…why? you can download the model
@jxmnop
-
Recovering Model Weights Exactly Post-LoRA: New Discovery
By
–
just learned that under certain conditions you actually CAN recover model weights *exactly*, post-LORA pretty crazy
-
New Vulnerability Enables Pre-Fine-Tuning Weight Recovery Attack
By
–
— dr. jack morris (@jxmnop) 15 août 2025
We uncover a new vulnerability- Pre-Fine-Tuning Weight Recovery
With a few LoRA fine-tuned models we recover the pre-fine-tuning weightsof SoTA models, undoing Stable Diffusion personalization training and Mistral alignment
Project: http://
vision.huji.ac.il/spectral_detun
ing
… -
Working on Qwen: Extraction technique and model development
By
–
that's a good point; probably should have said 'we extracted one' not 'it'? working on qwen tonight
-
Clarifying Base Model Distribution Recovery in AI Experiments
By
–
some people are getting confused about the experiment – we didn't recover the base model's *weights*. that might not even be possible. we recovered the base model's *distribution*, with some error. an important question is how much. trying to figure that out right now…
-
Big Data has been significant for quite some time now
By
–
that’s funny. you’re right that the Data has been Big for some time now.
-
Interns Scaling AI Experiments With Unprecedented Computing Power
By
–
the world of AI is crazy right now because an intern will be like "sorry, let me double-check the numbers on that tonight" and then spin up an experiment with the power of 10,000 washing machines
-
PhD Success: Mastering Flow State Over Years
By
–
getting a phd is mostly about developing a system for maximizing the time you spend in flow state, over several years
-
Demonstrating Model Memorization: Testing Prompts
By
–
here I was arguing that I'm uncovering things that are present in the model that it won't generate on its own but point taken. what are some prompts that'd make a more compelling demonstration of memorization? I can run them.
-
GPT-OSS Pretraining Accuracy Concerns Discussed
By
–
yes––planning on it! although i worry about the accuracy of this kind of thing since the pretraining of GPT-OSS seems to have been pretty unique