Yuechen Zhao
Post · November 2026

AI is hand-writing neural networks now

A frontier model wrote a network by hand in 8 minutes, and it beat one trained on 500 million moves.

Training a neural network normally takes mountains of data. MIT researchers had Astra write one by hand in 8 minutes instead. It scored 517 at Breakout -- while a classically trained network on 500 million moves scored a pitiful 25. 🤯

A day later, Legora's CTO Jacob Lauritzen got Fable to one-shot a 100K-parameter network that spots fake legal citations with 97.8% accuracy, again beating networks trained on up to 8,192 labeled examples. Plus, it used just 3,793 non-zero weights, about 1/28th of what the trained versions used!

Wait, what? These frontier models just thought about the problem really hard and hand-coded working neural networks from scratch?

HOW IS THAT POSSIBLE?

Turns out, it's not quite wizardry, but it's pretty darn close. A neural network is just a big pile of numbers that get multiplied and added together. Normally we let gradient descent nudge those numbers toward the outcome we want, and nobody really knows what any single one means. BUT if you know exactly what you want each neuron to do, you can technically set the numbers yourself. For Breakout, Astra basically wrote a tiny program -- layer 1 figures out where the ball will land, layer 2 moves the paddle there -- and translated it into weights. It's using the network as a (very weird) programming language.

Simple enough in theory. Actually pulling it off -- in one shot -- is a whole other story.

What's even more wild is that writing networks was only half the paper. The MIT team also handed Astra the raw weights of Qwen, an 800-million-parameter open-source model with a known bug: on hard math problems, it gets stuck in loops and never gives an answer. Finding that bug by reading the weights is like decompiling assembly code -- technically possible, but so painstaking that no sane human would attempt it. Astra traced the looping to one pattern inside two of Qwen's layers, edited just 0.016% of the weights (no retraining), and Qwen's MATH-500 score jumped from 7.9% to 48.1%!

The implications of these findings are potentially massive.

Instead of just scaling compute to train the next big model, should we unleash today's models on the weights themselves to spot patterns and make targeted improvements directly?

Instead of "distilling" big models into smaller ones, should we be asking big models to hand-write small models built for a single task, at a fraction of the compute?

Instead of continuing to complain that AI is a "black box" we don't understand, should we just have models look inside each other and explain what's going on?

But perhaps the biggest takeaway of all: current-generation models are so capable that it can be hard to imagine what they might achieve next -- and here once again, they've done something so wildly complex that no human brain could realistically hope to match it.

Thrilling. And more than a little humbling.

I write about what I’m seeing in tech and AI. Follow along for the next one.

← All writing