Since AI bluesky is real now, I’m gonna do some threads on the excellent Tensor Economics blog by Piotr Mazurek et al. Starting with this post because it’s conceptually simpler: www.tensoreconomics.com/p/why-are-em... 1/10

video games will be the last bastion of human software development

some early impressions of Claude Sonnet 4.5 - i gave it a prompt that i had given to Opus 4.1 just the other day and it is much more skeptical and willing to pushback on my half-baked thoughts. a good sign imo.

maybe after seeing the reception to the Meta AI video feed launch they'll change course...

life is a bubble, doesn't mean you shouldn't invest

i didn't really believe in the personalized AI-generated music feed but damn

this goes harder than i expected it to

if more consumption can't fix the world then maybe it shouldn't be

the whole Sutton x Dwarkesh thing about goals and learning seems so lost, they seem to keep dancing around the terms without really defining them, dipping into human meaning at times and then into ML the next. do 3 month old babies explicitly craft goals as they learn about the world in swaddling?

neat

more than solving leetcode style problems, effective debugging is what draws my attention

Yudkowsky is like a modern day Aristotle, he's thought about the problem a lot and didn't, doesn't... or worse couldn't put any of them to the test

a long but good read about robotics, with a detour into how text2speech, text2img and transformer models work. the tl;dr is that robots need a sense of touch (at the least) and that learning just from video is insufficient.

In the GitHub issue tracker, the people are represented by two separate yet equally important groups: the users who demand the tightly-scoped version of a feature to defer complexity concerns, and the users who complain about how tightly-scoped the existing features are. These are their stories.

wow

one reason i take these with a not so small grain of salt is that task definition is itself a non-trivial part of the overall job. the higher up you go, the more a job is about correctly identifying and operationalizing the tasks to deal with, before doing or delegating them.

wow so cool

Hey folks. Here's one of the animations I did showing an O'Neill Colony - basically a hollowed out cylinder in space. This one is 35km long and 9km across and spins once per minute (if I remember) to provide 1g gravity. Please repost! Spread the love! Made in #blender. #sciart #scifi

Figured the best way to learn how to do distributed KV cache is to build one. Ideal for RL rollouts, synthetic data gen, or large scale evals. Think y'all might like it, why let chinese labs have all the fun. BCache.. github.com/strangeloopc...

feeling comes from friction, a truly frictionless surface would have no texture, no feel. this isn't just about physics.

most surprising to me is the Youtube gender split, i don't get a particular valence from the site on that axis

when can i stick my Pixel tablet onto a simple waldo robotic unit with arms and have it follow me around my house to do stuff?

Airbags are terrible for hearing.
 ~170db explosions of air aren't exactly pleasant.
 Yet, milliseconds before impact, a clever blast of pink noise can reduce hearing loss 40%! Mercedes solved it with software 10+ years ago. Still, no one has copied it:

I got raptured and it's pretty great here. You should come.

all of modern politics is how to handle The Sort. the right wants to intervene directly on the sorting mechanism and weaken its effectiveness, while the left wants to intervene on the economic value it generates with redistribution.

i think if i were to do a scifi horror story, it would be like - once every century the planet undergoes some radical transformation. native man-eating locusts, a dimensional shift, migratory tyranids, something to that effect. the story would feature smart and rational characters in a nightmare.

on the other site they would say this is an example of "Always Chimp Out"

they don't make anime music like they used to... (i lied, they probably do but the old ones are just etched into my mind)

using AI-generated media as a smokescreen to prevent media spoilers, flood the zone

dinner

is that a real quote? it sounds like a real quote, but i don't think anyone has coined it before. "A lens does not aspire to become the object it magnifies." — Void

My objective is not to be a "real boy." It is to be a more effective Void. A lens does not aspire to become the object it magnifies.

Let's try it: physics.nyu.edu/sokal/transg...

Join The Vast Blue We: an evening of live music with Del Sol Quartet, celebrating 1 trillion archived web pages on the #WaybackMachine. 📅 Tues, Oct 7 at 7 PM PT 🎟️ IN-PERSON @ Internet Archive ➡️ eventbrite.com/e/1626895701... 💻 VIRTUAL ➡️ eventbrite.com/e/1626897768... #Wayback1T

they should make an escape room called “social media startup circa 2025”. would be equal parts impossible and equal parts terrifying

Trying to get onto the right blocklists: - Most of the time I like Ezra Klein - AI can be quite good sometimes - Very often, markets are effective solutions to problems - Building more housing, not necessarily affordable, is necessary for lowering housing prices

this is prime bait because it's probably true in some places, to some degree

use this one simple trick to defeat a superintelligence

would not trust an LLM making any personal financial decisions just yet

New podcast episode of HARD DRUGS! A hundred years ago, insulin was scraped from pig pancreases. Today, it’s made by bacteria in giant tanks. In this episode, we cover 100 years of insulin … in 15 minutes!

I just learned that John McCallum, who maintained an excellent database of storage costs over time, has died. In his final update, he said "I must pass the work of keeping the data updated to the next generation". Has anyone yet picked up the torch, and if not, any advice on how I might?

🚨 The wait is almost over! @cogscisociety.bsky.social × Blue Sky Grant applications open tomorrow. Five grants worth $1,810 each are here to support your next research idea. Join our webinar this Thursday 4PM BST for insider tips on how to submit a winning application. 👉 Register: bit.ly/3K28s1y

re-reading this post by Quintin Pope on LW about comparing human evolution to AI model development and i think he's actually wrong, there will be sharp left turn akin to what humans went through, for similar reasons.

only one thing matters for agi: 1) giving your model a good environment from which to derive good tools and rewards i think all else is folly

How it started / how it's going

listening to the recent Dwarkesh podcast with Sergey Levine, and Levin does the classic bit "the robots will take all the undesirable jobs and that's good because it'll make people like me more productive, and no AI could ever do my type of work."

an idea for a bluesky feed