Been thinking about kv tagging in systems like kubernetes and how if hashtags supported #key:value you would have entity-attr-value triples.

LLMs are melting classification because they moved everything out of trees and into high-dimensional space And it's melting academics' brains because classification is the entire Western tradition But do LLMs make *ontology* easier or harder?

LLMs really like to write imperative code from scratch, and you have to set up adversarial loops to get them to factor out pure functions and re-use them.

reading old cybernetics books and i love how goofy their illustrations are

Kind of opposite ends of the design space. Want public data to be difficult to lock in, difficult to censor, difficult to control? Spread it around as broadly as possible. Want data to be private? Touch as few computers as possible, for as little time as possible, ideally point to point.

*shakes fist* moravec!!!

don't skate to where the puck is, skate to where the puck was in 2023

I've come full circle on Elm's approach of combining state AND effects into a transaction object. There are atomic update patterns you just can't do with separate effects github.com/gordonbrande...

They can't even give you new numbers. They just combine existing numbers.

been working on a couple of open source libraries and generally just enjoying life lately. gonna be fun to share these when they are ready.

Gen X and millennials really entering the Facebook Boomer era six months until a third of that cohort's profiles read some variant of I hereby declare that Sam Altman does not have my permission to use my pictures, information, or posts, both past and future. By this statement I give notice th

The computer now has a natural language interface and a replicator for software You think I’m going to do anything other than ensure this technology gets into as many people’s hands as possible? The fight isn’t whether we use this. It’s whether we can self host these models

Used Attie to make a STEEP horizon scanning feed. Nice.

As I said elsewhere I am basically radicalized about this. We fight now for personal computing and personal agency or we lose another decade to closed clouds We push now for an open internet and open models. Nobody is going to hand it to us because they’re nice

This tweet triggered a wonderful research rabbit trail exploring narrative tagging systems this morning. I think a lot of AI and agent stuff might benefit from seeing itself as a text-based adventure game and learning from the patterns that gamedevs have developed for these over the years.

For genres where feel matters: 1. Double it or Halve It. (The Sid Meier classic) 2. Nonlinear, bursty movement almost always plays better. Mario’s jump is fast, then slow. Taper. 3. Do as much on-button-press as possible. 4. Make / ask for every config value to be live editable while playing

Build your save object early, and make it *extremely* flexible so you can add onto it later. Similarly, make sure your arrays that contain the most important data do not also contain objects that can't be converted to strings.

Combat design one: Human beings are really only capable of 'feeling' one thing at a time. You basically get *one* meaningful gameplay interaction per tenth of a second.

if you're making something multiplayer, get the multiplayer part working immediately where however many players can be in one session and see each other's names. then everything after that is immediately obvious when it's not working. do not make something offline and convert to multiplayer later

We actually got way more realistic behavior put of Halo Infinite’s bots when we added a couple hundred ms of delay for them to actually act on newly observed information. It’s a good hack for ai optimization too!

markdown is eating the world (laudatory)

Latest thing on LinkedIn is diatribes against AI written with AI. Think I'm being trolled by the algorithm.

how long before Claude outages cause market crashes?

putting off this upgrade as long as possible

Ad astra per agentis

anthropic, releasing mythos with limited availability

Lightning and thunder require time; the light of the stars requires time; deeds require time, even after they are done, in order to be seen and heard.

if you're so smart, why aren't you burning more thinking tokens

llms are bad at reasoning through concurrent code and I feel like this reveals something about me too

The Web Stream and AbortController APIs are pretty hairy, but very useful!

Scaling coding agents beyond simple tasks... - Write a prd.md (goals, not implementation details) - Pull prd into Claude Code. Chat about high-level architecture and interfaces (no code). Save to spec.md. - Pull in spec. Have agent write tests. - Pull in prd + spec. Have agent write implementation.

there's nothing that says a blob mentioned in a record must exist in the same repo/pds as the record, right? I can put the blob anywhere I want afaik. This is where RASL would be helpful though. If it's not available from getBlob, where do you go?

I’m gonzo for AI and this is a good idea. Wikipedia belongs to a deeper pace layer. Critical infrastructure. Move slow, edit things.

I'm begging you to stop gazing into the ai generated basilisk... please go back to watching subway surfer tiktoks

We propose that ‘antifragility’, coined by Nassim Taleb, is a great high-level concept for ecologists of all kinds for parsing and communicating the effects of environmental variability. An antifragile property benefits from variability. We explore its use in two mathematical models. (2/4)

Who needs to grow tomatoes when we've got pasta sauce?

hubrismaxxing

Karma is real, but applies to systems more than individuals. But then again, when you realize that you're a system, and that you're part of a system, and that this system is part of you...

*me, having a second cup of coffee* "you can just do things"

an agent is just an actor. multi-agent is just actor model.

Every so often I imagine Leó Szilárd having his vision and then posting a warning to LessWrong and the epic cope industrial complex this would have generated online

my agent factory

claude has started doing a funny thing where it writes notes to itself using comments in the bash tool

taking this to its logical conclusion and terminalmaxxing

it's quite simple. phone is fake computer

opinions be like

people only have good epistemics for the things where having bad epistemics kills you fast

this is just to say that the most valuable thing it needs to find bugs is information over time, especially if you're dealing with some mutable system. with the right log granularity and a high level idea of what algorithm should be doing, it'll find it

forced to concede that "I" am 90% language

This approach seems perfect for the vibecoding terminal era bower.sh/smol-contract