#model-training-costs

[ follow ]
#deepseek #manifold-constrained-hyper-connections #llm-scaling #reinforcement-learning-fine-tuning
Artificial intelligence
fromZDNET
1 week ago

How DeepSeek's new way to train advanced AI models could disrupt everything - again

Manifold-Constrained Hyper-Connections (mHCs) promise a low-cost method to scale large language models; DeepSeek delayed R2 due to performance and chip-access concerns.
Artificial intelligence
fromTheregister
3 months ago

DeepSeek didn't really train its flagship model for $294,000

DeepSeek's $294,000 figure reflects only reinforcement-learning fine-tuning compute, not end-to-end training, making true training costs roughly twenty times higher.
[ Load more ]