How DeepSeek Silicon Valley changed the AI ​​view

Chinese AI Laboratory Depth After releasing the open versions of AI models competing with the best technology, meta, meta and Google, in 2025, the first silicone valley provoked the stranger.

Deepsel claims that they have set up their models in a highly efficient and quick-time establishment, and these models provide a part of the prosecution of American AI companies. There is development Not only technological giants However, the highest level of the US government was afraid of China’s leading in the AI ​​gun race.

“I would not be surprised if there are many lab war rooms at the moment,” he said.

DeepSeek’s rising marks the point of an infection for Silicon Valley’s AI view. AI ceos, founders, founders, researchers and investors have great effects for the American American AI policy of Deepseek models to TechCrunch. Moreover, these experts say that models serve as an indicator of the acceleration rate of AI progress.

“Of course (DeepSeek) was too hungry,” he said. “But it’s still very interesting and there are so many things we can buy.”

New ways of AI thought

One of the main innovations related to the creation of the R1 of DeepSEEK was “pure reinforcement learning”, workers and an incorrect approach, workers and wrongly wrong approach.

Catanforoosh compared to a child to keep a child to avoid touching a warm plate with Deepseek accidentally burning.

“(A child) can touch a hot plate, can be burned and quickly teach you to do it quickly,” said CatanForoosh through the text. “This is a clean reinforcement learning – the review (…) is to learn to learn from testing and error based on the DeepSEEK method, learn through a single experience of the model.”

DeepSeek seems to have more than other advanced AI models more than in learning more strengthening. Openai also used reinforcement learning methods To develop O1The company announced the R1 that Deepseek revealed several weeks. Openai Upcoming O3 model The company claims claimed by the claim that the company claims better performance using similar methods.

The study of strengthening, according to Katanforoosh, today represents one of the most promising ways to develop AI basic models. The term “Foundation Models” usually applies to AI models taught with massive amounts of information such as pictures and text from the Internet. Most likely, other AI laboratories, especially the success of Deepseek, will continue to increase the level of strengthening to develop AI models.

A few months ago AI companies found themselves struggling to increase the activity of basic models. However, as the success of reinforcement learning and others, the success of methods, subtle regulation and test test time shows that AI progress can be removed.

“R1 gave me more confidence in the pace of staying at a high level,” he said.

A turning point For AI policy

R1, R1, who can load and work in a number of AI criteria in a number of AI criteria. When we played for the first time, when he saw that Openai and open existing models, it continues rapidly when it sees “closed” models.

This can force the United States to an open or completely open source, even completely open source to compete with China. Andreessen Horowitz (A16z), Deepseek’s TechCrunch’in Techcrunch’in a general partner of Techcrunch, which proves how the regulatory core of the last two years, proved that Martin Casado Techcrunch.

“I think it’s not only in our technical ability (USA),” Casado said in an interview. “Very competitive solutions are from anywhere, but especially China. Instead of hindering the U.S. innovation, it does not allow China. In fact, it does not allow our companies to use open source and spread our technology very much.”

Casado seemed to refer to the former President Bayden Recently recurrent AI executive power and vetoed california bill sb 1047Both opposed A16Z. A16Z, the American innovations of both events confirmed the priority of the Prevention of “Strange” to the EU Doomsday scenarios. In a wider, silicone valley generally succeeded To lower the “AI Doom Movement” in 2024. The real concerns around AI, A16Z and others have repeatedly said: America’s loss of competitiveness to China.

This scenario looks more different in the light of Deepseek’s rise.

It has a lot of open EU’s largest players, including nothing, but also a16z, databalks, autral and black forest laboratories. The company of VC can play a role in the AI ​​that advises the Trump Administration. Old A16Z partner Sriram Krishnan is now the Chief Policy Advisor for Trump’s AI.

President Trump said he had DeepSEEK on Monday “cabbage“When praising the Chinese AI laboratory for American AI companies, the Chinese AI laboratory for its open. EU matches a very closely with the position of A16z.

“DeepSeek R1 is the moment of Sputnik,” said A16Z co-founder Marc Andreessen X in xReferring to the launch of the Soviet space orbiting space ship, decades ago forced the United States to investigate its space program.

Deepseek’s rising also seems to change the mind of the open AI Skeptics, such as Google CEO Eric Schmidt. Last year, Schmidt concerned about the spread of Western open AI models around the world. But an op-ed published on Tuesday Schmidt said Deepseek’s rise was “turning point” He called on the Global AI race and the American open AI.

Looking forward

It is important not to overcome Deepseek’s success.

For example, some analysts teach one of the frontline models of DeepSEEK, the DeepSeek V3 teaches only $ 5.6 million – about $ 5.6 million – using Nvidia GPUs about 2000. The Chinese AI Laboratory, after all and Deepseek did not spruce overnight It was reported to have something There is a reserve with more than 50,000 skilled NVidia Hopper GPU.

Deepseek models are also defective. According to a test The information-reliability organization provides 83% of the time asked about newsguard, R1 error answers or news about the answers. A separate test R1 is the result of 85% of Chinese tips, and maybe the result of this Censorship of the government prepared by the AI ​​models in the country.

Then there are allegations of IP theft. Open says there are evidence This DeepSEEK would be truthful, using AI models. (Offior, Openai is currently in court for a number of parties allegedly caused by copyright violations in teaching their models.)

Again, DeepSeek transferred the needle with more efficient models – and updated. Lambert noted that unlike O1, R1 shows users ‘thinking process’. When Lambert saw the internal processes of some users, they observed that they believed or believed in the models of AI when they saw their internal process.

Now we need to see how American politicians and AI laboratories will respond.