llm / ai agents / machine learning
The Weight Edit Made My AI Worse
I asked Codex to apply Astra’s weight-editing rule to a Liquid 3 model trained to grade my trading agent. Correct grades fell from 5 of 29 to 3 of 29. On eight selected cases, DRY decoding reduced looping answers from five to zero and increased correct grades from three to five. Each of the five correct DRY answers still had a citation or source reference that failed verification.
Transcript
0:00MIT researchers claimed that they
0:03could have an AI model
0:04create another AI model by
0:07literally
0:07writing out its internal weights
0:10and biases from scratch without
0:12the need of an expensive
0:14data center
0:15I tried the same exact
0:16procedure for my own use
0:18case and it didn't work
0:20at all it actually
0:21did far worse
0:23For context
0:24I was fine tuning Liquid
0:263
0:27a new hyper efficient
0:301 7 billion parameter
0:31language model
0:32The goal was to get
0:34Liquid 3 to grade my
0:36AI agent to see how
0:38well it actually accomplished
0:40the task
0:42With my latest iteration of
0:44Liquid 3 I noticed a
0:46pattern It kept repeating itself
0:48over and over again until
0:50it ran out of
0:52tokens
0:52In this MIT paper notice
0:55the exact same problem
0:57Qwen repeated itself over and
0:59over again and for that
1:01reason
1:02it couldn't give a correct
1:03answer
1:04And allegedly
1:06if I were to use
1:07Astra to edit the weights
1:10it could improve the
1:12accuracy from 7%
1:13to over 47%
1:17So I tried this exact
1:19same procedure
1:21I compared the activations and
1:23edited 64 columns in each
1:26of two layers
1:28I then tested my original
1:30unedited model with the model
1:32I edited with GPT 6
1:34Astra
1:35And the results were so
1:37bad I had to stop
1:40midway
1:41The edit took us from
1:435 correct grades to 3
1:46and even more answers ran
1:49out of
1:51tokens
1:51I then tried something entirely
1:53unique
1:54I simply just changed the
1:57hyperparameters
1:58I didn't have Astra try
2:00some crazy tricks on weights
2:02and
2:03biases
2:03I just changed some parameters
2:05to penalize repeated sequences
2:08And this one trick decreased
2:10my degenerative repetition significantly
2:13I went from having 5
2:15out of 8 looping in
2:17the greedy baseline to having
2:190
2:20None of them did this
2:23degenerative looping
2:24This directly increased my accuracy
2:27from 3 of 8 correct
2:29to 5 out of 8
2:31correct
2:32So in conclusion
2:34the AI hype you see
2:35on Twitter doesn't necessarily align
2:38with real life
2:39But real AI engineering real
2:42machine learning engineering actually has
2:46an
2:47impact
Join the conversation
Loading conversation…