Post AcZGgmUDa33HdUwRF2 by weddige@gruene.social
 (DIR) More posts by weddige@gruene.social
 (DIR) Post #AcYylwL0a99ESkwb3Y by baldur@toot.cafe
       2023-12-07T10:10:51Z
       
       0 likes, 1 repeats
       
       “Is AI leading to a reproducibility crisis in science?”“AI” research itself has had a massive reproducibility crisis for years. But the more these tools spread into other fields, the more those fields will be affected by the inherent snake-oil nature of “AI” https://www.nature.com/articles/d41586-023-03817-6
       
 (DIR) Post #AcYylxUcHuQa2qRmfg by baldur@toot.cafe
       2023-12-07T10:15:49Z
       
       0 likes, 2 repeats
       
       And here’s the thing people keep ignoring when buying into the hype of the “AI” bubble:You can’t trust “AI” research.Very few of the studies genuinely show what they claim to show. Those that do usually have highly flawed methods. Those that have solid methods have contaminated test/training data. Many of them will tell independent researchers they can’t see the data or their code because “confidentiality”.Most of it isn’t reproducible at all.It’s an industry built on snake-oil.
       
 (DIR) Post #AcZGgmUDa33HdUwRF2 by weddige@gruene.social
       2023-12-07T10:53:51Z
       
       1 likes, 0 repeats
       
       @baldur The problems with AI and data leakage remind me of a story my driving instructor told me about 20 years ago. He had a pupil who couldn't read but was determined to get his driving licence. So the pupil studied the exam and memorised, for every possible question, what the picture looked like and where the answers should be marked.The pupil aced the test without knowing what any of the questions meant.
       
 (DIR) Post #AcZGgnvYDuVg6YjmU4 by polarisera@spinster.xyz
       2023-12-07T13:44:26.913320Z
       
       0 likes, 0 repeats
       
       @weddige @baldur Overfitting combatted by regularization. We have the tools for dealing with open data, open systems with known predictable and analyzable algorithms. We also know there are hard limits on what can be known and predicted based on any given data criteria. But that was bad for the tards of AI and the tards of the public. AI lets them whitewash and destroy decades of research into the limits and control of algorithms.