Each word you say is planned by a small patch of brain above your left ear, about a quarter of a second before you say it. It sits in a temporal lobe region textbooks describe as only for listening. Our paper today in @ScienceMagazine: https://t.co/EXOgl3YsNI
@UCSF@NeurosurgUCSF
AI coding agents can waste time and tokens relearning the same lessons. TokenEconomizer builds on past sessions to cut costs while maintaining task success. It’s a path toward coding agents that learn not only how to solve tasks, but also how to spend computation more effectively.
The open-source plugin is available now on @github. https://t.co/52jMrEw11J
Super interesting work from Microsoft and colleagues.
They take a closer look at the exciting research area of weight-space programming.
They had Claude Opus 5.5 write transformer weights directly in code, with no gradient training.
The smallest models that solved the algorithmic tasks had a median of 166 parameters.
The method is called interpretable weight-space programming. An agent proposes a computation, writes it as explicit weights and features, tests it, and revises it for 30 to 50 rounds. A person can read the result line by line.
All 26 runs on algorithmic tasks reached 100% accuracy on held-out data, often by writing a known algorithm into the weights.
They find that the model matters. Sonnet 5.5 needed a median of 41,600 parameters, and Haiku 5.5 needed 100,000.
Language tasks were harder. Those models stayed near 190,000 parameters and lost accuracy against learned baselines.
They then used the same loop on brain data. In the visual cortex, it predicted responses better than existing interpretable models. In language fMRI, it found a confound in the dataset.
https://t.co/JKYdkdLqwV
Chat with Paper: https://t.co/N58d4lTGNC
Our weight-space programming paper! The exciting thing here is that agents can iterate on mechanistic hypotheses much like humans do (but way faster). This yields models that are both highly accurate & interpretable, esp. in our brain encoding experiments
Excited to share what we've been working on since this summer- started in my internship at @MSFTResearch!
Modern DNNs predict neural activity surprisingly well, but are black boxes. We introduce interpretable weight-space programming (iWSP) and ask: can LLM agents build predictive models of the brain directly in code, with no training or pretrained weights, so every feature stays transparent? 🧵 (1/6)
@AIQuanting The interpretable models here are still worse than black-box models, but seem more accurate than alternative interpretable models (mostly because of exploring many more relevant hypotheses)
Excited to share what we've been working on since this summer- started in my internship at @MSFTResearch!
Modern DNNs predict neural activity surprisingly well, but are black boxes. We introduce interpretable weight-space programming (iWSP) and ask: can LLM agents build predictive models of the brain directly in code, with no training or pretrained weights, so every feature stays transparent? 🧵 (1/6)
(1/4) 4-digit multiplication is hard for transformers to learn. @yuntiandeng@YejinChoinka et al cracked it by phasing out CoT. @Elenal3ai@PresItamar et al hit ~100% using a running-sum loss.
An LLM just wrote it in 355 weights (40,000x fewer than SOTA) without training.
@DimitrisPapail Even worse for frontier labs, this means a lot of their formerly high-quality web data sources are now polluted by older versions of their models.
For anyone questioning whether this post is overhyped, I think they are actually *underselling* the approach.
We did similar experiments and found that Claude models can write entire, functional neural net weights from scratch! https://t.co/8kLTXHPNY6
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq
@wundram Maybe we can represent a lot of mechanisms as code, but I don't think we know how to translate mechanisms for hard tasks (like MATH-500) into anything effective besides neural networks yet