Can AI Improve Qwen by Editing Its Weights?

0:00On X an extremely viral

0:02tweet claimed an AI model

0:04could create another AI model

0:06by literally typing out the

0:09weights and biases and showing

0:11the end result

0:13outperforming the model created using

0:15our traditional extremely expensive training

0:19that requires all these data

0:21centers. If you know anything about

0:23AI you understand how insane

0:25this claim actually is The

0:29MIT team behind the original

0:31study had GPT 6 Astra edit

0:34the weights of Qwen 3.5

0:37an open source Chinese model

0:40The score on the math

0:43benchmark went from 8

0:45in the original Qwen model

0:47to 48%. And I wanted

0:51to see if this was

0:53a hoopla or if Codex

0:55could really train an AI

0:57model from scratch

0:59So here's exactly what I

1:01did I took the original

1:04base model from the paper

1:05and I first reproduced the

1:08results just to make sure

1:09I understood what was going

1:12on and then I tried

1:14it for my specific task

1:16I had Qwen try to

1:18grade the output of an

1:21AI agent basically saying did

1:23this trading agent do all

1:26of the correct things or

1:28did it make mistakes I

1:31used the same exact procedure

1:33that was presented in the

1:35paper and I tried to

1:37see did Qwen improve Did

1:40Qwen do a better job

1:42like the paper would suggest

1:45and I did this for

1:48several hours and what I

1:50found was that Qwen did

1:52no better than random changes

1:55to the weights and biases

1:57While it did very well

1:59in this one math benchmark

2:01the results were not reproduced

2:04by my specific experiment

2:08What this shows is while

2:09weight space programming is a

2:11promising future for frontier models

2:14it's not quite there yet

2:16at all

2:17You literally did better by

2:19randomly changing the weights

2:21And this is your reminder

2:23to not trust whatever you

2:25find on social media