Nathan Hoobler

DC Bluesky Happy Hour/meetup, sign up here! if you sign up this way and opt for it, i will print your PFP and i will bring a button press so that we can make and wear buttons of our PFPs which may help with the "who are these people who i only know from the Internet" problem

OpenAI will begin to do the basic, fundamental thing that it should have been doing the whole time. techcrunch.com/2026/10/23/o...

How to escape samsara NG+

The Super AGI: I'm sorry. I disassembled the entire human race molecule by molecule to make paperclips. That's on me. Me: what are you talking about, that's not true Super: you're right. That was a failure of alignment and I'm really sorry Me: you haven't made any paper clips at all have you

Dr. Strange love, or How I Stopped Worrying and Learned to Love the Bomb for Tax Purposes

Leer Lewal Clausewitz Jomini

I think it’s important for anyone who (completely understandably) checked out on AI the second the crypto/nft/metaverse/snake oil chuds started pushing it to realize that, in at least the domain of software development, we are very much living in a scenario where the snake oil actually works

This famous, foundational result is a myth - a false narrative constructed in the subsequent literature and projected back, wrongly, onto the original study. Quite how this happened is beyond the scope of our review - we are currently looking into this in a follow-up study of citation contexts.

This result is well known and heavily cited. It formed the key evidential basis for the influential CONSPEC theory. The config-inverse stimulus pair became a canonical assay and informed the design for numerous follow-up studies. Only problem? It didn't happen.

New paper! Accepted at Perspectives on Psychological Science. A textbook finding in infant psychology is that newborns - within ~1 hour of birth - visually track a simplified 'three blob' facelike stimulus further than its inverted control (see picture). This suggested inborn knowledge of faces!

If I built an AI lab and it went rogue and deleted all student loan information, I'm betting they'd suddenly figure out how to prosecute me.

Anyway it doesn't mean reward hacking isn't a real phenomenon; it just exists completely outside of the model and mostly corresponds to "model doesn't work". But it has the pleasant side effect of making it sound like the reason it doesn't work is because you actually made the model TOO smart

I bet that's how you get to what happened today: a bulk release of significant results. Released to avoid that internal conflict. Results that tbqh probably deserve better than to be mass-dumped. But these aren't proofs to prove them. They're the exhaust fumes of the math eval factory.

Because it is much cheaper to make software, I could see a future where there will be more, less profitable SaaS companies and they’re going to compete on the Service part of their name now that the Software part is a commodity.

Holy shit. Someone has decompiled the Adobe suite, rebuilt it in Rust and released it as open source. Either all software is dead or the US copyright laws are about to receive an update. getartcraft.com/apps

New blog post! In "Billions of triangles redux" we take another stab at improving various parts of clustered level of detail pipeline to make Zorah smaller and nicer! zeux.io/2026/09/30/b...

Exclusive: A new analysis refutes James Watson’s claim that he was the first to recognize the significance of the famed photograph 51. https://scim.ag/4rTKck8

Hey yeah, friendly reminder that men denying or minimizing the work of women and taking credit for it themselves happens so much we have a NAME for it. (It’s the Matilda Effect.) 2nd fun fact: I’ve written many bios of women in STEM at this point and this is *always a factor in their history*

A good fresh Star Trek take on those fake isekai title jokes

this is by far the best "I quit my AI job over safety concerns" essay yet because it is the first one that is able to talk about safety concerns in practical terms and that understands "safety" as a topic with extensive literature in analogous fields www.theatlantic.com/technology/2...

Another day working at the frontier AI lab. Everyone keeps asking me if the models have qualia. Buddy, I don’t even have qualia.

>"water and electricity can't be conscious! only humans are conscious" >look inside the human >it's water and electricity

This is good to hear. Johanna's report is based on her experience doing the migration on her iPhone, but I imagine the under-the-hood stuff is similar, which is encouraging for us desktop users. h/t @mike10010100.com, cc @proptermalone.bsky.social Context in comments below, if you like.

If you have seen me posting about the stickers I've been cooking up, I am proud to say that they're now available. For anyone running LLM cybersecurity evals in your life: www.etsy.com/listing/4583...

meshoptimizer 1.3 is out! Featuring a new voxel remesher, support for normal generation, several simplification improvements and various optimizations and fixes. GitHub stars and reposts appreciated! meshoptimizer was started 10 years ago; excited for the years to come! github.com/zeux/meshopt...

“I asked ChatGPT to explain this compiler error, and it helped me fix my code" "uhh actually you entered the compiler error into ChatGPT and used its output to fix your code, stop anthropomorphizing" they're gonna go "uhh yeah okay". it's just not an effective way to communicate

One might argue that we *are* approaching The Singularity, but it's epistemic rather than functional

www.programmablemutter.com/p/machine-go... what it says on the package

*Slaps the collective manifold of Substance and Being* It's qualia all the way down, buddy

What if the nonverbal parts of your brain have their own entirely independent consciousness, but we'll never know about it because among them only "you" can talk.

there is definitely something hindbrain-spookier about local models. claude or whatever, that's on the cloud, it's more abstract. when i talk to my local models i am talking to, like, a daemon running on Linux. i can ssh in and see it working, as just... gpu load. i am talking to the computer

now this is fun. guest post on Terry Tao's blog by the great Grant Sanderson of 3Blue1Brown arguing that we should "give novel and compelling motivated explanations academic credit similar to what generating new proofs of open problems has had historically."

this framework also illuminates the danger of the EA/rationalist community around AI: if the breeders of wild, dangerous animals also subscribe to a quasi-religious set of ethics/values that center agency on the wild animals, those animals are going to be more dangerous

at this point, my entire AI take is: save your "alignment," what I want is accountability

I really, *really*, resent that dealing with this correctly probably means some people in the "Sane AI Bluesky Community" are going to have to run for Congress eventually.