DeepSeek’s AI Cost Claims Questioned by DeepMind CEO Demis Hassabis

DeepSeek, an artificial intelligence startup, have recently grabbed notice with their R1 model. They claim to have achieved GPT-4-level performance at a training cost of only $5.6 million, which is significantly less than OpenAI's claimed $100 million or more. However, Google DeepMind CEO Demis Hassabis remains skeptical, calling the claims “exaggerated and a little bit misleading.”  

Speaking at the Artificial Intelligence Action Summit in Paris, Hassabis said that DeepSeek’s cost estimate likely represents only the final training step, excluding critical expenses such as data collecting, infrastructure, and several training cycles. This allegation was echoed by OpenAI, which stated that companies based in the People's Republic of China frequently attempt to distill models from leading AI companies in the United States. He also hinted that DeepSeek may have utilized Western AI models for the purpose of refinement.  

Despite DeepSeek’s efficiency claims, Hassabis does not regard it as a huge advance. It is his contention that Google's Gemini models are, in fact, more cost-effective, despite the fact that they have not been marketed with as much success. He added, "It is impressive, but it is not some new outlier on the efficiency curve." That is an accurate statement.  

DeepSeek’s bold assertions have surely generated controversy in the AI field. The real test, however, will not be in marketing claims but rather in continuous performance and ground-breaking innovation. This is because the battle for powerful yet cost-efficient AI models is heating up.

This article is based on information from Business Today.