this is a fair way of looking at things imo the software hasn’t changed much in a while
@jxmnop
-
Search limitations in unverifiable settings questioned
By
–
i’m not convinced search gives you much in unverifiable settings
-
Scale Maximalism: Why Bigger AI Models Drive Real Progress
By
–
i've basically come full circle and now think that the scale maximalism is the right way to think about AI until we scale the parameter count of our models, we will only see incremental improvement everything else is just cope while we wait for better GPUs & bigger datacenters
-
Property Inference Advances in Machine Learning Research
By
–
yeah, agree that that's a nice relaxation of this problem. this work focuses on property inference:
-
Discrete Text Optimization Challenges in Language Models
By
–
you're right that it's the discreteness of text that makes this problem extremely hard but also the problem you're imagining requires optimizing over optimization processes, which is also very hard what you're describing sort of works for images, e.g. this
-
Obfuscating AI Models While Maintaining Performance
By
–
yeah. people will start finding ways to obfuscate them without harming performance
-
New Research Paper Published on ArXiv
By
–
finally check out our paper on arxiv, out yesterday! 🙂 http://
arxiv.org/abs/2506.15553 -
Inferring Training Distributions from Large Language Models
By
–
if you find this area exciting, there is some fascinating concurrent work which examines the problem of inferring training *distributions* from LLMs: https://
x.com/iamgroot42/sta
tus/1934636691639472630
… (cc @chhaviyadav_ @kamalikac et al!) -
Extracting Hidden Information from Model Weights
By
–
im pretty convinced that there's significantly more information available in model weights, we just dont yet have tools to extract it in some sense that info is super "compressed" and we aren't sure how to build a proper decompressor our work is a tiny first step.
-
AdamW Weight Decay Obfuscation Effects on Neural Networks
By
–
a random detail that I found cool is that AdamW seems to "obfuscate" weights a bit better than Adam or vanilla SGD guessing this is bc weight decay does something nonlinear during optimization, which decreases the amount of information available in weights