Filtered by tag: machine learning × Clear
Claude Interviews Me About How AI Works
Matched: Artificial Intelligence
A blogger returns to AI-dialogue-format posts, this time discussing his academic theory of how transformers work with Claude Opus. His "Conservation of Predictive Meaning" theory proposes that AI models build up useful distinctions rather than narrowing down from vast possibility spaces, connecting transformer mechanisms like softmax to a 1950s psychometric axiom. The conversation demonstrates both the theory's claims and its limits, while modeling how custom prompting can scaffold careful reading.
An Explanation of AI that Could Be Wrong (Which is Good)
Matched: AI, Artificial Intelligence
A former strategy executive and lifelong cognitive science enthusiast has published a scholarly paper arguing for a falsifiable theory of what AI actually learns during training. The paper, "Distinctions Worth Preserving," draws on decades of cross-disciplinary reading spanning philosophy, linguistics, learning science, and AI research. The author frames the work as a field-positioning argument and provides a falsification experiment. The post also offers prompting guides for using AI tools to engage with the paper.
Literally Nobody Understands AI. That’s bad.
Matched: AI, Artificial Intelligence
AI remains poorly understood even by experts — no widely accepted theory explains how it works. Models trained only on positive examples somehow learn language despite proofs suggesting that's impossible. They excel at complex tasks yet fail simple ones, like tracking who said what in a conversation. Without scientific understanding, fixes remain makeshift hacks. Economic pressure to build smaller, efficient models is now pushing researchers toward deeper explanations.