Source: Zinc Finance

Last week, in a show, OpenAI CEO Ultraman said, "In order not to let OpenAI be surpassed, there is pressure every day. Plan to be with DeepSeek's leadership Meeting with the layers. ”

I have to say that it is rare to see Ultraman Sam in such a low position. Even if he was jointly transformed by the company's directors and senior executives and was kicked out of the position of CEO, he was calm He calmly took out his phone to take a selfie, turned around and went to Microsoft to regain power.

Now, under the popularity of DeepSeek around the world, Ultraman Sam and OpenAI have undoubtedly suffered an unprecedented impact.

Since the first year of ChatGPT-4 launching AI, OpenAI has been regarded as the "absolute leader of the AI ​​world" and is the catch-up target of countless Internet giants and AI startups.

Now, the pursuers are in turn become chasers.

The Spring Festival has just passed, and the whole world has ushered in the "DeepSeek Moment", which has exceeded 100 million users in just 7 days.

For comparison, the time ranking of users in various Internet products around the world shows that ChatGPT takes 2 months and second only to DeepSeek; TikTok takes 9 months and ranks third; Pinduoduo takes time 10 months, WeChat took 1 year and 2 months, ranking fourth and fifth.

The number of users is just an appetizer for DeepSeek to surpass ChatGPT. What scares Sam Ultraman the most is that under the impact of free DeepSeek, no user is willing to choose paid ChatGPT.

When facing the company's biggest challenge, OpenAI also had several strong struggles to survive.

The first card played by Ultraman Sam was the price reduction.

Just when DeepSeek caused OpenAI's API subscription volume to plummet by 37% in two weeks, OpenAI, which has always been pampered, panicked and announced an urgent announcement of an API price reduction of 50% and opening up some source code.

But this did not alleviate OpenAI's constant avalanche. Even the Wall Street Journal came out to publicly mocked: "This is just a stress response before the collapse of the closed source empire."

In order to alleviate the decline, Ultraman Sam made a series of tricks. , played the second card: relax content review.

Last Tuesday, February 14, OpenAI launched a shocking bomb: the official unblocking of ChatGPT's "adult content" generation function.

In foreign countries, chatbots have been in a highly criticized situation. Especially in November 2024, there was a sensational "AI instigated suicide" incident in the United States. Google's big model Gemini directly sent a 20-year-old college student: "Human, please die, pleasePlease. ”

The outside world has been calling on AI companies to strengthen content review, but OpenAI has done the opposite and further lifted the ban on ChatGPT's "adult content" generation function.

This move Implementation has caused ChatGPT application downloads to surge by 280%, but the user's retention rate plummeted to 11% the next day.

The data will not lie, which shows that "Qizhao traffic" is just a flower in the mirror, and it is difficult to Retain real users.

All the tricks are declared invalid, OpenAI finally came to the intersection of fate:

Yes Continue to close the source, maintain the status quo, and blindly believe that your product strength is enough to match the free DeepSeek?

Or overturn all your business plans and turn to open source?

Already, Sam Ultraman and OpenAI have always been firm "AI conservatives": insisting on closed source, delaying open source, and pursuing profit maximization.

The "API+ subscription system" that ChatGPT has always implemented is also OpenAI's powerful commercial moat. The so-called "API+Subscription System" is very simple to understand, it is a paid membership.

20 At the end of January 23, ChatGPT launched the Plus subscription option of $20 per month; by the end of 2024, ChatGPT added a Pro subscription option of $200 per month.

Paid users can enjoy various Permissions, such as priority access, shorter queue time, higher concurrent use, etc.

After OpenAI established the "order of the AI ​​world", almost all AI companies and big model companies were " Touch OpenAI to cross the river", and use "API+Subscription System" as the unified standard in the AI ​​industry.

But the Internet is the most The classic business war model is to defeat payments with free. Especially in the Internet world, it is undoubtedly a very difficult task to want users who are used to free to pay.

According to Mistral, an open source AI developer abroad Disclosure that the vast majority of customers are using smaller free models, and less than 10% of customers use the largest paid models.

As DeepSeek progressed step by step, Ultraman Sam was on Reddit During the Q&A campaign, it was publicly admitted: "OpenAI is on the wrong side of history on the open source issue, and now it needs to come up with a different open source strategy. ”

The day before yesterday, on February 18, Ultraman Sam became a "ridream man" on a foreign social platform, suggesting two directions of OpenAI's open source plan :

First is a fairly small o3-mini-level model that still needs to be run on the GPU; secondly, a side-end model suitable for smartphones.

Sam Otter Man has undoubtedly turned on the switch for the full transformation of OpenAI into open source, in the past six yearsSince then, OpenAI has not opened sourced any large models since GPT-2.

I didn't expect that the task that Musk had not completed was easily solved by DeepSeek.

But, fortunately, although the difficulties popped out one by one, Ultraman Sam and OpenAI still have the last trump card to turn over.

That is: GPT-5.

Switch the time to two years ago. On March 14, 2023, OpenAI released GPT-4, opening the "first year of AI". That was a "ChatGPT moment" that was no less than the current "DeepSeek moment".

Two years have passed, and the call for GPT-5 is getting higher and higher, but Ultraman Sam only handed over a "semi-finished" GPT-4o on May 14, 2024, the new generation of flagship GPT- 5 Always in a difficult state.

Ultraman Sam doesn't want to launch the GPT-5. It's too difficult to make a GPT-5 that meets users' psychological expectations.

In mid-2023, OpenAI launched the first training codename "Arrakis", but the test results show that larger-scale AI big model training takes an extremely long time, which will lead to overall cost. Soaring.

From May to November 2024, OpenAI conducted several small-scale trainings and second large-scale trainings one after another, but the problem of too little data and insufficient data diversification still exists.

Ultraman Sam confessed: "The cost of training GPT-4 is about $100 million, and the cost of training AI models in the future will reach $1 billion. Now, the GPT-5 month-long training has been 500 million US dollars spent and failed to achieve ideal results. ”

After the rise of DeepSeek, OpenAI was forced to overturn the closed source logic and move towards full open source, and it is bound to be re-established Plans to design GPT-5.

The latest news at the moment is a hint from Ultraman Sam on social platforms, saying that the GPT-5 will be released "in a few months" and may be launched in late 2025.

I just don’t know if OpenAI can still withstand the bombardment of many open source models led by DeepSeek.