Source: Alphabet List

OpenAI’s 12-day marathon live broadcast event is over, but this “marathon” is a bit unworthy of its name.

The most anticipated GPT-5 is still absent, and the belated arrival of Sora lacks surprises. The next-generation inference model o3 has not yet come out of the pan and will have to wait until January next year.

At the same time, pressure from the outside came one after another: Google took advantage of the OpenAI marathon event to raise its sniper rifle and fire bullets that hit accurately one after another. Not long after the OpenAI event ended, Musk's xAI announced the completion of a $6 billion Series C financing. Together with the Series B financing in May, the company has raised $12 billion.

This was originally a marketing opportunity for OpenAI to end 2024 in a relaxed and cheerful atmosphere, but it mirrored the external attacks and perfectly outlined OpenAI's experience this year.

In this year, OpenAI is still excellent, but it may have retreated from the altar to the world. In 2025, there are bound to be more challenges.

At the end of the year, before Christmas, we will conduct live broadcasts for 12 consecutive days, and a series of new products and functions are coming!

The above is the impression before the start of OpenAI’s live broadcast event at the end of the year. But the reality is: the 12-day live broadcast is not continuous and is closed on weekends; the live broadcast is not long, sometimes only ten minutes.

The most important thing is that in terms of content, OpenAI did announce a series of new features and products, especially the video generation tool Sora is finally online, and the launch of the next generation inference model o3 and o3 mini.

However, GPT-5, which is most anticipated by the outside world, is still absent.

Just after the 12-day live broadcast event officially ended, the Wall Street Journal broke the news that the development process of GPT-5 (codenamed Orion) was slow and did not achieve the expected results.

Specifically, this report stated that the Orion project has been in development for 18 months and has conducted at least two large-scale trainings, each requiring several months to process large amounts of data to make Orion smarter. . But relevant people said that new problems will arise every time during training.

Even based on Orion's optimal performance, although its performance is better than OpenAI's existing model, it is not worth it when combined with the cost - according to the estimates of the "Wall Street Journal", six Months of training alone can cost as much as $500 million.

Of course, this situation also made Microsoft, OpenAI’s largest funder, “the first to be disappointed.” Microsoft's original expectation was that Orion would see a new model in mid-2024.

Looking back at the end of November 2022, ChatGPT had just come out, and GPT-4 was launched in March of the following year, and it is still a crushing existence. Soon after, in mid-2023, the GPT-5 project was launchedDevelopment has already started, and since then, the outside world has been looking forward to the launch of GPT-5.

The greater the expectations, the greater the disappointment. The hindered launch of GPT-5 has also become a conspicuous symbol of OpenAI’s transformation in the eyes of the outside world.

While GPT-5 was delayed, OpenAI also made other efforts, such as launching the inference model o1, and then launched o3 in this live broadcast event, but it lacked the initial launch of GPT-4 The feeling of amazement.

In addition to the model changes behind ChatGPT, OpenAI’s Sora was officially announced at the beginning of the year and finally launched during the live broadcast event at the end of the year. One year is indeed a long time in the booming AI industry. "Friends" have already made arrangements, and Sora is no longer "condescending".

2024 has become a critical year for OpenAI’s image transformation. It cannot be said that OpenAI is no longer excellent, but at least it has fallen from the altar to the world. The "crushing" advantage seems to be no longer there.

"Friendly businessmen" are rising up, making it increasingly difficult for OpenAI to "stay ahead".

As far as this live broadcast event is concerned, OpenAI’s most popular o3 model and Sora have both been blocked. The most powerful competitor among them is Google, which was once controversial for its slow movement in the AI ​​wave.

On December 9, the third day of the live broadcast event, OpenAI finally announced the official launch of Sora, which can be enjoyed by ChatGPT Pro users who pay $200 per month.

Different from the high attention caused when Sora was officially announced at the beginning of the year, this time Sora was like a pebble dropped into a lake and did not cause much waves. The performance of up to 20 seconds and the highest image quality of 1080p has limited appeal. After all, the biggest competitor Runway, as well as Keling and Conch AI have been launched for a long time, and the pricing is not that expensive.

On December 17, Google came to “add insult to injury” again and suddenly launched Veo2. Veo2 can currently create videos of up to 8 seconds and 720p, and promises to reach more than 2 minutes and 4K resolution in the future. CEO Sundar Pichai specifically emphasized that Veo2 "has a better understanding of real-world physics and motion" when advertising on social media, which is a mockery of who is not telling the truth.

This time, the limelight was completely taken away. After Sora was launched, user discussions were not enthusiastic. Many reviews pointed out that it still suffered from the deformation of biological movements, unnatural hands, and garbled text that were demonstrated in the demonstration at the beginning of the year. Condition. Veo has attracted a lot of praise, especially for its coherent biological movement, natural and rich character expressions, and stable long panning shots.

Under pressure, as the live broadcast event came to an end, OpenAI announced that it would provide unlimited access to Sora to all subscribers during the Christmas holiday.

Google’s resistance is not limited to this.

On December 20, the penultimate day of the OpenAI “12-Day Marathon” live broadcast event, Google announced the launch of Gemini 2.0 Flash Thinking. This is a multi-modal reasoning model that is visualized through the thinking chain process and highlights the transparency of AI thinking and rapid problem solving.

Friends who are familiar with OpenAI products should easily see that Google’s new model is aimed at OpenAI’s inference model o1.

According to the preliminary evaluation results of the independent benchmarking website lmarena.ai, the overall score of Gemini 2.0 Flash Thinking surpassed the o1 preview version and ranked first in the overall score.

In addition, Google’s AI actions at the end of the year include: releasing a new generation of Vincent graph model Imagen 3, and integrating the previously released Astra project and Mariner project into Gemini2.0.

Since the launch of GPT-4, Google has continued to carry out internal reorganization to concentrate its AI firepower. In April last year, Google merged Google Brain and DeepMind into "Google DeepMind", with DeepMind co-founder Demis Hassabis in charge. In August this year, Google absorbed Noam Shazeer, the founder of Character.AI, and a “small group of colleagues” into Google DeepMind.

Today, Google has swept away the embarrassment of “not rushing” at the beginning of last year and has become one of OpenAI’s biggest competitors. According to data from the unified LLM API service platform OpenRouter, Google's share among platform developers has grown to 50%, while in September this year this number was still about 5%.

Google is just one of the rising "friends". The most representative one is Anthropic, which started as an OpenAI "rebel". It released Claude 3.5 this year, with the support of Amazon. xAI, founded by OpenAI's "enemy" Musk, released Grok-2 this year, and it was opened to all users instead of only X subscribers. It was also rumored that it would launch a separate application. And the giant Meta continued to solidify its "open source AI" strategy this year and continued to increase its supercomputing deployment.

Menlo Ventures conducted a survey of 600 IT decision-makers in U.S. companies and announced the results of the 2024 survey: Enterprise spending on generative artificial intelligence has soared by 500% this year. US$2.3 billion in 2023 to US$13.8 billion.

In the field of enterprise artificial intelligence, OpenAI’s market share dropped from 50% to 34%, Anthropic doubled from 12% to 24%, Meta’s share remained at 16%, and Google also dropped from 12% to 24%. A significant increase of 7%, reaching the 12% level.

In 2024, OpenAI’s competitors will become stronger and sharper, leaving OpenAI surrounded by layers of people.

Just talking about product competition does not give a full picture of OpenAI’s situation.

Competition in the AI ​​industry is not static. Underneath the weakness of OpenAI's products is the company's intense structural transformation and personnel turmoil.

Just during the OpenAI live event, another key figure left, namely Alec Radford.

He has been working at OpenAI for 8 years. His idea of ​​combining the Transformer architecture with massive data revolutionized OpenAI research and directly contributed to the success of the subsequent GPT model.

In fact, throughout 2024, OpenAI experienced personnel turmoil, with at least nine executives leaving. There are three points that best illustrate the "turbulence":

First, the 11 people in the OpenAI start-up team have dropped sharply to only 2 people today. Second, former chief scientist Ilya Sutskever and former chief technology officer Mira Murati left. Third, search director Shivakumar Venkataraman left. He previously led Google’s search advertising team and was hired by OpenAI with a high profile to lead the search segment just 7 months ago.

Originals have jumped ship one after another, and newly recruited key figures have also left at the speed of light, which shows the degree of turmoil in OpenAI. Except for a few entrepreneurs who left OpenAI, most of them were absorbed by "friends".

At the beginning of this wave of AI, the competition for talent has already begun. OpenAI has "feelings and ideals" on its left hand and high salaries on its right hand, making it a "holy place" for aspiring researchers. However, there is now a question mark as to whether the magic of OpenAI can continue.

An important next step for OpenAI is to become a truly for-profit company and get rid of the "control" of a non-profit board of directors. This matter was just a rumor last year, but this year it has been put on the table by OpenAI. To embrace profit, it is inevitable to give up some uniqueness, just as Musk has been attacking: OpenAI initially had the original intention of fighting against technology giants such as Google and creating AGI that benefits mankind. Founded in the form of an institution.

The process of reorganization is actually a re-adjustment of the company's priorities and primary goals, and this process will inevitably cause friction with the "elders".

High salaries may not last forever.

OpenAI itself does not have stable hematopoietic ability. This year, OpenAI completed US$6.6 billion in financing. According to financial documents disclosed by the company, it is not expected to become profitable until 2029, when revenue will reach 1000 billion US dollars. According to media analysis such as The Information, which has reviewed the documents, OpenAI's losses next year may reach US$14 billion, nearly three times this year's expected losses.

As for Microsoft, the largest “funder”, the relationship with OpenAI has become increasingly subtle. On December 24, local time, Reuters quoted people familiar with the matter as saying that Microsoft is working hard to get Copilot away from dependence on OpenAI. In addition to training its own small models, it is also actively customizing other third-party models. Some media simply summed it up as: Microsoft hopes to establish an "open relationship" with OpenAI.

On the other hand, OpenAI’s high-paying strategy has also been targeted by Musk and attacked.

In November, Musk’s lawsuit against OpenAI was upgraded. This time not only new defendants were added, but new evidence was also submitted. In a revised 107-page complaint, it wrote about OpenAI’s high salary and job grabbing behavior: “OpenAI is trying to create a shortage of AI talents from competitors through aggressive recruitment and high salary. Moreover, OpenAI plans to recruit 1,500 employees. It cost him $1.5 billion.”

The fact that Musk resumed the lawsuit against OpenAI is also worth pondering. The lawsuit first occurred in March this year and was subsequently withdrawn. Months later, Musk refiled the lawsuit in federal court. In mid-November this year, the scope of complaints further expanded.

At the same time, Musk himself and his xAI are making rapid progress. He himself became one of the biggest beneficiaries of the U.S. presidential election, becoming Trump's "confidant" and will be in charge of a new "efficiency department" after Trump takes office.

On December 24, xAI officially announced that it had completed a $6 billion Series C financing and announced its investment lineup, including Nvidia, AMD, Morgan Stanley, and Sequoia Capital. Together with the US$6 billion Series B financing in May this year, xAI’s total financing amount has exceeded US$12 billion. According to CNBC, xAI’s target valuation is US$50 billion.

If OpenAI CEO Altman could laugh it off last year and remain elegant in the face of Musk’s attacks, in 2024, this elegance is no longer there.

When raising funds this year, Altman tried to reach a closed agreement with investors, urging investors not to invest in OpenAI’s competitors.

A few days ago, Altman called Musk "obviously a bully" in an interview and said that Musk's high-profile dispute with OpenAI has become a "sideshow." It is rare for Altman to "attack" Musk so bluntly in public.

At the end of the year, taking advantage of Christmas, we held a live broadcast marathon. OpenAI might have expected to end 2024 in a relatively relaxed atmosphere and bring the outside world's attention back to the product. itself.

However, competitors are attacking and challenging Gaozhu on the road ahead. Such efforts do not seem to meet expectations. Instead, people saw the pressure on OpenAI and saw that a challenging 2025 is coming to OpenAI.