The new DEPSEK model ignited the smart source, R1, the sale of NVIDIA shares and caused the consumer application to rise to the highest application stores.
Last month, Dibsic said it trained a model using a data center of about 2000 H800 graphics processing units in NVIDIA in just about two months at a cost of about $ 5.5 million. Last week, she published a paper showing that the performance of the last model matches the most advanced thinking models in the world. These models are trained in data centers that spend billions of dollars on faster artificial intelligence chips and price in NVIDIA.
The reaction through the technology industry was on a high -performance Deepseek model, the lowest cost. Pat gelsiner, for example, Take to X with bliss, posting“Thank you Depsik.”
Of course, Geelsiner, former CEO of Intel, is a engineer and current president to start his subscription to public subscription, Gloo, a platform for churches. He left Intel in December four years later and an attempt to chase NVIDIA with GPUS AI, the Intel, Gaudi 3 AI.
Gelsiner wrote that Deepseek should mention the technology industry in the three most important lessons: low costs mean the broader adoption; Creativity flourishes under restrictions; And “wins open.” Dibsic wrote will help in increasingly reset the primary model work world. “Openai and the anthropology are both a closed source.
Gelsiner Techcrunch told R1 that R1 is very impressive, Gloo has already decided not to adopt and pay Openai. Gloo builds Amnesty International called Kallm, which will provide Chatbot and other services.
“My glue engineers run R1 today,” he said. “They can run O1 – well, they can only access O1, through application programming facades.”
Instead, within two weeks, Gloo expects Kallm rebuilding from scratch “with our own basic model that is all open.” “This is sexy.”
He said he believed Deepseek will make Amnesty International accessible to everyone, and will not be artificial intelligence everywhere. good Amnesty International will be everywhere. “I want Amnesty International Best in my OURA episode. I want the best Amnesty International in hearing. I want more artificial intelligence on my phone. I want better artificial intelligence in my compact devices, such as my voice recognition in my EV.”
Perhaps GELSINSER’s happy reaction was in contradiction with others who were happier than the logical constituent models now have a higher performance and more affordable. Artificial intelligence grows more expensive, not less.
Another reaction must be by clarifying a diplisk has imposed its numbers in one way or another and the training should be more expensive. Some believed that she could not say she is using high -end chips because of the restrictions of exporting AI to China. Others were fixing holes in their performance, as they found sites where other models were better. Others still believe that the next model for Openai, O3, will outperform the R1 when it is released that the current situation will be fixed.
Geelsiner ignores all of this. “You will not have complete transparency, given that most of the work has been done in China,” he said. “But still, all the evidence is that it is cheaper than 10 to 50x in their training from O1.”
Deepseek proves that artificial intelligence can be moved forward “through creativity engineering, not to throw more devices and resource account in the problem. This is exciting.
As for this is a Chinese developer with everything it means, such as concerns about privacy and control, gelsiner metapherapy shakes his head.
He said: “The presence of the Chinese reminds us of the power of open ecosystems, perhaps an embarrassing touch for our society, for the Western world.”
