People who complain that Baidu is "too anxious", you don't understand

Under the expectation of everyone, Baidu finally handed over the first answer sheet of "Wen Xin Yiyan" a few days ago.

As the first domestic and even global Internet giants to bravely "stand" out to benchmark the existence of ChatGPT, Baidu's conference directly attracted full attention, and also ushered in many doubts.

Source: Network

From before the press conference, screenshots of chat records confirming forged and saying that Baidu employees were required to answer questions on behalf of "Wen Xin Yiyan" began to circulate; After the press conference, the media directly took the "loss of XXX billion yuan in market value" as a summary of Baidu's achievements released this time; Even after getting started with actual measurement, the actual use of word of mouth has rebounded, and many people have begun to complain that "Baidu is too urgent".

The all-round complaining and discussion really made Baidu fiercely popular for a long time.

Although Baidu did choose to record rather than live demo, the key is that ChatGPT allows people to form their own "high expectations" in advance: everyone does not want China to miss this important opportunity for innovation and change, so it is natural to compare the "Wen Xin Yiyan" on high expectations that can match GPT3 or even GPT4.

Baidu's choice to release Wen Xin's words as soon as possible is precisely considering the timing - in order to seize the key AI transformation of the big model, Baidu has no time to grind and must rush forward. On the contrary, those who complain that Baidu is "too anxious" really do not understand the importance of this change, let alone see the urgent situation.

Big models, a new era of AI that is about to enter

Source: Baidu Intelligent Cloud

When it comes to the landing application of AI, many people are not unfamiliar.

Chinese technology companies, including Baidu, have helped a considerable number of enterprises achieve basic AI capabilities including language understanding, text review, text recognition, image review, and image recognition by developing their own AI technology stacks and building open capability platforms.

Compared with the AI capabilities of these single-function and subdivided industry scenarios in the past, ChatGPT and "Wen Xin Yiyan" not only bring a leap in natural language processing (NLP) technology, but also a manifestation of the arrival of the "big model" change in the computing power era.

In the last wave of computer vision technology with CNN convolutional neural network model as the core, many companies that want to actively try AI have encountered the same problem: they want to use AI, but the cost of independently collecting and annotating data is too large; Most AI datasets and neural models are also not directly available across domains.

In the end, the entire AI market entered an "island" development rhythm, and after several industries that were easy to apply AI (such as security, translation, speech recognition, and autonomous driving) were covered, the commercial application of AI fell into "stagnation". Behind this "stagnation" is the limited "intelligent ability" of the previous generation of AI, which can only solve relatively simple specific problems.

The various performances of ChatGPT shocked the audience, which is the best proof that the "big model" route has once again achieved a breakthrough in the "intelligent ability" of artificial intelligence.

To give a simple example: in the past, in order for artificial intelligence to learn to drive autonomously, we provided it with driving-related picture data, hoping to directly "train" it to learn to drive. The results of the training are frequently wrong in special cases (irregular road signs, other vehicles that do not obey traffic rules), which hinders the further promotion and application of autonomous driving.

Now, we can first use more computing power resources and time to throw all kinds of pictures, videos and even text information to it, let it continue to learn itself, and finally form a basic large model (Foundation Model), and then apply it to specific applications such as autonomous driving for application-related optimization and adjustment. It is equivalent to letting it learn to recognize the world and achieve a breakthrough in the level of "intelligent ability" before it learns to drive.

Taking the official investment of "Wen Xin Yiyan" as an example, the training data includes trillion-level web page data, billions of search data and picture data, tens of billions of daily average voice call data, and 550 billion facts of knowledge graph. The joint training of these huge and complex knowledge finally contributed the thinking chain and complex reasoning ability to "Wen Xin Yiyan", making it possible to solve various difficult problems.

In addition to the improvement of "intelligent ability", the "big model" trained by a variety of data also has its own innate advantages in application landing and commercialization.

Different from the "island" landing development of artificial intelligence in the past, each large model has a broad application scenario, and while landing in the new application scenario, it will bring more original data to the large model itself, which in turn continues to expand the intelligent capability boundary of the large model.

Finally, a "snowball" development route of large model capabilities of "continuous application, continuous absorption of data, continuous training, continuous increase of capabilities, and continuous expansion of application" is formed.

Even if the "big model" has far more demand for computing power and funds than in the past, at least it can confirm that along this route, there is more hope for artificial intelligence to gradually approach a higher level of cognitive intelligence, and even help human decision-making and exploration.

If you don't rush, it's really late

In February this year, when foreign media broke the news that Microsoft's $10 billion investment in OpenAI did not exchange for any shares, but only phased dividend rights (when the total profit reached $150 billion, Microsoft's dividend rights automatically stopped), the exaggerated cooperation agreement once puzzled many people.

The real reason is that like many tracks that can "snowball" in the past, the first movers in the "big model" track will have a near-monopoly advantage.

According to industry estimates, the training of large models often requires thousands of top GPU acceleration cards at the beginning, with a total value of more than 50 million, and the electricity, manpower and hardware depreciation of a single run is close to 5 million yuan. Many small and medium-sized companies that have been in the artificial intelligence industry in the past are likely to be unable to train competitive "big models", and the increase in the threshold will greatly widen the gap between leading companies and catch-ups.

Therefore, since the official announcement of the investment in OpenAI in January this year, Microsoft has held a series of conferences in just a few months, and realized the landing of various capabilities of ChatGPT in a series of Microsoft's blockbuster products and solutions, which can be seen that it is actually very "anxious" and "radical".

At this time, it is normal for us to look at Baidu's "hurry". Being able to quickly sprint and output results for the first time in a global technology company proves Baidu's own deep accumulation and investment in large models and AI research.

Baidu has many years of accumulation in key NLP model resources such as "knowledge graph", "semantic computing", "reading comprehension", etc. Among them, the development of knowledge graph can be traced back to 2014, through the gradual expansion of the development and application of multi-source heterogeneous knowledge graph, Baidu released the "Wen Xin Big Model (ERNIE) 1.0" in 2019, and ERNIE 3.0 developed by it has repeatedly won in global artificial intelligence semantic understanding competitions, and now it has become the key skeleton of "Wen Xin Yiyan".

In addition to the "Wenxin Big Model", Baidu also has the largest deep learning framework in China, "Baidu Feijiao", which by the end of 2022, it has gathered 5.35 million developers, created 670,000 AI models, and served 200,000 enterprises and institutions, ranking first in the comprehensive share of China's deep learning platform market. As the largest search engine service provider in China, Baidu has unique advantages in data resources.

In terms of key R&D investment, Baidu is also increasing its weight year by year, with core R&D expenses of 21.416 billion yuan in 2022, accounting for 22.% of its core revenue, the proportion of which ranks first in the entire technology industry.

Although Baidu's accumulation and investment have been considerable, the external "help" required for the growth of the "big model" has become the "last straw" that urges Baidu to deliver its results in such a short period of time.

Different from the past Internet era to create functions and create application ideas, large models have two key requirements "reinforcement learning" and "prompt", the former refers to the real application data of a large number of users, again integrated into the training data of the large model, simply put, it must be "used first, in order to become smarter".

"prompt" refers to the range of answers that the large model may answer from the beginning, so that it can better understand the user's problems during operation, which is equivalent to typing "small copy" in advance and memorizing some basic knowledge points, which is the homework that the large model must do to cover new application scenarios and industries.

By allowing Wen Xin to go online quickly and gradually carry out internal testing, Baidu will be able to have more user feedback and integrate it into training as data to optimize and improve the actual performance of existing applications. By quickly establishing an ecosystem, it can help Wenxin Yiyan to quickly land within the scope of its existing capabilities and obtain actual users, but also to expand application scenarios and industries in a planned manner, determine the next training direction of the entire large model as soon as possible, and realize the coordinated development of AI engineering and commercialization.

In other words, the press conference of "Wen Xin Yiyan" is not only a "handover of papers", but also a "throttle" that Baidu has stepped on with all its strength, and it is also the starting point for the next accelerated development. Looking at the product ideas of the past Internet era, it is really outdated.

Source: Baidu

The moment of innovation is coming, and if you fight hard, you should encourage it

Compared with the performance of the product, although OpenAI is very strong and Microsoft is fully committed to it, Baidu still warms up and releases it with a high profile, and actively enters the large model market, which is the most worthy of the attention of the general public.

Like major scientific and technological challenges such as self-developed chips and self-developed large aircraft in the past, the battle of AI "big models" is about to gradually become a war that China cannot lose. The only way to break the technological blockade and achieve independent development is to go all out to innovate, and it is a lifetime of innovation.

In particular, the current strong combination of "OpenAI + Microsoft" has threatened the subsequent development of the "big model" track of domestic enterprises. Baidu can be the first to stand up at this time to face the pressure output results, and it should encourage and provide assistance, not ridicule.

Only ordinary users are more tolerant of existing technology gaps, actively use and give feedback; The industry actively uses it to rapidly expand the vitality and business potential of the entire ecology; Policymakers tailor a more encouraging business environment and other parties to work together, and private enterprises like Baidu, which shoulder the responsibility of innovation, are likely to win the final game.