We should share our LLM use more
By sharing I mean:
sharing personal AI policies/preferences, formal or informal (I care most about disclosure. I write every word of this blog + all social media + all correspondence. LLM text seeps into the software I make since I find it a pain to remove it all)
using LLMs with another person, over zoom or even (gasp!) in the same room ( mea.media was a delight to make for this reason, making it IRL alongside Elle Griffin)
sharing rough token usage (I'm on the $100 / month claude plan. Lately I max it out in 4-6 days and then enjoy 1-3 days of peace)
sharing our anxieties ( What a time to be alive! I don't think this LLM + society integration business is going well so far.)
sharing our hopes
( I want:
one attention surface where if I check that spot I know I haven't missed anything I care about
more time to make art + music with friends
climate tech + health tech + I can understand to participate anywhere in my life I care to
for "being smart" to not matter so much)
How we currently share
Raw Transcripts
The transcript is what happened, it sometimes makes sense to share these. Especially when it reveals something hard to describe.
Kevin Roose's raw Sydney chat transcript was published verbatim in the NYTimes (interesting re-read in 2026, fwiw)
Terence Tao talking Jacobian Conjecture I found interesting to scan. A good reminder of the latent capabilities in these models, and interesting to see his prompting strategies.
(maybe I'm just nosey, but I would glance at a random LLM transcript of anyone. What kind of prompter are they? Similar level of curiosity to glancing at the bill your friend picked up to see how they tip)
We share screenshots of funny turns of phrase, or examples that elicit interesting behavior ("How many R's in strawberry?" type stuff).
We share transcripts when anything remarkably bad happens.
But raw transcripts are mostly too large to read. Too large even to meaningfully throw to LLMs, since throwing an LLM a raw transcript is basically just resuming that session.
Self-Report: "You know, Gemini and I had a chat about just this topic..."
Terrible. "Well that's not what Claude said..." is never something you want to hear.
We share our LLM use when we use LLMs, and then share our work
When we use LLMs, sharing our work becomes sharing LLM output. And so we disclose. But that's imperfect and prone to misunderstanding.
How we could share
With things like the substack-pangram integration we're saying as a society (rightly!) that there's digital places slop shouldn't be.
Complimentary to that could be a place where the slop should go. A hybrid space, human readable but also something you can throw your LLM at.
I see some good ideas close to this space.
I think Matt Webb's RSS for Vibe Coded apps is a very interesting idea.
tools.simonwillison.net is a neat pattern.
I've thrown my hat in the ring with my slog (slop-blog). Slog entries are stand-alone html+css+js packages that can be downloaded and thrown into any LLM system (or even web interface) to be remixed. Basically the stuff that falls out of my experimentation that I think might be interesting to others.
potential RSS-for-vibe soft requirements/features
social something
I don't care what strangers are using LLMs for. I am interested in what my friends are using LLMs for, because it's a signal to what they're thinking about.
existing projects can be converted one-shot ( /convert-to-RSS-for-vibe skill or something)
Lovable apps made by people who don't know SQL from sequel, tools.simonwillison.net items, slogs from my slog, it should be (IMO) Big Tent and not fussy.
remix as first class / machine-readable along with human friendly
You can browse it like a hackernews clone. You can click on anything and tell your LLM, "This, but make it more {adjective}." Or, "Make the app halfway between {app1} and {app2}."
You can also ask, "Is anybody thinking about / working in the area of {arbitrary interest}"?