Is there a way I can try out induced fit on my machine, can I make decisions easier for myself. Crons that run regularly, taking in immense data, churning and thinking with a bunch of tokens, and then surfacing to me with a list of options.
Gwern in guardian angels, what's his argument, how can I take it further?
See Induced fit.
Research Raven
Find me interesting nuggets of information across the web, curate for me, tell me new things. Should have an understanding of my "taste", what I find interesting, beautiful, cool, novel, etc.
Posting Peacock
Convince me to post more of my work.
See what I have done, and draw my attention to the parts other people would find interesting.
Should have an understanding of an "angle" (ie what actually is interesting, what my voice is, etc).
But the thing is the models kinda suck at this bruh so maybe we need some kind of hard coded "signal" to grep for? Or maybe a couple passes, the first pass is a nugget of information.
Daily readings
- What to read today: ultrathink about my recent writing, what topics did I write about yesterday, what notes, etc.
- Generate a list of 3-5 options of what I should read today given specific goals. What is related to what I am currently interested in etc.
- In the list have a one liner about why. I can click (maybe multiple), and it can send them to my kindle, and schedule reminders to do them later in the day!
- Once I "start" reading something, lock my attention in on it, nothing else, only notes.
- Generate a short quiz to make sure I comprehend.
Writing output commitment
- Maybe it's just some visual indicator of my writing output for the day.
- Eg you still have 3,000 more words to write today! You have "one" productivity block left, opening X right now is blocked.
What essays do you want to write about today? This is currently done by me just following my energy. Is there a way to help me lower the friction to just start writing about something?
Outreach
Given topics I am interested in, who in my network should I talk to? This kind of feels like Posting Peacock: given the work, who would find this interesting, identify when the work is at that level for a private email, or a public blog post, tweet, etc.
- open threads manager
- find new interesting people to message depending on current work/interests!
Cogsec
- don't click that link, respond to that message
- hide specific posts
- hide ads
- hide specific UI elements I won't like
- hide all colors except for images? Force e-ink?
- hide notifications
- machine wide mute list of words? Banned from even rendering on my screen?
This is more for cogsec, and tbh twitter is probably the best and worst thing for my cogsec!
- Also have something remind me when I am on twitter for too long.
- Maybe I just have insane autoscheduled pomodoros?
- The agent is always watching! We can feasibly do this very well on Arch.
- I often find when I start playing music, it dips my productivity, I get distracted and it can take me a little to notice this is happening. Maybe this is where having a system that can tell me the time since my last output. What metric would this be? Words/lines per hour? Is there some streak mechanic?
I listen to the same tracks dozens of times in a row. I want to make something that will tweak the ambient sound based on things like my wpm, if I am working vs doing something passive etc. Sound could be a fun surface for a guardian angel.
See Playgrounds and a live ad-blocker.
A list of options
Ok step one I think is generate as many multiple choice options for myself in a day as I can, outsource the hard stuff to that.
See The surface or the system?.
LLM shapes
The biggest QOL things I think LLMs still need to have ironed out:
- interruptibility (live steering)
- multi agent collaboration in a shape that doesn't add friction (like a bureaucracy would)
- live updates on progress? What is it doing at each stage?
- being able to live steer in a direction that feels right
I think the biggest issue with the shape of current LLMs is that directing them doesn't feel smooth or natural yet. A lot of people like to say we are at the same stage the computer was before the mouse.
When I think of the total application space of this technology we have made barely a fleck of progress, I would argue barely 1% of the way there.
But I think the big unlock from a lot of these tools is that they allow you to move asynchronously and lower the friction to explore other directions and categorize.
For example text editors are linear, ideas are not. This is a problem.
Ideas are currently trapped in a linear sequence. You are limited by your level of self awareness to notice when you are deviated. Having a way to monitor this would be great! This is a good place to start. Then having a gentle nudge, an intervention, a gardener helping you prune.