Top all-time · #5100
gradient science
Company blog · gradientscience.orgaiscience
gradient science is the company blog at gradientscience.org, writing about ai and science. Since 2019, 15 different Hacker News readers have linked it 18 times, most often “Adversarial Examples Are Not Bugs, They Are Features”. It is still writing: the last post was on 2025-03-06, and you can follow it by feed.
Too few sampled comments (4) to say how often HN recommends it.?
Why HN links it
Gradient science posts summarise research papers, such as the adversarial examples paper, and one commenter cites its work on what tiny models suggest about AGI. HN readers link it so others can skip the PDF.
Summarised by Claude from a sample of 4 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
Research highlights and perspectives on machine learning and optimization from MadryLab.
Most-linked posts?
- Adversarial Examples Are Not Bugs, They Are Features3 people · 4 mentions · first linked 2019-05-09
- PhotoGuard: Defending Against Diffusion-Based Image Manipulation2 people · 2 mentions · first linked 2022-11-04
What Hacker News says?
I appreciate you're trying to wind it down so I'll try to get to the point, but there's a lot to unpack here. I'm not evaluating these models on whether they are AGI, I am evaluating them on what they tell us about AGI in the future. They show that even tiny models, some 10000x to 1000000x times smaller than what I think are the comparable measures in the human brain, trained with incredibly simple single-pass metho…
Comment linking /adv/
on “Microsoft is investing $1B in OpenAI”Submitted Debuggable Deep Networks: Sparse Linear Models as a story · 6 points
HN discussionI enormously appreciate the resources OpenAI provides to start out in DRL such as this one. However, OpenAI has (purposely?) left out the brittleness of their algorithms to parameter choice and code-level optimizations [1] in the past. As a researcher myself, I would be more than surprised to hear that OpenAI did not explore this behaviour themselves. Instead, my guess would be that these "inconveniences" would do h…
Comment linking /policy_gradients_pt1/
on “Spinning Up in Deep RL”Submitted PhotoGuard: Defending Against Diffusion-Based Image Manipulation as a story · 3 points
HN discussionYes, because your validation set performance is imperfectly correlated with performance on more samples (because it's of finite size), so it will tend to overestimate within-distribution. There's also the problem that you probably want to deploy it on data which was not exactly collected the same way. So you have issues of both internal and external validity. In practice, at least for CNN classifiers on ImageNet etc…
Comment linking /data_rep_bias.pdf
on “Why neural networks struggle with the Game of Life”Submitted How Does Batch Normalization Help Optimization? as a story · 3 points
HN discussionThere's a good summary of the paper by the authors here, for people who don't want to digest the pdf: http://gradientscience.org/adv/
Comment linking /adv/
on “Adversarial Examples Are Not Bugs, They Are Features”Submitted Adversarial Examples Are Not Bugs, They Are Features (2019) as a story · 2 points
HN discussionSubmitted GSM8K-Platinum: Revealing Performance Gaps in Frontier LLMs as a story · 2 points
HN discussionSubmitted Are there certain types of failures that pre-training alone cannot help with? as a story · 2 points
HN discussionSubmitted Show HN: Model-Aware Dataset Selection with Datamodels (DsDm) as a story · 2 points
HN discussionSubmitted Predicting Predictions with Datamodels as a story · 2 points
HN discussionSubmitted Adversarial Examples Are Not Bugs, They Are Features as a story · 2 points
HN discussionSubmitted PhotoGuard: Defending Against Diffusion-Based Image Manipulation as a story · 1 points
HN discussionSubmitted Identifying Statistical Bias in Dataset Replication as a story · 1 points
HN discussionSubmitted Adversarial Examples Are Not Bugs, They Are Features as a story · 1 points
HN discussionSubmitted Representations by robust networks align much more closely with human perception as a story · 1 points
HN discussion
