The reason why all AI is crazy about Deepseek

As of a few days ago, I heard only nerd nerds (I call this one). DeepseekSimilarly called, Chinese AI subsidiary Luxury capital managementQuantum analysis (or quantum) company that was first released in 2015.

However, within the last few days, it is definitely the most discussed company in Silicon Valley. This is mainly due to the release of Deepseek R1, a new large language model that performs the same “reasoning” as OPENAI’s current best model O1. Step -level, or “chain of thinking” fashion.

Not only that, Deepseek R1 has a higher or higher score than O1 of Openai in various third -party benchmarks (tests to measure AI performance in answers to questions on various subjects), and are transmitted. According to Trained at a slight cost (about $ 5 million according to what is conveyed) Under the strict inoperative measures imposed by the United States, the lawn of Openai, there are much less graphic processing units (GPUs).

However, unlike the O1, which can only be used for PLUS TIER ($ 20 per month) Chatgpt subclan and more expensive tier ($ 200 per month), DeepSeek R1 was released as a completely open source model. The AI ​​code chart that shares the AI ​​code has been rapidly raised. Hugging the most downloaded active model of the community。

Also, thanks to the fact that it is completely open source, many models have already been running on a mobile device or making it so small enough to combine them. Other open source models are trained by fine -tuning multiple variations. Even if you want to use it for development purposes, DeepSeek’s API cost is more than 90 % cheaper than the O1 model equivalent to Openai.

The most impressive thing is that you don’t even need to be a software engineer to use: DeepSeek Free website and Mobile app for users It is very similar to Openai’s Chatgpt using ChatBot interface equipped with R1. Again, by connecting this powerful inference model to the web search, Openai has not yet been executed by DeepSeek’s undercut or “mogging” Openai (Web search is currently not powerful. Only available in models).

Ironically and closed

In order to democratize AI to the masses, it may be quite delicious or sarcastic as the opening goal of opening. As NVIDIA Senior Research Manager Jim fans put it on X“We live on the timeline where non -US companies take advantage of Openai’s original mission. It is a truly open Frontier research to strengthen everything. That is meaningless. The most interesting results are possible. High sex.

Or as X user @suspedededrobot PutRefer to it DeepSeek seems to have been trained About questions and other data generated by Chatgpt: “Open is stolen and rich from the Internet, and Deep Shek is stolen from them and returns to the masses for free.

However, it is not only fans who sit down and pay attention to the success of Deepseek. The availability of the open source of Deepseek R1, its high performance, and the fact that it came out from nowhere to challenge the former leader of the generated AI is based on my conversation and reading. I sent a shock wave. Various engineers, thinkers, leaders. As my double curve heading suggests, if “everyone” is not crazy about it, it’s certainly about the town of technology and business circles.

a Message posted blindlyThe app for sharing anonymous gossip with Silicon Valley is the success of the meta, as the meta has exceeded how fast the meta is to become an open source AI king using a lama model. It suggests that it is in a crisis.

“This changes the whole game”

X user @tphuang wrote with persuasive power: “Deepseek has commoded AI outside the top end. In the first photo, the moment of a light bulb. R1 is much cheaper than US labor costs. It will be automated in the next 5 years. Please be careful later Why DeepSeek’s R1 is more attractive to users than O1 of Openai

“There are three huge problems: O1:
1) Too late
2) Too expensive
3) End user/lack of control for trust in Openai
R1 solves all of them. The company can buy its own NVIDIA GPU and run these models. You don’t have to worry about additional costs or slow/unresponsive Openai servers. “

@tphaung also posed Convincing as a question: “Is DeepSeek LLM how Android has become the OS world?”

Web entrepreneur ARNAUD BERTRAND did not engrave the surprising meaning of DeepSeek’s success. write in X: “We do not exaggerate how deep the whole game will change, and not only with AI, but also a large -scale accusation of the United States misplaced attempts to stop China’s technology development. Without it, DeepSeek may have been possible (as you say, the need is the mother of the invention).

Censor problem

However, others have heard about the rapid rise of Deepseek, and claims that it is inevitably subject to the censorship requirements of the country’s law and content because it operates a startup from China. Masu.

Certainly, it turns out that the use of my own DeepSeek in the US iOS app is so. Do not answer questions about Tiananmen Square( As a result, at least 200, perhaps thousands of deaths.Get nicknameTiananmen Square Massacre]In Western media.

Ben Hylak, a former Apple Human Interface Designer and a co -founder of the AI ​​Product Analytics platform, was posted on X. When asked about this subject, Deepseek R1 was able to enter the circuit loop.。

As a member of the media itself, I am very serious about the freedom of speech and expression, and that is definitely one of the most basic and untouched reasons for the champion.

However, note that Openai models and products, including Chatgptt, have refused to answer any questions in any range of harmless content, especially related to human sexuality and the theme of NSFW in adults. Please do not.

Of course, it is not a comparison between apple and apple. And some people may be skeptical of the resistance to the ultimate value and usefulness of Deepseek. However, it does not deny its performance and low cost.

And in the times 16.5 % of all US products are imported by ChinaIt is difficult to be careful about using DeepSeek R1 based on censorship concerns and security risks. In particular, when model code can be freely downloaded, offline, using on -devices in a safe environment, and can be freely adjusted freely.

I definitely detect an existing crisis about “Fall drop in the western part” and “China’s rise” and motivate me to discuss animation about deep -shak, and others have already connected it to H. I amOW US user participated in the app Xiaohongshu (also known as “Little Red Book”) When Tactoku is temporarily banned in this country, you will only be surprised at the quality of the life of China in the video shared there. The arrival of Deepseek R1 occurs in the context of this story. This rises (and clearly by many metrics) that China appears, but the United States has declined (and by many metrics).

There are few Chinese AI models that shake the world

It is not the Chinese AI model that threatens the superiority of the Silicon Valley giant. They, like Openai, are gathering as much money for the ambitions of developing artificial general information (AGI). Economic valuable work.

Just yesterday, another Chinese model by Tactoku’s parent company– DOUBAO-1.5-PRO -It was released with a performance that matches Openai’s non-rational GPT-4O model at the third party benchmark Again, 1/50 cost。

Chinese models are very fast, and even models outside the technology industry are careful. economist The magazine performed a work on DeepSeek’s success And other Chinese AI efforts and political commentators Matt BRUENIG posted to X“The summary of the NLRB document uses the outline of Gemini, Chatgpt, and Claude for almost a year. Deepseek is better than everyone. The chat bot version is free. The price of the API is the price. , Openai’s API price is 99.5 % below. [shrug emoji]”

How does Openai react?

Sam Altman of Openai’s co -founder and CEO The company said today O3, the second inference model family that has not been released yet, was introduced to ChatGpt for free users. Openai seems to open up its own path with a more unique model. This sets the industry standard.

But the problem is as follows: How much can Openai will continue to lead by creating and releasing a new state -of -the -art AI model because deep sheps, baitdance, and other Chinese AI companies are grabbing their heels? And when it falls, how intense and how fast it will decline?

However, Openai has another historical precedent, which is there. Just as Google’s open source Android went to Mobile, if DeepSeek and China’s AI models actually became LLMS -For a while, the share of the market lion was locked down, the Apple iPhone was locked down. You need to see how it looks and how it looks. The house approach managed to open up the high end of the market From there it will steadily expand downwardEspecially in the United States, it currently owns almost 60 % of the domestic smartphone market.

Nevertheless, for everyone who spends a lot of money to use the main lab AI model, DeepSeek has the potential to have the same functions much cheaper and have much better control. 。 Enterprise settings may be enough to get a ball game.