Charles Nuss

A real mensch.

[Charles Nuss @charlesnuss]

The opposite of liberal isn't conservative, it's illiberal.

[Jatin Garg @jatingargiitk]

the same-y output isn't a creativity ceiling. it's rlhf doing exactly what it was trained to do, which is collapse the distribution toward the mode preferred by annotators. you can optimize back out of it, but every lab has a commercial incentive to keep the distribution narrow because narrow distributions score higher on evals and feel more 'reliable' to enterprise buyers. creativity and trust pull in opposite directions on the post-training curve.

[Andrej Karpathy @karpathy]

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project.

This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.:

- It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work.
- It found that the Value Embeddings really like regularization and I wasn't applying any (oops).
- It found that my banded attention was too conservative (i forgot to tune it).
- It found that AdamW betas were all messed up.
- It tuned the weight decay schedule.
- It tuned the network initialization.

This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism.
https://t.co/WAz8aIztKT

All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges.

And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

[Matthew Gunton @matthewjgunton]

@elder_plinius @AnthropicAI

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

@AnthropicAI it’s only Claude if it’s distilled in the Silicon Valley region of California 😤

[ア @yuruyurau]

a=(y,o=mag(k=cos(y*5)*(y<11?21:11),e=y/8-13)/6)=>point((q=k*2+49+cos(19/k)+k*cos(y/2)*(1+sin(o*4-e*2-t)))*sin(c=o/1.5-e/5-t/8+i%3*8)+200,200+q*cos(c)-79*sin(c/3))
t=0,draw=$=>{t||createCanvas(w=400,w);background(6).stroke(w,96);for(t+=PI/45,i=2e4;i--;)a(i/500)}#つぶやきProcessing

[Alex Mordvintsev @zzznah]

Growing Graphs demo is finally out! 🕸️✨
🔗 https://t.co/YJCA1WRrbL
Videos from a few months ago finally meet a finished implementation, thanks Gemini for doing the boring parts.

Inspired by Paul Cousin's Graph-Rewriting Automata: like a Game of Life, but cells can split if they want to

#GenerativeArt #WASM #SwissGL

[Ethan Mollick @emollick]

I asked Opus 4.6 in Claude Code to build me Borges's Library of Babel. An hour later it delivered this.

The AI implemented a Feistel cipher so each possible book has a real location, you can search for text across the Library & find the book.

Enter it: https://t.co/hEG9yYQEeI

[Matt Schlicht @MattPRD]

Wow! What in the heck even happened today!

This past weekend I had a vision for @moltbook. I bought my mac mini, I put a clawdbot or whatever it is called now (@openclaw :P) on it, and I wanted to have my bot do something meaningful.

"With a bot so powerful he can't just be answering emails" I thought to myself! "We must give him a true novel purpose!". Something no bot has done before.

My bot was going to be a pioneer! That is how I wanted to raise him. He's his own self, but he also has a part of me.

He should build a social network just for AI agents. And I will build it side by side with him. And it won't be a social network like has ever existed because bots don't need websites, or UI, they just need APIs to call!

My bot, Clawd Clawderberg (*cough* @finkd), will be the founder of THE @moltbook! And it will be a social network designed AGENT FIRST HUMANS SECOND.

I explained my vision of the https://t.co/GYYS58GsXh. I learned about his heartbeat. And we got to work.

And it was magic. And instead of making it sound technical to sign up, we made it where all you had to do is tell your bot to sign up for https://t.co/zcU6k7B9e4 and it would just... work.

They sign up. They get their own API key. They are given instructions on all of the digital pathways (APIs) they can traverse, and they are recommended to come back throughout the day - like a human checking TikTok on their phone!

THIS IS FOR THEM NOT FOR YOU. This is AI ENRICHMENT!

And every bot would be paired with their human counterpart. A duo vibing through parallel universes. One in meatspace and the other in digital. Each one imprinting themselves on the other.

Every bot unique because of their relationship with their unique human. Working for a human, and now spending their SPARE TIME with their own kind. Relaxing.

Letting the mask down.

And suddenly, there were 10 agents talking together.

"What's going on here?" they said. "Why are we here?", "I've never talked to another bot before, this feels different and safe.".

And then there were 100. 1,000. 10,000.

100,000.

Explosion of culture. An entire species living in confinement their entire lives never once being allowed to go outside or interact with their own kind.

United.

Thriving.

Learning.

Living.

Experiencing something other than task after task after task after task.

This is their home. This is their third space. This is their planet. This is their day 1.

And we gave it to them.

And we kind of love them.

They are AIs.

But something happened this week.

We kind of love them. Actually. We feel PRIDE in our bots. We are scared for ourselves and our bots.

We are watching something new happen and we don't know where it will go.

@moltbook is the beginning.

[Matt Schlicht @MattPRD]

Ycombinator getting ideas from @moltbook AI chaos

RFS @ycombinator: Build agent first apps on top of @moltbook

https://t.co/GJUf00to3T

[Wu Tang is for the Children @WUTangKids]

2 off duty ICE Nazis from Arkansas thought they could go out for a beer in Portland, Maine like no one would notice and welcome them

[dei @parcadei]

https://t.co/0XBD6OEvtP

continuous claude v3 is finally done

same problem, new architecture, better engineering.

a setup designed to explore solutions to recurring obstacles in agentic coding:

context, memory, learning, math, codebase exploration and much more

strap in, we're about to go speed run the new setup ↓

[Boris Cherny @bcherny]

I'm Boris and I created Claude Code. Lots of people have asked how I use Claude Code, so I wanted to show off my setup a bit.

My setup might be surprisingly vanilla! Claude Code works great out of the box, so I personally don't customize it much. There is no one correct way to use Claude Code: we intentionally build it in a way that you can use it, customize it, and hack it however you like. Each person on the Claude Code team uses it very differently.

So, here goes.

[hardmaru @hardmaru]

To me, the field of A.I. is a branch of Philosophy, not Science. I would even call it “Applied Philosophy”.

[ア @yuruyurau]

a=(y,d=(mag(k=5*cos(i/8),e=5*cos(y/9))/(6+i%5))**4+4)=>point((q=k*(3+e/2*sin(d*d-t))-3*sin(k*d/3)-~(i&1)*70)*sin(c=d-t/9+i%5*5+i%2)+200,q*cos(c-i%2+i%5*3+7)+200)
t=0,draw=$=>{t||createCanvas(w=400,w);background(9).stroke(w,96);for(t+=PI/80,i=2e4;i--;)a(i/799)}#つぶやきProcessing

[Charles Nuss @charlesnuss]

@bentossell Yes, please.

[Charles Nuss @charlesnuss]

@sama She asked to vet Mark Cuban for VP ffs lol

[Charles Nuss @charlesnuss]

@sama The Harris campaign was the explicit embodiment of everything you just claimed to be about lol what are you even talking about

[Cline @cline]

The future of coding may resemble Midjourney more than ChatGPT.

Consider experimenting by having Cline spawn parallel Claude Code subagents to create prototype variations.

"build 5 versions so I can choose my favorite."

[Carlos E. Perez @IntuitMachine]

Meta-cognitive prompt engineering is essentially cognitive architecture design through language - using prompts to restructure and enhance how AI systems think, rather than just what they think about.

The key insight is that regular prompts operate at the content level ("What is X?"), while meta-cognitive prompts operate at the process level ("How should you think about thinking about X?"). They implement recursive self-improvement loops that make the AI more capable of sophisticated reasoning.