AI for Prediction: Why You Need a Statistician Not a Philosopher
summary
Strong claim: real-world prediction needs statisticians not philosophers. Simple experiments ask an LLM to predict from 120 train / 30 test rows and show LLM accuracy is far below standard baselines (RF, GBT, AdaBoost), surprisingly matching nearest-neighbor-like behavior and collapsing as feature count grows. Fix is Large Tabular Models—architectures with cell embeddings, two‑dimensional attention and trained on billions of tables (TabPFN, TabiCOL, Tab FM, Mitra V2, Nexus)—which produce state‑of‑the‑art wins on tabular benchmarks.
AI-written summary · about
chapters (12)
- 0:00Prediction as survival
- 1:34Human vs tabular prediction
- 3:43Real-world prediction challenges
- 5:31LLM raw tabular performance
- 7:58LLM decision boundaries
- 10:05LLM performance vs features
- 11:47LLM prediction limits
- 12:29Large tabular models intro
- 13:10In-context learning shift
- 16:19Transformer tabular architectures
- 19:09Advancements in large tabular models
- 20:10Future of prediction
related talks
54mKeynote: Can you trust your (large language) model?
38mRole of Statistics in the era of Big Data
42mQuestions for Theory in the New Age of Machine Learning
24mHow to successfully become a Data Scientist
18mAI Is Not Magic: The Physics, Math, and Hardware Behind It
46mPlaying the long game
38mAI in action: How to use practically in iGaming beyond the hype?
49mGetting to True Predictive Risk: Will Data Accuracy Thwart AI’s Potential?