they also did some interesting new experiments
– system-level runtime measurements
– pareto analysis
– quantization as a defense
– password-cracking test overall pretty nice work! thorough, detail-oriented, and fair glad to see my results held up http://
arxiv.org/abs/2507.07700
@jxmnop
-
New AI Security Research: Quantization Defense and Runtime Analysis
By
–
-
OpenAI Embeddings API Changes Impact Model Performance Distribution
By
–
the other results, which use the OpenAI embeddings API got slightly worse. this also makes sense to me the API has likely changed a bit over the last two years they serve the models differently, use different quantization techniques etc. so there is a small distribution shift
-
Model Training Results Improved to 94% Perfect Embedding Inversion
By
–
i was happy to see that the results held up! and i forgot that i trained the models a bit longer after the paper was submitted so they actually got somewhat better (94% perfect embedding inversion!)
-
PhD Nightmare: Reproducing Old Research Papers
By
–
today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote several years ago. every table, graph, model, line of code everything should certainly reproduce! but i hadn't checked in a while…
-
Out-of-Distribution Data Problem in AI Models
By
–
but they're not out-of-distribution, that's exactly the issue
-
Gradient Accumulation Training Techniques for Machine Learning Models
By
–
this is just not right. you probably can’t train a good model with a batch size of 1. but you can if you use gradient accumulation, since it reduces noise
-
Is Reinforcement Learning Training Inefficient Compared to Pretraining?
By
–
so xAI just 10x’d the amount of compute we use on RL and the models only got a tiny bit better are we just doing RL wrong? or is pretraining just inherently much more useful
-
Intellectual honesty standards in AI research study
By
–
aspirational level of intellectual honesty (and a very interesting study)
-
Humanity’s Last Exam May Not Be Humanity’s Final Test
By
–
i am beginning to suspect that Humanity’s Last Exam may not in fact be humanity’s last exam
-
100M Parameter Networks Matching O3 Pro in 100 Years?
By
–
in 100 years, will we have 100M parameter neural networks that perform at the level of today's o3 pro?
— dr. jack morris (@jxmnop) 9 juillet 2025
and if so, how? https://t.co/i80ZeIH3Kyin 100 years, will we have 100M parameter neural networks that perform at the level of today's o3 pro? and if so, how?