StrawberryMatcha

@walnutinfra | yes i do luvvv strawberry matcha & weird ai stuffs

[StrawberryMatcha @strbrmatchaa]

Econ can be an interpretiblity tool for the agents just as it has been for the humans.

Economics developed theories to explain the behaviors for centuries whereas RL allowed us to set the utility function(=reward function) explicitly.

This paper inspired me to think about that possibility: https://t.co/GdPL8u7E76

If you let agents bid for queued user claims, they show the exact behavior that econ calls as 'Nash equilibrium.'

Econ for agents are sth worth exploring. I'd like to do a small side project on this, would update later.

[StrawberryMatcha @strbrmatchaa]

@societies_ai Interesting. Governments would def want these if they become more accurate. What was the method to simulate personas?? I also wonder why it was 2,721.

[StrawberryMatcha @strbrmatchaa]

Just saw this thread and made a really simple slop detector using multimodal Jev. Interesting✨

We might see less slop with Jev yeaaah!

[Nathan LeClaire @dotpem]

can you use jev to de-sloppify? my early experiments suggest it could help

i suspect codifying our tastes and preferences into "semantic linters" will become common. jev seems to thrive on comparisons, so the LLM can explore multiple drafts with jev as a judge.

[StrawberryMatcha @strbrmatchaa]

@dotpem wow. maybe i should try make jev put boxes over the most suspicious sectors. it might help de-sloppify further.

[StrawberryMatcha @strbrmatchaa]

interesting ✨

there is a neuron that decides the model's self recognition.

[Sauers @Sauers_]

God neuron

[StrawberryMatcha @strbrmatchaa]

@Sauers_ wow. but how do you detect which neuron to steer & how much to steer? Is it using some kind of interpretibility tool?

[StrawberryMatcha @strbrmatchaa]

what is happening. can't believe my eyes.

[Elon Musk @elonmusk]

Turns out the future does have flying cars

[Nick @nickcammarata]

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had

with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works

https://t.co/zHij5zwKGT

[StrawberryMatcha @strbrmatchaa]

should watch this. super cool.
generative ui is near.

[Cheng Lou @_chenglou]

My dear friends, I’d like to show you my latest toy project

Introducing Handwired: turn your regular JavaScript code into a neural net!
- no training
- exact outputs vs the JS program
- fully differentiable

All runnable in a browser, without GPU

[StrawberryMatcha @strbrmatchaa]

@doyc_1 werent they the ones who predicted the ny election??!

[Lukas Ziegler @lukas_m_ziegler]

About a robot that became a 3D printer! 🖨️

A team of designers in Shanghai developed a 6-axis robotic 3D printer inspired by spiderwebs at Tongji University.

The project is about merging design and fabrication processes, drawing from spider threads' biomimetic strength and structure. 🕸️

The printer utilizes a unique spindle-knot extrusion technique, enabling self-supporting structures and greater spatial flexibility than traditional layer-based 3D printing.

The device features a robot from KUKA and a robotic end-effector with movable printheads, controlled by an Arduino-based system with precise temperature and extrusion controls.

Despite being a work in progress, the project shows the potential of combining robotics and 3D printing.

~~ 

♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

[StrawberryMatcha @strbrmatchaa]

pretty cool✨

loved the post talking about applying this to the animal language.

[Dominik Schnaus @dominik_schnaus]

DINOv2 has never seen a caption, and Qwen3 has never seen an image. We still aligned their embedding spaces without a single image-caption pair.

It even works when the images and the captions come from different datasets.

Project page: https://t.co/2YHSkXVz1D

⬇️

[Yifan Zhang @yifanzhang_]

Introducing LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches

https://t.co/jRYWtIIhSW

Democratizing RL training through gradient compression!

[StrawberryMatcha @strbrmatchaa]

@atomic_chat_hq @liquidai decision models are about generalized intelligence provided at such speed&cost

NOT just about speed

3B cant have same amount of generalized intelligence as Luna

[StrawberryMatcha @strbrmatchaa]

@rezoundous distilling.

[Kate Deyneka @katedeyneka]

these guys solved the 70-year-old Pierce–Birkhoff conjecture - an open problem in algebraic geometry - for under $500

wow!

[StrawberryMatcha @strbrmatchaa]

lowkey uncanny

[explos1ve @explosss1ve]

THIS IS GENUINELY THE MOST TERRIFYING THING I'VE EVER WATCHED

they took the full wiring of a fruit fly, 166,700 neurons, and plugged it into an android hanging from the ceiling

the fly it came from has been dead for years

1,000 artificial muscles, a skeleton, no face

the fly brain was built for six legs and two wings

this body has two legs, two arms and nothing to fly with

nobody told it

> minute 1, a twitch in the fingers
> minute 4, the toes start curling like it's testing what it's attached to
> minute 9, the knees lift one after the other, it's trying to walk on air
> minute 20, someone walks underneath it and the kicking speeds up, you can hear the cables creak
> minute 31, the lights go off and every muscle locks at once

it found the rhythm a fly walks with and forced it onto human legs

and when the lights went off, it did exactly what a fly does when a shadow passes over it, it froze and got ready to launch

then came the part that made the room go quiet

> it brought both hands up to its chest
> started rubbing them together
> slowly, over and over
> the way a fly cleans itself

there is no code in that body that tells it to do that

it's 166,700 neurons doing what they've done for millions of years, inside a body they were never meant to have

right now the cables are the only thing holding it up

it learned to walk in one afternoon

nobody wants to find out what it learns in a month

and the next brain they're mapping is a mouse, 70 million neurons, about 400 times bigger than this one

[marfin @marfinxx]

this is pure f*cking treasure

A Stanford research group found how to orchestrate Claude Opus 5.5 and GPT-6 Astra together, and it completely breaks the resampling wall

most developers think test-time compute means running 32 blind retries on one model. you burn API tokens repeating the same 2-3 error clusters in loops

Stanford's paper (arXiv:2608.05643) splits the compute between two different brains:

> breadth search: GPT-6 Astra runs high-entropy exploration (T=0.8) to discover divergent solution paths
> step audit: Claude Opus 5.5 scans each step at T=0.0 and catches the exact logic fallacy
> local surgery: Opus rewrites only the broken step instead of dumping the whole reasoning chain
> verifier-free consensus: majority voting locks the proof without relying on drift-prone PRMs

the benchmark metrics outclass standard Best-of-N sampling:

> 91.2% recovery rate on detected reasoning flaws
> under 1.8% regressions on correct branches
> MATH500 jumps from 29.6% to 58.0% (+28.4% absolute gain)
> AMC-2023 improves by +7.5% with zero reward models

Astra explores the maze. Opus repairs the logic

you stop burning tokens on blind retries. you run verified reasoning depth