Can AI Write Another AI’s Weights?

0:00Social media is blowing up because

0:01someone claimed to find the holy

0:04grail of AI programming

0:06Researchers on Twitter claimed that frontier

0:09models could create smaller models by

0:11literally just typing out the weights

0:14and biases

0:15I decided to put this to

0:16the test and here's how In

0:19Experiment 1

0:20I took Qwen 3.5’s 0.8 billion

0:23parameters and used the researcher's code

0:25to reconstruct the weight edits GPT-6 Astra

0:29made

0:29In their study

0:31Qwen 3.5 went from 7.9% on

0:33a math benchmark to 48.1% So

0:37I decided to test on 4

0:39math problems and in my test

0:42Qwen went from getting 1 of

0:434 correct to getting 3 of

0:454 correct

0:46And the way it did it

0:47was Astra saw the activations that

0:50caused Qwen to repeat itself

0:53It then mentally edited those weights

0:55and biases so that Qwen would

0:58stop getting in a repeat loop

1:00This one small edit was enough

1:03to get Qwen to finally start

1:05answering with the correct answer

1:07In Experiment 2

1:09researchers gave Astra the source code

1:11of a tiny version of Breakout

1:13and a blank neural network

1:16Astra figured out how to predict

1:18where the ball would land and

1:20moved the paddle there It then

1:23typed out the weights and biases

1:25to make a neural network do

1:28exactly that without a single practice

1:31game

1:32The end result was the model

1:34created by GPT-6 Astra averaged 517 points

1:39And the model created the traditional

1:41expensive way the way it requires

1:44a data center and a GPU

1:46averaged just 25

1:48So that got me thinking

1:49what if I could use the

1:51same methodology to train an AI

1:54model for my specific task

1:56I've been wanting to fine tune

1:58Qwen 3.5 billion parameters to catch

2:02mistakes that my trading agent makes

2:05The longest example in my dataset

2:08was 262,000 tokens

2:11So in order to get Qwen

2:13fine tuned on this dataset of

2:151,000 questions

2:17I would need to rent an

2:18NVIDIA H200

2:20And just two epochs would have

2:23cost me anywhere between $65 and

2:26$100 in roughly 12 to 19

2:29hours of rental time So instead

2:31of renting an expensive GPU in

2:33a data center

2:34I decided to try GPT-6 Astra

2:37I had Codex inspect Qwen's mistakes

2:40in its activations and literally edit

2:43the weights and biases on my

2:45Mac

2:45With this approach

2:47Qwen 3 5 did no better

2:49than random changes The final version

2:52scored 5 of 9

2:55while the original version scored 6

2:58This goes to show that you

2:59can't blindly trust what you read

3:02on Twitter

3:03While weight-space programming is a

3:05promising future

3:07it's just not there yet for

3:08real world problems

3:10It's honestly not even close

3:13Always do your own research