Image source: generated by Unbounded AI

The OpenAI Spring Festival Gala series across the ocean continues. On the 9th day of the continuous press conference, OpenAI officially released the o1 model API.

A brief inventory of the special Christmas season features that have been released: the full-blooded version of the o1 model VSChatGPT Pro subscription plan, the Ace Pigeon that detonated the world at the beginning of the year - the video generation model Sora is fully open, the new Canvas writing function, ChatGPT is officially connected to the video calling and screen sharing functions of Apple Family Bucket and GPT-4o, and is a complex project project to reconstruct the AI ​​collaboration model...

Compared with previous OpenAI model updates, these releases It's completely different. This iconic company is transforming into a product company.

As the most important reference for Chinese companies in designing technical routes and financing narratives, when OpenAI transformed, the Chinese followers who followed OpenAI closely also quietly "replaced" ".

1. Wisdom spectrum, more OpenAI than OpenAI

Previously, Zhipu has always been the most high-profile follower. In the Zhipu Financing Roadshow PPT recently released by the media, where there is a technical roadmap, there is OpenAI.

As a Tsinghua-based company established in 2019, Zhipu has launched a wealth of products in many aspects involving large model links, including general GLM large models of various sizes, as well as general models based on The dialogue products and AI assistants that are coming are smart and clear, representing Code model CodeGeeX, Vincent graph model CogView and multi-modal dialogue model VisualGLM, multi-modal understanding model CGVLM, intelligent agent product AutoGLM and even hardware, from algorithms to products, Wisdom has achieved complete self-development.

This company, which was involved in the development of large models before OpenAI’s ChatGPT became a blockbuster, emphasized that it had the same vision as OpenAI’s heroes. He insists on the research and development of pre-trained large models, and also wants to seize the fact that there are both B-side and C-side users in the API market.

In the second half of 2023, most of the industry’s answers to the next stage of the 100-model war will focus on using industry data to train vertical large models, but Zhipu chose the same as OpenAI, believing in the use of general large models. ability. It also continues to keep up with the release pace of OpenAI. Entering 2024, Zhipu has successively released CogVideoX, a video generation model that benchmarks Sora, and end-to-end voice models GLM-4-Voice and GLM-4- that benchmark GPT-4o. VideoCall.

Such a strategy comes from the financing resultsLook, it worked.

Zhipu’s latest round of 3 billion yuan in financing is consistent with the investment reasons stated in September when Zhongguancun Science City Company, a market-oriented investment platform established by the Haidian District Government, led the investment. , used to support the research and development of large wheel base models. These financings have made Zhipu’s valuation exceed 20 billion yuan, making it one of the most highly valued large model companies. In addition to state-owned assets, investors behind Zhipu also include Hillhouse, Qiming, Legend, and major manufacturers including Meituan, Alibaba, Tencent, and Xiaomi.

And this following has also begun to diverge recently. When OpenAI turns to a product company, Wisdom is still not that to C. Their external stance is that they still insist on spending energy on pre-training models, and the key information revealed in the financing is also the commercialization brought about by toB-based revenue. Income doubled. Compared with 2023, Zhipu’s commercial revenue has increased by more than 100% since this year. Its MaaS (Model as a Service) platform API annual revenue has increased by more than 30 times year-on-year. This platform has attracted 700,000 enterprise and developer users, the number of paying customers has increased by more than 20 times, and its penetration rate in commercialization and enterprise business has increased. Come higher and higher.

Compared with the capabilities that OpenAI cares about in its transformation into a product company, the customers that Zhipu has obtained obviously still value model capabilities more. This makes Zhipu look more like the previous OpenAI than today's OpenAI.

“Although OpenAI also has toB, its toC capability is also pretty good, and it is no longer completely comparable to Zhipu from the product line.” A leading VC partner said comparing the current situation of the two.

2. The Dark Side of the Moon has become OpenAI’s latest personal follower

One day before Zhipu released the financing news, Dark Side of the Moon released its "visual thinking model" k1, and launched the latest version of the web version as well as Android and iOS apps.

In the introduction, after the recent mathematical model k0-math benchmarked OpenAI's o1, k1 once again benchmarked OpenAI's o1 and GPT-4o.

The Dark Side of the Moon has become a new close follower of OpenAI.

As a head model company that rarely speaks publicly on the model side, it updated two k-series models within a month.

After Dark Side of the Moon became famous at the beginning of the year, founder and CEO Yang Zhilin set up a domestic AI technology idealism banner for him, which is most optimistic about the large model To C, "We are only optimistic about To C, because the company The goal is AGI (artificial intelligence), and the prospect of AGI in To C far exceeds that in To B," Yang Zhilin said many times in the first quarter of this year.

When talking about OpenAI, he mentioned in an interview at the beginning of the year that model companies should understand the relationship between AGI and products. “For us, it is not a means and an end., both are purposes”; “ChatGPT has not yet fully established continuous evolution based on user data”; “A company that does not care enough about users may not be able to achieve AGI in the end. ”

This shows that at that time, he wanted to break away from the narrative logic of OpenAI and tell his own story by talking about OpenAI.

Looking at it today, such a narrative also has more realistic considerations. . In the recent controversy surrounding Zhang Yutong, Many details have been released. In fact, this information also reveals how Yue Dark uses the resources it has raised.

For large-scale enterprises, financing is crucial, and it is necessary to raise funds with the top companies. Funds of the same size directly determine whether you can sit on the table in the initial stage. It was Zhang Yutong who played a key role. Several consecutive key financings made Yue Dark worth nearly 21.7 billion, making it one of the Six Little Tigers. As Kaifu Li said, these funds are enough for each company to do pre-training. But the implication is whether to do it and how to do it. Not necessarily.

From Yang Zhilin’s previous statement. Judging from a series of interviews and conversations, he is a person who "can afford to wait". For example, he talked about making the decision to focus on Kimi, saying that he made the decision in February, March or April this year, based on his judgment on the US market and his own observations. Propose to focus on Kimi. “It is important to make a product the best it can be. Cutting down on business is essentially controlling the number of people. Now if we do three businesses together, we will literally turn ourselves into a big factory, with no advantage. ”

For example, in terms of models, Dark Side of the Moon has never introduced its own models as a whole, but has chosen one of its capabilities—context length—to focus on display. Compared with Zhipu’s product line, this It's obviously also more cost-effective.

For example, in his previous statement, Yang Zhilin was very concerned about the number of people in Yuedan, "We have the smallest number of people in this group of companies."

In OpenAI, it represents the pre-training stage, not everyone. Everyone can afford to invest in OpenAI. Investors told Silicon Stars that the logic of Dark Side of the Moon is that Kimi is easy to use, not the logic of selling models, so they do not launch models frequently. Of course, the work of the model is very important, and it does not need to closely follow OpenAI. It’s thankless, but in a period when everyone must rely on benchmarking against OpenAI to accumulate better resources, this not only makes Yue Dark different from other large model companies that have benchmarked against OpenAI, but also makes its financing needs go further. The other way.

Until OpenAI. The complete difficulty of GPT-5 and the launch of o1, as well as the discussion about Scaling law hitting the wall, are destined to be affected in terms of financing due to Zhang Yutong's incident, and the strategy of benchmarking OpenAI has also changed.

On the stage of NeurIPS 2024 recently, Ilya Sutskever used his trademark simplicity to reveal the latest insights of artificial intelligence in 15 minutes. The most shocking thing was Sc represented by pre-training.Discussion of the "end" of aling law: "Pre-training as we know it will end. (Pre-training as we know it will end.)". Immediately afterwards, Noam Brown, a great figure in OpenAI and a core member of o1, also made some remarks, saying that o1 represents a new Scaling model with inferential computing as the core, which pushed the discussion of the end of Scaling law to a high point.

Yang Zhilin keenly captured this point. In a public sharing after the launch of o1 in September, he spent a lot of space explaining the newly launched o1. He said that the efficiency improvement of AI today is amazing, reasoning. The cost has dropped rapidly. He also said that when working on Kimi, he hopes to think more closely about products and models, “And this Sca ling has also undergone some changes. Originally, most of the scaling occurred in the training phase, but now most of the calculations, or more and more calculations, will be transferred to the inference phase."

After Dark of the Moon. They began to release their own o1 series, and used the strategy of "defeating GPT" that was commonly used when releasing models by companies such as Zhipu. After the recent release of the visual thinking model K1, Yue An said that its mathematical, physical and chemical capabilities surpassed OpenAI o1. On K0-mathe, Yue An said that in the industry's most commonly used mathematical ability benchmark test MATH, the first generation of k0-math The model performance exceeded OpenAI's o1-mini and o1-preview.

When OpenAI all acquiesced that pre-training has become less important, it naturally became reasonable not to invest heavily in pre-training. When OpenAI became a product company, it long ago declared that it would fight for the product. Chinese companies will naturally follow closely.

According to Silicon Star people, at the same time as Wisdom Spectrum, at least two of the leading large model companies are advancing their own new round of financing plans, but whether to release financing news may have their own considerations. In terms of large manufacturers, Byte and Alibaba also have investment plans in large model companies at the end of the year. Combining the actions of various technology lines and product lines, although there are some voices of difficulty, generally speaking, everyone is still anxious to compete, and there is no real winner.

In today’s competition among large model companies in China, OpenAI is still a target that cannot be avoided. Especially in the financing process, your secret connection with the world’s first-class companies determines many things. .

The strategy of following OpenAI will continue, and everyone will continue to get what they need.