Sleeves-up AI

LLM-generated SITA? Not yet.

I've been sitting on many months of SITA material, in part due to a stubborn fixation on using LLMs to drive my writing. It was a "learning experience" for most of that time, but efforts have finally started to bear fruit.

While all of the major API providers can produce serviceable "C+" writing with simple 0-shot prompts, achieving my "B+" standard for personal writing still takes a lot of iteration and duct tape.

My current workflow uses Obsidian Copilot and a battery of custom prompts and style guides. The plugin isn't very useful out-of-the-box (few AI tools are), but provides enough UI and primitives for me to cobble together task helpers such as:

• Expanding outlines and creating variations
• Edit my inline comments
• Polish and proofreading

I'll run these 20+ times over the lifetime of a draft, especially when I've lost "writing momentum." It's not life-changing, but I'd be moderately disappointed if this ability went away tomorrow.

AI is still very much a tool here, but the process and proof of concept has convinced me that AI will be capable of "A+" writing sooner or later — given a finite but tedious amount of user effort. One of the main reasons we haven't seen AI deployed more widely, likely boils down to the fact that most people aren't willing to spend so much time spelunking through rabbit holes to get to acceptable quality.

But as AI capabilities increase, easy-of-use improves, and we all become more skilled at context management...

"Hyperjournaling" / Looper Pedals for Thoughts

I don't identify with the "ADHD" label (if anything, I have the opposite affliction), but I still absolutely relate with having more ideas than time to chase them.

I envision something like a musician's looper pedal for thoughts — where I can set the baseline and structure for an idea, and then the resident AI (dare I say "agent"?) picks it up the background and runs it out to the next logical stopping point.

This likely takes the form of a journal or document space, where I brain dump notes and gradually relinquish control of those notes as the AI infers my intent. An intuitive UI for managing such journals and spaces would be an inbox, where I could review the AI's progress, assumptions, and reasoning.

This, I think, would be the next logical evolution of interfaces like Claude Artifacts or ChatGPT Canvas. And if they don't built it first, I might just get fed up and do it myself.

Armchair Speculations

LLM Sampler Settings

I've only done a shallow exploration of llama.cpp sampler settings so far, but this intuitively feels like a good research direction. LLMs produce a distribution of next possible tokens, but our algorithm for selecting the winning token seem quite... rudimentary.

Not that complexity is needed or desired here. But I sense that we're leaving a lot of intelligence on the table by merely discarding parts of the range and choosing randomly.

Technology's Lost Innocence

Politics is far outside of my circle of competence, but it's been interesting to watch "tech" undergo a political awakening the past few years.

The unspoken anthem of Silicon Valley is that technology gives introverted nerds a path to change the world — without dealing with messy people problems, such as marketing and politics. This was never actually true but directionally appealing (at least to introverted nerds who actively dislike marketing and politics).

But between AI, crypto, and regulatory pressure, the zeitgeist has shifted. And I sense an altogether different soundtrack playing in the background.

Links and Listening

A reporter tries AI friends — and found them surprisingly compelling.

On organizational dysfunction and why the optimal amount of fraud is not zero.

A guide to storytelling, with delightful examples.