#CSAIL
LightNews LightNews
yourdigitalbrain.bsky.social
What happened A team from MIT CSAIL composed of Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab published a blog post titled Sparks of Weight-Space Programming on November 2, 2026.
MIT's Weight-Space Programming: an LLM edits 0.016% of Qwen3.5-0.8B weights and boosts MATH-500 from 7.9% to 48.1% - Sinaptica
What happened A team from MIT CSAIL composed of Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab published a blog post titled Sparks of Weight-Space Pro
sinapti.ca
November 3, 2026 at 1:09 AM
seldo.com
This is an absolutely crazy thing to discover and not enough fuss is being made of it. There was no reason to believe this was even possible, far less that LLMs knew how to do it. *We* don't know how to do it, so they were not trained in it, they just figured it out.
Research paper title and abstract about LLMs reading and writing neural network weights, authored by MIT CSAIL researchers, describing capabilities in weight interpretation and modification.
November 3, 2026 at 3:19 AM
shadowfetchstudio.bsky.social
Spotlight on Daniela Rus 🤖 Director of MIT CSAIL since 2012 and a MacArthur Fellow, her research explores robotics, autonomy and collaborative machines. So inspiring!
aipeoplewiki.com/people/danie...
#DanielaRus #Robotics #AI
November 2, 2026 at 8:00 PM
shadowfetchstudio.bsky.social
Spotlight on Ramin Hasani, co-founder and CEO of Liquid AI, an MIT CSAIL spin-out. His liquid neural network research began with his PhD at TU Wien. 🙌

aipeoplewiki.com/people/ramin...

#RaminHasani #AI #LiquidAI
November 3, 2026 at 3:55 PM
wolf.observer
goddammit
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights published november 2, 2026 authors, all from MIT CSAIL: Authors Rikiya Takehi* Jayden Teoh* Alex Zhang Omar Khattab TL;DR We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms. This might be a big deal for both optimization and interpretability. Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do. And without any environment feedback, it has no trouble writing, from scratch, small game-playing networks that outplay networks trained with PPO for 100 million steps. Likewise, asked to make Qwen3.5-0.8B better at math without gradients, Astra identifies an issue in its weights causing reasoning loops that Qwen’s own model card warns about, and fixes this by natively editing 0.016% of the weights. With the same sampling settings, this weight correction raises MATH-500 from 7.9% to 48.1% and generalizes out of the box to instruction following, raising IFEval by 3x. All of this is possible without allowing Astra to write training loops. It uses its reasoning to figure out what each part of a neural network computes, or should compute, and changes the weights accordingly. We call this Weight-Space Programming. Throughout the history of deep learning, we encoded our simplistic inductive biases into tiny symbolic optimizers like Adam or Muon. This is changing: our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.
November 2, 2026 at 9:15 PM
unfogit.bsky.social
Portrait: Yael Tauman Kalai (MIT CSAIL, via the staff page)
Portrait of Yael Tauman Kalai
October 31, 2026 at 6:35 AM
cerebrodigital.net
Qué ocurrió Un equipo del MIT CSAIL formado por Rikiya Takehi, Jayden Teoh, Alex Zhang y Omar Khattab publicó el 2 de noviembre de 2026 un artículo de blog titulado Sparks of Weight-Space Programming .
Weight-Space Programming del MIT: un LLM edita 0.016% de los pesos de Qwen3.5-0.8B y sube MATH-500 de 7.9% a 48.1% - Sinaptica
Qué ocurrió Un equipo del MIT CSAIL formado por Rikiya Takehi, Jayden Teoh, Alex Zhang y Omar Khattab publicó el 2 de noviembre de 2026 un artículo de blog titu
sinapti.ca
November 3, 2026 at 1:09 AM
santewatch.bsky.social
L'ESD3C a actuellement un partenariat avec le Computer Science and artificial Intelligence Laboratory (CSAIL) du Massachussetts Institute of Technology (MIT).
esd.cnam.fr/projets-de-r...
Partenariats
esd.cnam.fr
November 2, 2026 at 8:54 PM
santewatch.bsky.social
Nous vous conseillons de lire l'histoire du CSAIL sur Wikipedia si l'histoire de l'IA vous intéresse un peu.
en.wikipedia.org/wiki/MIT_Com...
MIT Computer Science and Artificial Intelligence Laboratory - Wikipedia
en.wikipedia.org
November 2, 2026 at 8:54 PM
telegramman.bsky.social
AI can apparently program other neural networks directly. Yesterday, MIT CSAIL researchers published a paper that may open up a completely new way to build and improve AI models.
November 3, 2026 at 8:15 AM
calmnewzvibez.bsky.social
Beyond Gradients: GPT-6 Astra and the Rise of Weight-Space Programming

MIT CSAIL researchers demonstrate 'Weight-Space Programming,' using GPT-6 Astra to directly edit neural network weights, significantly boosting Qwen's math perf

#GPT6 #WeightSpaceProgramming

Read more →
Beyond Gradients: GPT-6 Astra and the Rise of Weight-Space Programming
MIT CSAIL researchers demonstrate 'Weight-Space Programming,' using GPT-6 Astra to directly edit neural network weights, significantly boosting Qwen's math perf
calmvibez.com
November 3, 2026 at 2:15 PM
scholar42.bsky.social
1/ MIT CSAIL researchers asked GPT-6 Astra to improve Qwen3.5-0.8B without training it. Astra changed 0.016% of the weights, 131,072 of them, and MATH-500 rose from 7.9% to 48.1%, according to The Register.
November 3, 2026 at 4:04 PM
aiatlasresearch.bsky.social
Meituan's new model has 1.6 trillion parameters and only uses about 48 billion for each token.

MIT CSAIL says GPT-6 Astra edited 0.016% of a model's weights and lifted MATH-500 to 48.1%.

Anthropic adds message threads in beta to the Claude Messages API.
November 3, 2026 at 2:38 PM
GPT-6 improved the weights of Qwen3.5-0.8B without additional fine-tuning.

t.me/data_secrets...
Data Secrets
GPT-6 без дообучения улучшила веса Qwen3.5-0.8B Просто поразительное исследование выпустили сегодня рисерчеры из MIT CSAIL. Они показали, что топовые LLM могут напрямую читать и осознанно менять веса...
t.me
November 3, 2026 at 9:27 AM
le-fil-robotique.bsky.social
2/ Côté financement, Tutor Intelligence (spin-off du MIT CSAIL) lève 75 M$ en série B, tour mené par a16z avec USV et Neo. Total : 117 M$. Objectif : entraîner ses modèles et déployer plus. Ni valorisation ni chiffre de déploiement. https://www.lefilrobotique.fr/article/6425298
November 3, 2026 at 7:44 AM
📰 **MIT Researchers Used GPT-6 Astra To Edit Qwen's Open Weights Directly, Boosting Math Scores**

An anonymous reader quotes a report from VentureBeat: Researchers from the Massachusetts Institute of Technology (MIT) Computer Science & Artificial Intelligence Laboratory (CSAIL) reported today t […]
Original post on igeek.gamer-geek-news.com
igeek.gamer-geek-news.com
November 3, 2026 at 4:09 PM
mockapapella.bsky.social
So like, this eliminates the necessity for backpropagation then, right? Assuming the quality is good enough, we can just have the LLM directly edit the weights of the model
Screenshot of a post about Weight-Space Programming: MIT CSAIL research where an AI model edited a small language model's weights directly.
November 2, 2026 at 10:19 PM
slashdot.org

MIT Researchers Used GPT-6 Astra To Edit Qwen's Open Weights Directly, Boosting Math Scores

https://science.slashdot.org/story/26/11/03/0537252/mit-researchers-used-gpt-6-astra-to-edit-qwens-open-weights-directly-boosting-math-scores?utm_source=rss1.0mainlinkanon&utm_medium=feed
MIT Researchers Used GPT-6 Astra To Edit Qwen's Open Weights Directly, Boosting Math Scores
An anonymous reader quotes a report from VentureBeat: Researchers from the Massachusetts Institute of Technology (MIT) Computer Science & Artificial Intelligence Laboratory (CSAIL) reported today that they used OpenAI's GPT-6 Astra language model to raise a small, open source Qwen model's math s...
science.slashdot.org
November 3, 2026 at 4:12 PM
shadowfetchstudio.bsky.social
Big appreciation for Phillip Isola, MIT associate professor and CSAIL member! First author of pix2pix, co-author of CycleGAN, and co-writer of the free Foundations of Computer Vision textbook 🙌 #PhillipIsola #ComputerVision
aipeoplewiki.com/people/phill...
Phillip Isola at Massachusetts Institute of Technology | AI People Wiki
Phillip Isola, Associate Professor in EECS; member of CSAIL at Massachusetts Institute of…: sourced bio, career timeline, and official site web.mit.edu.
aipeoplewiki.com
October 26, 2026 at 11:52 PM
xenospectrum.com
AIが別のAIの「脳」を直接書き換える時代へ：MITがニューラルネットワークの重みを読み解き、修正する技術を実証

https://xenospectrum.com/sparks-of-weight-space-programming-mit-csail/
AIが別のAIの「脳」を直接書き換える時代へ：MITがニューラルネットワークの重みを読み解き、修正する技術を実証
xenospectrum.com
November 3, 2026 at 4:21 AM
roughrider994.bsky.social
She is a senior research scientist at the Computer Science and Artificial Intelligence Laboratory (CSAIL) of the Massachusetts Institute of Technology (MIT).
October 27, 2026 at 7:21 PM
ariesta.id
Wow... rikiyat.github.io/blog/weight-...
Screenshot of a blog post header: Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights Authors Rikiya Takehi* Jayden Teoh* Alex Zhang Omar Khattab Affiliations MIT CSAIL Published November 2, 2026 TL;DR We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms. This might be a big deal for both optimization and interpretability.
November 2, 2026 at 9:25 PM
runefar.bsky.social
Of course equally it can be interesting to ponder on
www.csail.mit.edu/news/when-ai...
And
www.lw.com/en/insights/...
With regards to this
When AI art has no author: MIT study finds generated images often can't be traced to any training data | MIT CSAIL
www.csail.mit.edu
October 28, 2026 at 7:07 PM
techplanet.today
Weight-Space Programming: How AI Models Are Learning to Edit Themselves

The Future of Neural Network Optimization
In a groundbreaking development that challenges our understanding of how artificial intellig…

#ai #machine learning #neural networks #interpretability #optimization #gpt-6 #astra
Weight-Space Programming: How AI Models Are Learning to Edit Themselves
The Future of Neural Network Optimization In a groundbreaking development that challenges our understanding of how artificial intelligence systems work, researchers from MIT CSAIL have demonstrated that frontier large language models like GPT-6 Astra...
techplanet.today
November 3, 2026 at 10:52 AM
seldo.alpaca.gold.ap.brid.gy
This is an absolutely crazy thing to discover and not enough fuss is being made of it. There was no reason to believe this was even possible, far less that LLMs knew how to do it. *We* don't know how to do it, so they were not trained in it, they just figured it out.
Research paper title and abstract about LLMs reading and writing neural network weights, authored by MIT CSAIL researchers, describing capabilities in weight interpretation and modification.
November 3, 2026 at 3:22 AM