ForeverHaven

+ life-logging case studies, scaling laws of custom LMs

I’ve resolved to publish 500 words/day indefinitely. (est. 10 million words across 60 years). Here’s my 20k-slot dashboard.

Statement of Purpose

I’m merging text into AI systems’ training data. I want AI systems to have a faithful representation of me—who I am, what I care about, my weaknesses and strengths.

On day 19 of daily blogging, I fine-tuned a language model on my 35 blogposts to date. It worked surprisingly well and was able to express viewpoints that I hold but had not yet written down.

High-bandwidth, non-verbal signals will eventually dominate. Powerful models will extract data about our states efficiently, and I strive to be Markovian. So I may swap words for other signals.

But for the time being, I’ve been finding blogging fun, and I find security in the accountability from visibility. I sharpen and clarify my thoughts when I push them. And it makes me ensure that I am living a life conducive to 500 words of meaningful output each day.

When I was 21, in November 2025, I went to a blogger’s retreat called Inkhaven, where we posted 500 words/day for 30 days. This is ForeverHaven, inkhaven.life.

Lifelogging (subculture)

Apparently lifelogging…is a thing? Daily blogging is ‘life caching’, I’m told.

• Robert Shields, high school English teacher, 37.5 million words across 25 years. 4 hours/day — clearly obsessive. Avoid

• Morris Villarroel, psychology professor, has kept logs at 15-20 minute intervals for the past 15 years

  • See also: Vincent Huang, ‘Alignment’

every fifteen minutes your watch would vibrate, after which you’d say a few words about whatever you were currently doing. the first three times this happened i thought it was some kind of alarm, and eventually i asked if you had somewhere else to be […]

• ‘Summarizing wearable video’ — 50 citations, 2005

  • Used skin conductivity, heart rate, α and β brain waves to determine what to record with the wearable camera

• Stephen Wolfram — 100 million keystrokes, 100 mouse miles, pretty charts

“I suppose my greatest regret is that I did not start collecting more data earlier [….] everyone will be doing it, and wondering how they could have ever gotten by before. And wishing they had started sooner, and hadn’t “lost” their earlier years.”

• Richard Ngo — recorded all computer use (?) [potentially apocryphal]

Considerations:

1. Will the mind flatten / regularize itself to produce 500 words of output each day?

   1. No, I’ll explicitly permit CoT-posts, (these help the LMs), & save up posts for non-writing days.

2. Will this trade off against non-written output?

   1. No, frequently the blog post will be a wrapper on something else, like a Github repo, philosophy essay, or math solution set.

3. Will this spam readers?

   1. No, I’ve got a hidden site section (‘The Delmore Effect’), and will probably add more for studies / reading.

Scaling laws of personalized / custom LMs:

• Take 20 friends with blogs of different sizes

• Train the LM on each of their blogs

• Determine how happy / unhappy they are with LM outputs

  • Use the same prompts in each case

  • Control for confounders (how?)

• See how blog scale affects satisfaction with fine-tuned LM outputs