Why more parameters alone aren’t enough to make AI smarter, according to the Chinchilla study?

More parameters ≠ automatically smarter AI Many models like GPT-3 (175B) or Gopher (280B) have huge parameter counts but too little training data They are "undertrained" like a student with a big brain but only one textbook Compute optimal training Model size (parameters) and data volume must grow proportionally Larger models only perform better if they see proportionally more data otherwise, computing power is wasted If I double the parameters, do I also need twice as much training data? Data volume is just as critical and even saves computational resources? The study directly refutes the assumption that parameters alone determine AI intelligence. Why is this revolutionary and is this still up to date? More sources: Training Compute-Optimal Large Language Models Chinchilla (language model) Neural scaling law Chinchilla Paper explained What is the Chinchilla Point? Revised Chinchilla scaling laws – LLM compute and token requirements Scaling Laws for LLM Pretraining Reconciling Kaplan and Chinchilla Scaling Laws Scaling Laws and Emergent Abilities in LLMs Scaling Laws for LLMs: From GPT-3 to o3