Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Claude Fable 5.2 tests direct weight writing on MNIST

AI summary

In a single-author experiment, Claude Fable 5.2 generated code that set all 5,994 weights of a fixed MNIST convolutional neural network from scratch under rules that prohibited fitting, regression and searching, while allowing filters and templates to be drawn manually. The setup also excluded PyTorch and practice games, providing only training loss, accuracy and a confusion matrix after each attempt. Claude had up to 50 steps, during which it refined hand-designed features, measured feature frequencies in the training set and rebuilt the final layer as a naive Bayes-style classifier. The network reached about 90% accuracy. The author interprets the result as a chaining of historical, hand-designed model-building techniques rather than a fundamentally new training method, and compares it with GPT-6 Astra’s Go network.

In-depth insight

The experiment tests whether an LLM can directly write and iteratively refine neural-network weights. It also shows that the reported result depended on hand-crafted features and measurements from training data, rather than on a conventional automated optimization procedure.

The writing aspect of WSP turns out to be quite a history lesson.

Image 1

Intro

Weight-Space Programming still has me pretty intrigued. The basic idea is that frontier LLMs can read, write and edit the weights of neural networks. There's nuance to what's going on, but it feels like there's a lot of new ground to explore.

In the original blog post, beyond the reading and editing experiments, the authors had GPT-6 Astra build a network to play Go from scratch, then improve it by playing practice games, and editing the weights based on analysis of its losses. That loop of writing, testing and editing feels like a new way to train a network, something like LLM Descent. But is it?

In this article, I run my own weight-writing experiment with Claude Fable 5.2 on a familiar image classification problem. Look closely at what the model builds, and it turns out to be an interesting lesson in historical model-building techniques. With that new lens on weight-space programming, I also study the Go network that Astra built in the original article, and find something very similar.

#Claude Fable 5.2#GPT-6 Astra#Weight-Space Programming#MNIST

Original · 16 min readRead the original
Group chat

No comments yet. Start the conversation.