the.com/ai training
teaching a machine to guess the next word by making it fail millions of times first.
means the process of feeding a model massive data and adjusting its internal numbers until its guesses stop being embarrassing.
from borrows the word from animal training, but here the treats are gradients: tiny corrections computed via calculus (backpropagation) that nudge billions of parameters closer to right answers.
costgpt-4 class runs reportedly exceed $100 million
data appetitetrained on trillions of words, still hungry
real bottleneckelectricity and gpus, not ideas
forgettingmodels can overwrite old skills learning new ones
for instance
gpt-4 — openai, 2023, trained on internet-scale text plus human feedback
alphago — deepmind, 2016, trained by playing itself millions of times
stable diffusion — 2022, trained on billions of image-text pairs scraped online
llama — meta, open weights, trained on curated public datasets