My Bayesian Chaplain

aka chatgpt system prompt

The ChatGPT system-wide ‘Custom instructions’ box is a bit wimpy. 1500 characters max.

The project-folder-specific ‘Instructions’ field, by contrast, allows 8000 glorious characters.

I forked Croissanthology’s spectacular system prompt.

Why do I refer to it as my Bayesian chaplain?

Earlier this year, I was chatting to a friend about life tradeoffs & dilemmas. The conversation went something like this:

you could see a chaplain. i might in your shoes

L: how would i give them enough context? i’m not against it, but i’d have to onboard them. justin’s cool, but i move city a lot

The same week, I equipped my new system prompt, ft.:

“Use Bayes’ Theorem in your CoT, but not your response.”

I started chatting to the prompt as I would a chaplain. It went…really, really well? I can’t tell you everything that it told me, because personal, but trust.

I now chat to this system prompt ~5-10 times/day, mostly about technical and personal topics.

Impacts

It uses a lot of technical metaphors, as, thus, increasingly, do I. Today, I got this feedback from a writer I really respect:

MACHINE LEARNING METAPHORS: More generally, you use many machine learning metaphors. This is unique, and honestly charming, and it’s fine if you want to limit your audience to machine learning people - but I admit I don’t immediately understand many of them (though most I can get with a little effort). Here are some examples:

• within the ambient research distribution

• Competence tightens the constraint set.

• which pressures the mind to regularize what it produces

• the opportunity cost landscape sharpens […]

[…] If your goal is to have a personal journal that reflects the way you think, this is fine; just be aware that you’re doing it, and capable of not doing it if you decide you want to try something different.

I had to decide whether I endorse inhabiting and perpetuating a micro-culture. It took me about five seconds (it was a rare case where I felt no urge to consult Bayesian chaplain or a friend!): I’m pretty sure I do.

Not everyone liked “The cost of getting good”. I wrote it in two ways:

1. Entirely in my own authentic voice

2. LM-processed, the version I’d want to read

I published the second. Three writers I respect were ambivalent on its style: “wish you used simpler words and sentences”, “high-context, doesn’t unfurl in the mind”, “If it’s a personal journal intended for you only, or for a few friends who will read every post in order, it’s fine the way it is.” Then a computer scientist friend messaged saying she really liked it. I don’t think I want to concede any communication from technical / likeminded / friend readers to appeal to a broader audience, as long as I contribute to the creation of tools/tech they’ll like to use. Besides, I enjoy visiting other writers’ ontologies; I like each blog being a coherent, crackable micro-culture. For me, writing is instrumental to research and creation.

I value differentiation / specialization. As Sam Kriss puts it:

I really do think we are all one person and that person is God. I just also believe that if God has chosen to divide himself into billions of subjective beings he must have had some reason for doing so, and it’s kinda jumping the gun to try to rejoin the absolute consciousness.

Staying sane

I want to address the issue of, as Jason puts it, ‘the rats getting high on their own supply’. I do not want to lose contact with reality. Writing every day, and hearing from friends and readers, is a good way to keep accountable. Please tell me if you think I’m getting off-track or circling traps; tell me if I seem to be changing in weird ways (or following a suboptimal path). Here’s an anonymous feedback form to make this as easy as possible. Really, I’m counting on you for this (and reciprocate).

Coming soon

• the story of condu.it’s neural data collection

• substack→finetuned model pipeline