Nvidia’s ‘Nemotron-4 340B’ model redefines synthetic data generation, rivals GPT-4
https:// venturebeat.com /ai/nvidias-nemotron-4-340b-model-redefines-synthetic-data-generation-rivals-gpt-4/ Nvidia’s ‘Nemotron-4 340B’ model redefines synthetic data generation, rivals GPT-4
Redefines synthetic data generation? From "making fake data" to what? (I get it, it's hyperbole.)
Training on synthetic data is the most garbage-sounding thing I've ever heard. Isn't Reddit and Wattpad fake enough? At least those posts were generated by actual humans.
I guess I'll wait and see, but I'm really skeptical that this will be a good thing.
2Its a promising idea. Limitless training data in any domain you want. Whether or not it actually pans out is a whole different story.
6
340B is fucking huge, holy shit. How big is GPT-4?
1The rumor is 1.76 trillion, or 8x220B (mixture of experts) to be specific: https://wandb.ai/byyoung3/ml-news/reports/AI-Expert-Speculates-on-GPT-4-Architecture---Vmlldzo0NzA0Nzg4
2