InTowards AIbyHarish K·6d agoWhy Mitra-v2 Treats Regression as a 1,000-Class Classification ProblemI wanted to understand why a regression model would predict 1,000 probabilities instead of a single number, so I tested the idea with small…A response icon1A response icon1
InAI AdvancesbyHarish K·Nov 7I Tested MiniCPM-OCR-Jev to See If AI Can Catch Its Own OCR MistakesI tested a 2B-parameter OCR verifier on 432 document regions and manually checked 864 outputs to see whether it could recognize mistakes…A response icon2A response icon2
InThe Quantastic JournalbyHarish K·Nov 2I Rearranged the Columns of a Dataset. Xiaomi-TabLDM Changed Its PredictionsI tested Xiaomi-TabLDM across four datasets, three train/test splits, and nearly 100,000 predictions to see whether changing column order…A response icon1A response icon1
InArtificial Intelligence in Plain EnglishbyHarish K·Oct 7The 1948 Paper That Predicted ChatGPT Would Be Boring Without RandomnessA Bell Labs mathematician generated nonsense on purpose, and it may reveal the hidden dial behind AI “creativity.”A response icon1A response icon1
InThe Quantastic JournalbyHarish K·Oct 6The Matrix Is Mostly Empty: An Intuitive (and Hands-On) Guide to Matrix FactorizationImagine a spreadsheet where the rows are users and the columns are movies. Each cell contains a rating from 1 to 5, but most of the cells…A response icon2A response icon2
InAI AdvancesbyHarish K·Sep 15I Tested Google’s TabFM Against XGBoost and AutoGluonGoogle’s TabFM almost removes task-specific training from the workflow. The accuracy surprised me, but the inference cost was even more…A response icon6A response icon6
InAI AdvancesbyHarish K·Aug 24The Empty Sphere: How a Memorial to a Dead Mathematician Became the Backbone of Modern SimulationIn 1934, a Russian mathematician published a paper with an unusually personal title for a work of pure geometry. It described a shape that…A response icon1A response icon1
InAI AdvancesbyHarish K·Aug 20The Model That Failed Without BreakingHow River lets machine-learning systems keep learning after deployment
InTowards AIbyHarish K·Aug 11I Tested 10 AI Models for OCR Across 20 LanguagesA hands-on comparison of 10 AI models across 20 languages, covering OCR accuracy, complex documents, speed, and unexpected failure modes.A response icon3A response icon3
InArtificial Intelligence in Plain EnglishbyHarish K·Aug 10I Gave Frontier AI Agents a Real Ransomware and Cybercrime InvestigationEvery time a certain model gets released, I observe the same metrics. 91.4 on one benchmark, 89.7 on another, better at coding but only…A response icon1A response icon1