Rand Arete

Rand Arete

@nulldivide.com

algebraist | archer | acrobat we keep us safe 💖💜💙 he/they

1K followers925 following625 posts

neither the soup nor the cake can be trusted. one tells lies. the other is a lie. nothing is real.

Frances: Meh and Loving It@francesmeh.reviews

a good soup makes you feel like everything's right in the world. to be clear though, it's not. the soup lies.

i work in dev tools and infra. building tools is kind of my thing. i build tools for myself and for others. i build tools for building tools. the way things are going right now, though? exciting, but exhausting.

Rand Arete reposted

Bit of a long shot but do any of my followers have real human 3d artists they've worked with before? Need a couple of grey-box low-detail spaceships built out and would love to commission some Kenny-ish stuff #indiedev #3dart #shiptember maybe

sometimes i stop and consider myself like i'd consider an rpg character, and... fuck. i've got a *really* weird build. it's mostly intentional, and i'm not mad about it. but also, wtf.

sometimes i'm just a simple creature: sitting at the brewery across from my place, eating some tacos, reading a book, watching as a storm rolls in. happy.

Rand Arete reposted

This is the coolest article I've ever written. Genuinely. All of you should read it.

ngrok@ngrok.com

Compression and LLMs are trying to solve the exact same problem: predicting what comes next. @anniesexton.com's interactive essay walks us through the basics of compression to reveal its surprising overlap with every language model you've ever used.

agentic memory should be more than just periodic tool use to recall something from a memory store. working memory should keep track of concepts and tasks in flight. proactive memory should retrieve possibly relevant things as context evolves, even pre-warming/filling caches.

it feels like structured state space models and things like kalman filters have a lot of conceptual overlap with these needs, and i've built some experimental systems that do this (and work decently, with room for improvement), but i really need to do a deeper dive into the literature.

hard workout chased by a cold beer and a book. nice sunday so far. later... puppets doing a robot apocalypse musical? or something?

pffft. get thee some moar screens, my friends. embrace the infocalypse.

Arrdem@arrdem.tirefireind.us

I used to have a triple head setup but this is more useful

when you forget to eat all day, train hard several times, and then stuff yourself with sushi and sake.... that hits hard, in a good way.

agents love "correction as content" when writing docs or code, and end up with a ton of "it's now this, not this other thing [ad nauseam]." and of course, when they start a new session or compact, they tend to fixate on the corrected thing you've just flooded their context window with.

currently jamming on... 1. my long running agent: Abel 2. agentic state management: supastructure 3. learned mixture of models routing: waveguide 4. all the machinery on my gb10: phlogiston ... and RFLX, because thinking with holes never stops

Rand Arete reposted

New lecture for the book! Nominally about synthetic data, but mostly is a walk through of the distillation literature from the Hinton 2015 paper to multi-teach on-policy distillation of today! At 7.4 hours of video in my post-training brain dump and counting :) youtu.be/6nyJ8y8ghsE

youtube.comOn-Policy Distillation & Using Synthetic Data in Post-Training | RLHF Book Course, Lecture 7This lecture starts slow, but covers key trends and training methods that came out of advancements in synthetic data. The core of this lecture is the math behind on-policy distillation, the most…

very few things have caught me off guard and made me laugh in the way that this did

Grace@gracekind.net

Thomas Nagel’s what it’s like to be a fleet of waymos

ok, fallback to Opus is now closer to 80% of the time. is this a ridiculously overwrought safety system? limits on capacity to actually serve Fable right now? either way, it's just silly.

Rand Arete@nulldivide.com

hah. i've had Fable 5 fall back to Opus 4.8 nearly two thirds of the time, for very benign technical document review tasks. this safety model seems to be "see nothing, do nothing, say nothing."

hah. i've had Fable 5 fall back to Opus 4.8 nearly two thirds of the time, for very benign technical document review tasks. this safety model seems to be "see nothing, do nothing, say nothing."

Rand Arete reposted

Replying to @disconcision.com

Participants will work through a combined tutorial on value probes and functional programming in Hazel, and then complete some writing and debugging tasks; 2 hours total. Due to time constraints basic functional programming experience is required. Interest form here: docs.google.com/forms/d/e/1F...

docs.google.comLive Probes in Hazel: Study interest formThank you for your interest in our study. We are evaluating Live Probes, a new feature in the Hazel programming environment that displays runtime values inline in source code. As part of our tutorial ...

Rand Arete reposted

We’re doing a user study to find out how adding always-on live values alongside code effects writing and debugging experiences for functional programming. See reply for sign-up details