just coded up a model that gets even higher fraud detection rate than stripe's transformer (97->100%). AMA
@jxmnop
-
Vatican invests to overcome LLM data wall challenge
By
–
and of course since 2023 we've been dealing with a problem called the data wall, where a large portion of all text data known to humanity has already been indexed, cleaned, and trained into today's large language models. that's why the vatican is planning to invest in new data
-
Online Learning: The Missing Paradigm in Modern AI Systems
By
–
a paradigm of modern AI that feels very behind is that model weights are trained every few months, but frozen between and during interactions real intelligent systems (such as human brains) update their parameters online maybe we'll get there in a couple years
-
Physicists’ Hidden Impact on Modern AI Development
By
–
i was surprised to learn how much of modern AI was developed by physicists: – scaling laws – alphafold
– RLHF (dario got his phd in biophysics)
– diffusion models
– normalizing flows
– VAEs
– neural tangent kernel
– hopfield networks
– information bottleneck theory
– neural -
AI-Powered Astrology Using Personal Ad Data and Predictions
By
–
i always thought someone should buy all of our ad data and then use it to make 'real' astrological predictions using knowledge of us + AI
-
ML Course Critique: Theory-Heavy Approach and FastAI Framework Bias
By
–
why? I completely disagree. The ML course is a little tangential and too theory-heavy. and FastAI course is too much of an ad for the FastAI framework.
-
Balancing Theory and Practice: 50/50 Split for Learning
By
–
he's right though, doing exercises is good! maybe 50/50 is a good split. the books have good exercises within them as well, but don't require you to write any code iirc
-
Textbooks Have Zero Hallucination Rate Unlike AI
By
–
you're not wrong. the only difference is that textbooks have a 0% hallucination rate
-
Mixture of Experts: Historical Context and Evolution
By
–
oh cool i didn't know about this! apparently MoEs are from the 90s. they're still not in the textbook. i had thought the first real implementation was from Shazeer et al. 2017:
-
Deep Learning Textbooks Outdated: Need New Comprehensive Resources
By
–
crazy to me that the one seminal textbook on Deep Learning was written in 2015. ten years ago. before we invented transformers, pretraining, RLHF, reasoning, diffusion, MOEs… i think AI researchers these days may be too busy to write textbooks. but we need a new one, badly