AI models just got a lot less mysterious. A frontier model can now look at another model’s raw numbers and work out what that model does. The creator of this video, Matthew Berman, broke down a new research paper on “weight space programming,” and I was blown away by how far the researchers pushed it.
Quick refresher: weights are the millions or billions of numbers inside a neural network. They’re normally set by gradient descent, a fixed trial-and-error training process. Nobody fully understands why the final numbers work, which is why people call AI a black box.
🧠 The big idea
In weight space programming, a general-purpose LLM reads and writes those numbers directly, with no optimizer in between. You can ask it in plain English to find a failure inside a network, fix it, or write a new network from scratch. The star of the paper is OpenAI’s GPT-6 Astra.
🔬 What Astra pulled off
- It reads a model’s purpose from raw numbers. With only the weights and no list of options, Astra identified networks trained on Breakout, Pong and Pac-Man. It called the Freeway network “Frogger,” which is a very close guess for a road-crossing game. On Breakout, it reached 99% confidence using just 846 tokens. It also figured out which shapes three image models generate. And it told English, Japanese and Python models apart using only their middle layers.
- It fixes broken models with tiny edits. One trained model drove a simulated car that kept getting stuck on a ramp. In a single try with no tools, Astra changed 3.5% of the weights, and the car succeeded in all 1,024 episodes. The fix? Back up first to build speed. It also repaired Qwen 3.5, an 800 million parameter model that kept getting stuck repeating itself. With 50 forward passes (test runs of the model), its MATH-500 score jumped from just under 8% to just over 48%. GSM8K went from 13.8% to 67.1%.
- It writes working networks from scratch. With no training data, gradients or tools, Astra wrote networks for Breakout, Seaquest and Asterix. All three beat models trained traditionally for 100 million steps. Against 500 million steps of training, it scored 20 times higher in Breakout, came close in Asterix and reached 60% in Seaquest. It even wrote a 1.5 million parameter Go network, using code and a sandbox. With just 50 forward passes, that network beat an equal-size model trained on 20 million games!
⚠️ Where it still falls short
Astra couldn’t tell which of two models was sycophantic (too eager to agree with you) from the weights alone. It failed all four attempts. Once it could see where the models paid attention while reading text, it got all three pairs right.
The math gains also come with a catch. Astra stopped the Qwen loops, so the model could finish answers it already knew how to solve. Gemma 3, a bigger model that rarely loops, saw no improvement.
🚀 Why this matters
The author sees some big possibilities here, though none of these are proven yet:
- Better safety: understanding models down to specific weights could make them easier to align.
- Stronger small models: a tiny model on your phone could send its weights to a bigger model and get back a version tuned for your task.
- Self-improvement: if a model can edit another model’s weights, it might one day edit its own.
That last one is exciting and a little unsettling, if I’m honest. Watch the full video to see Astra’s reasoning, like how it spotted a paddle in a wall of numbers. It’s worth your time.