Really cool concept. Use AI, build public data to train future AI.

How to avoid politics on Bluesky 🦋 [THREAD] When scientists left X, many hoped for a quieter space. But politics will take over your feeds unless you actively do something about it. This guide will help you if you just want to follow your scientific interest! mikeyoungacademy.dk/researchers-...

Open AI exec claims something akin to RLaaS is the future AI development. As I've argued, that's good because RLaaS is better for business and safety: www.lesswrong.com/posts/f3EgXB... x.com/theinformati...

From the Apple team. I'd like to see this applied to 1D text as well. arxiv.org/abs/2509.14476

we are the only usable social media platform — no, mastodon does not count — that is not controlled by a passive or active ally of the government

I think simple economics tells us that superintelligent AI is not going to be threatened by us. open.substack.com/pub/joshuaga...

Very cool, I'll add this to my post on AI scaling. Though note that the infinite compute asymptote also has to do with the entropy of the data. You can't push it down arbitrarily far with algorithms.

for all people mock it wikipedia is genuinely one of the wonders of the modern world

This one's in my to-read pile (H/T toughSF) arxiv.org/abs/2203.07925

My solution to having a lot of open tabs is to simply read them all.

Some TPOT and e/acc folks have started popping in again lately. Hi! Your unhinged posts and chatbot romances are essential to our healthy AI ecosystem, so please check out my old post on constructing your feed to make bsky enjoyable.

A good principle RE policy: restricting people from doing something *necessarily* makes them worse off. It doesn't matter if it's regulations or taxes or H1B or minimum wage. It may be worth it, but you need a compelling justification. The burden of proof is on those that would control others.

the internet has never been more real life than it is now and tpot's dumbest poster is in charge

Against defeatism.

I should point out that the H1B stuff is just him making up laws that he wants. So I don't take it very seriously.

100%. You can call something a benchmark, let AI engineers sweat over it, and 6 months later an LLM can get those questions correct 70% of the time. That's a revolution all on its own.

Truly the most decel president ever. Just completely anti-progress.

New geometric algebra video up. I do think we're gonna reformat all of math with this system. You can write most of physics with a ruleset you can teach to elementary school kids. www.youtube.com/watch?v=iSct...

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

That’s how censorship operates in the USA (and many places) government pressure is used to force people & groups to self-censor. It’s a constant pattern: The majority of censorship is self-censorship, but the majority of self-censorship is deliberately cultivated by an outside power. It’s the goal

China is not catching up to Taiwan in leading-edge semiconductor fabrication. Because of this, export controls really do work.

This wasn't insider trading right? Right?? bsky.app/profile/hars...

Sorta wonder if direct instruction + LLM-generated games can boost human training and performance significantly.

If domain adaptation provides a moat against LLM providers is a million (billion) $ question. Watching Cursor's trajectory is probably one of the best ways to answer it

Incredible contribution to our knowledge of global working hours. 1/ Women supply 35% of total hours worked globally 2/ Hours worked do not substantially with income across countries. BUT this stability is produced by reductions in male hours which are offset by rises in female hours

This kinda stuff could be pretty useful for making semiconductors. Industry already struggles to make extreme UV light.

I’m Not a Robot, a game about solving CAPTCHAs, is out now! good luck :) > neal.fun/not-a-robot/

my political beliefs: - everyone should have real freedom, not just the in-group - everyone should have a decent standard of living - abundance + tech go brrr - YIMBY urbanism with lots of trains - staunch bodily autonomy (drugs, transition, etc.) - immigration is good, actually

What if a 100-meter asteroid was discovered just two weeks before impact? The collision would unleash the force of thousands of nuclear bombs. Could we stop it, or would millions die? www.youtube.com/watch?v=dKm7...

Most foundational models use softmax attention, which scales quadratically with input length—a major bottleneck. Linear attention has existed since 2020, yet large-scale models rarely use it. Why? minimax-01 finally makes linear attention work at scale. Deep dive here: 📌 youtu.be/iRuvGU-Sk3c

Yeah the admin thrives off looking scary, not actually being scary. They're weak and need people to self-censor. So ignore them while staying prepared. bsky.app/profile/hars...

feed is all politics today please send good RL papers you've read recently

I reference data science at the singularity constantly as it's not just a model of why data science blew up, but also how any field or org can organize itself around data, code sharing, and benchmarks: arxiv.org/abs/2310.00865

Trains a model to choose among LLM responses to get better performance. With some work sampling and picking LLM responses could give current models a performance bump. arxiv.org/pdf/2509.06870

Dig energy is trying to make at-home ground source heating/cooling cheap. Combined with heat pumps, residential heat can be electrified. techcrunch.com/2025/09/09/g...

Using LLM's to clean up text datasets to remove sensitive information or malicious responses while trying to maintain diversity. Neat. arxiv.org/pdf/2509.08653

Hidden in this blog about the Muon optimizer is a bigger idea. NanoGPT speedrunning is an accessible, competitive benchmark for LLM training. That's huge bc open code, data, and competitive benchmarks are what started the AI revolution in the 1st place. kellerjordan.github.io/posts/muon/#...

@timkellogg.me is data-pilled: "I’ve recently learned and fully accepted that ALL major LLM advances come down to data. ... any meaningful progress has come from higher quality, higher quantity, or cheaper data." Solid linkpost! timkellogg.me/blog/2025/09...

Well that's good news.

Global farmland has been declining since the early 2000s — but where is that land going? Researchers from @oxfutureoffood, @OurWorldInData & @oxmartinschool Director, Charles Godfray explore rewilding, lab-grown food & the future of farming.

Jon from Asianometry visited SF again after a year away. Some interesting takes on the AI and semiconductor scene. A thread with some thoughts ... www.youtube.com/watch?v=cru2...

No doomers, no crab buckets, no techno-pessimists in this tent. The future will be better and weirder.

1. I haven't seen anyone in my feed celebrating the violence. Though I'm sure it exists. 2. Digging up niche, despicable views on social media and being mad about them harms the epistemic commons. People should stop doing it.

Feeling sick to my stomach on hearing that Charlie Kirk has died. Political violence like this makes us all worse off.

My feed is almost entirely people condemning violence and telling others not to celebrate this.

So Russia shot 19 Shahed drones at Poland. Probably just to see what they can get away with. It's only going to get worse. Poland should respond by providing air defense over western Ukraine.

So true. Get rid of these regulations, let people cook. As an aside I kinda think if we had cheap/good meal prep ($1/meal, zero prep time, quality equivalent of a $30 meal today) we would be close to a limited form of utopia. Accelerate robochefs please. abio.substack.com/p/america-co...

Ozempic is a form of value drift. It changes your personality and motivation slightly. All tech does this, and it's probably a good thing, but something to consider.

I don't understand why results like this don't tank NVDA stock. For a given performance, models are shrinking rapidly. At some point, gap between small and frontier models will be negligible for your problem of interest. Better to iterate on the small model and enjoy the lower costs.

Very good news. forum.effectivealtruism.org/posts/dNSXKs...