tokens.

  • Tier 1a: Human-to-human (not me specifically)

Second favorite kind of tokens, just really good writing. The kind of writing where you feel delighted and valued as a reader.

  • Tier 2: Human-to-Machine , Machine-to-Human (me, specifically)

Most of social media is human-to-machine. We write not to other people, but instead write messages that we hope will survive (and maybe thrive!) amid the pressure of an algorithm.

All LLM text is machine-to-human. But at least when I'm doing it to myself, it's LLM text specifically for me.

The trade-off between human-written but not for me vs machine-written but for me? Unclear to me.

  • Tier 3: Machine-to-Machine

The garbage tier.

Sci-Fi Job Posting (in a world that takes tokens seriously)


Generic knowledge Worker

  • Pay $80,000-120,000

  • Perks: free coffee & snacks, no compulsory daily LLM use!

  • LLM token budget:

    • 0-1,000,000 / day

    • no daily compulsory LLM use, but we do expect you do burn at least 1,000,000 tokens a week

  • your token expectations:

    • output: 2-20K expected daily output

      • inclusive of all synchronous meetings + messaging within company (average daily scrum output per worker is 1K tokens, for reference )

    • input: 20-100K.

      • we proudly guarantee minimum 25% of your input tokens will be human generated!

      • As an additional sweetener: overtime pay automatically triggers on any day your token responsibilities surpass 100K

      • We have insurance for mental errors made my employees who have had to process more than 80K tokens in a day

(numbers totally made up)