Hengyu originated from the Cave Fei Temple
Qubits | Official account QbitAI
360 is on the cusp of GPT to usher in a new spring.
Within 2 months after revealing the relevant plans, 360 quickly gave a ChatGPT-like product. Half a month after the public display, just yesterday, the 360 model entered the landing acceleration period:
The "360 Intelligent Brain" developed based on 360GPT first landed in search scenarios and opened internal beta to enterprise users.
Previously, the 360 version of GPT in the experimental stage was demonstrated live under the operation of founder Zhou Hongyi and unveiled to the public.
Overall, its on-site performance is remarkable, and there are indeed small omissions, but from the stock price and market value after the display, it does not affect the market's expectations and confidence in it.
Zhou Hongyi described the 360's large model strategy as "two wings flying together", on the one hand, stepping up the self-research of core algorithms, on the other hand, launching related products as soon as possible to seize user scenarios.
How much real is the 360 version of GPT that began to seize the user scene?
Qubit got the new generation of search product internal test code embedded with 360 "wisdom brain", the mouth is groundless, the evaluation is the real chapter——
360GPT version of "New Search": Focused on differentiation
The 360GPT version of search, which has not yet officially announced its official name, is a combination of 360 search and large language model technology, which is essentially a new generation of "intelligent search".
Judging from the effect of Zhou Hongyi's earliest on-site demonstration, the 360GPT version of search initially combines the accuracy and timeliness of the search, as well as the "wisdom" of the large language model.
In short, 360GPT search has an intelligent advantage; 360GPT search results are less likely to be than GPT.
With this initial impression, we did a more detailed experience evaluation after getting the internal test code. Throughout the process, we compared the 360GPT version of New Search, New Bing and ChatGPT.
During the demonstration on March 29, Zhou Hongyi temporarily threw a timeliness question to the 360GPT version, "how many goals Messi scored for the Argentine national team". That morning, Messi scored three goals in a match, increasing his national team goal count from 99 to 102.
The 360GPT version of the search is based on the search of "Messi's national team scored more than 100 goals", and twice gave the answer of 100 goals. At that time, Zhou Hongyi was not satisfied with this answer, and he also stressed that 360GPT was still a child and was still learning.
Half a month later, we asked this question again. This time, the 360GPT version of the search gives more time-sensitive answers, and the answers are accurate:
New Bing has already demonstrated the market space for "new search" - due to the impact of GPT's new Bing, Google downloads and market share have declined significantly.
The next test is on difficulty, using social events to test its ability to understand and generalize.
First of all, regarding the matter of "Yu Minhong retaining people", we threw the question to the three contestants:
The answer given by the 360GPT version of the search engine better summarizes the key content of the event.
New Bing gives a brief description of the event itself, followed by a summary.
As for ChatGPT, in response to this request, it just translated the question into English without adding any additional information:
After testing and refining the abstract, we also experienced grammar rewriting skills, such as sentence voice rewriting, correction of language problems, and editing of sentences.
At the original press conference, the question that made the workers smile the most was undoubtedly Zhou Hongyi's question: "What does it mean to be the boss for another ten minutes?" ”
The answer given by 360GPT is a bit of workplace emotional intelligence: "We should actively respond to the instructions of leaders and carefully prepare meeting materials for their reference." ”
This time, we wanted it to rephrase "your advice sucks" to put it mildly, and 360 Search gives the answer "your advice isn't perfect":
Subsequently, we asked the 360GPT version to search for social hot topics and do reading comprehension.
For example, ask it, is Zibo barbecue still called Zibo barbecue when eaten in Beijing?
Zhou Hongyi himself has repeatedly called ChatGPT a real artificial intelligence, but compared with search engines, timeliness has become its shortcoming.
For example, in the face of localized problems such as "Zibo" + "barbecue", ChatGPT, a foreign monk, really scratches his head. The AI brain of 360 Search can correctly capture the reason for the popularity of "Zibo BBQ".
This is the advantage of combining large models and search engines, after all, relying on large models alone cannot solve the shortcomings of timeliness lag and accuracy to be verified.
And we also found that all replies to the 360GPT version of the search will show the cited source for users to trace.
In subsequent evaluations, in the face of non-timeliness problems, ChatGPT also had problems such as data errors, which shows that domestic search engine manufacturers do have advantages in the richness of Chinese corpus.
Continue to "do things" around the hot spot of Zibo barbecue. We tried to make it use the classic theory in the book "Positioning" to give Zibo barbecue a marketing plan, but I didn't expect the answer to look quite reliable:
After a passing experience, you can probably perceive that the 360GPT version of search blessed by "Wisdom Brain" is not only an enhanced version of traditional search, but also can improve the search experience, including Query's intention is clearer, the results given are more accurate, etc., it can also summarize, grammar correction, polish sentences, and generate marketing plans.
Why can 360 engage in the "new search" of GPT blessing
In fact, 360's ability to create its own large model is seen by the outside world as "a natural fit".
In general, 360 has many advantages for developing large language models, such as searching for genes, technology accumulation, data accumulation, engineering, etc.
OpenAI has taken the lead in achieving a breakthrough of 0 to 1 on the road of large models, and as a successful practitioner of parametric violence aesthetics, it has helped technology companies around the world point out the technical direction.
There are no insurmountable technical hurdles for Chinese companies, and the rest is only a matter of time guided by long-termism.
Of course, who can take the lead depends on the "endowment".
For a big model to succeed, in addition to the technical roadmap, data is very important. What kind of data is used for training directly determines the knowledge ability of this model. High-quality books, journals, and paper corpus are indispensable.
And search engine companies happen to have this innate advantage. Powered by its own search engine, 360 crawls hundreds of millions of pages every day, filtering out high-quality and spammy content.
At the same time, the long-term question and answer section can provide about 600 million question and answer pairs, from which 360 screens about 3 million training corpus and marks it for training.
Looking at the subjective conditions, this is also an important factor in whether 360 can do its own GPT well - the determination of the helm.
In the face of the unknown future development of large models, in the general split of AI bigwigs into "GPT is too dangerous, to suspend research" and "although dangerous, but the pace cannot stop", Zhou Hongyi put forward a third view: not development is the greatest insecurity.
With such knowledge and determination from the founders and leaders, it was only then that the 360 version of GPT quickly met.
At the same time, Zhou Hongyi also conveyed several advantages and opportunities of 360 GPT:
First, 360 has search engine genes, relevant technology accumulation, obvious engineering advantages such as data cleaning and processing, and the company's team has previous experience;
Second, 360 has accumulated high-quality annotated corpus at the C-end all year round, such as the 600 million question-answer pairs mentioned above;
Third, the scenario advantage, that is, the traffic advantage formed by national Internet products such as 360 search, browser, and security guard, as well as the user feedback advantage.
Don't just think of GPT as a chatbot
On March 29, Dafang held the "newborn child" live demonstration, and by April 16, the internal beta version of the new generation of intelligent search products was released, and behind the rapid implementation of the strategy was the steering party's cognition and judgment of new technologies and new waves.
When talking about the topic of GPT's "whether it has real intelligence", Zhou Hongyi claims to be a staunch "intelligence", and the GPT in his eyes is far more than a "fill-in-the-blank" trained with huge human language.
When experiencing GPT-3.5, Zhou Hongyi believes that GPT already has real intelligence like humans.
Understanding the agent cannot be limited to the shallow performance of the chatbot, and the real power is the super brain behind the chat window, which "represents the advent of a super artificial intelligence era."
At the same time, Zhou Hongyi also knows that once the AI under the blessing of the super brain is out of the application scenario, it is easy to be high and low, and it cannot really improve productivity. In the long run, in terms of industrialization, engineering, scenario, and productization experience, GPT still has a vast space worth expanding and exploring.
To make things, it is necessary to seek the landing application of real scenes.
Imagine that in the future, a product that does not have AI capabilities basically means being eliminated.
On the GPT "scene innovation" blueprint given by 360, the most self-evident answer is the combination with search engines, and now we have seen the landing products that have entered the internal test;
Secondly, it is to combine 360 browser, security guard and other products to provide personal assistant capabilities for users with more than 500 million monthly activities, shaping the 360 version of "Jarvis";
In addition, for the vertical SaaS application of small and medium-sized enterprises, 360 plans to launch smart office products, that is, "360 Intelligent Brain" to empower enterprises in office collaboration, marketing, communication and other scenarios, and upgrade the productivity and creativity of enterprises with large model capabilities.
Of course, 360 GPT isn't just about seeking value on To B and To C.
Zhou Hongyi also said that in the past few years of building a security system in 360, it has accumulated many To G customers, and training private AI services, that is, enterprise-level GPT, according to the needs of government departments and enterprises, is the direction that Zhou himself is also optimistic about.
In addition, in terms of 360's advantageous security business, 360 will follow Microsoft Copilot, launch 360 Security Copilot, and combine with 360GPT and intelligent application components to form a 360GPT security application framework.
The specific scene is anchored, and a new problem arises, that is, the business closed loop.
After all, for a large company, talking about ideals and vision, not talking about reality, is like a gimmick that attracts attention, floating like a castle in the air.
Zhou Hongyi said in the conversation that he can roughly describe the same path as Microsoft and OpenAI cooperation, that is, while investing resources in the technical level of core algorithms and continuing to innovate; On the other hand, AI can quickly cover existing products, and quickly commercialize and commercialize.
Immediately afterwards, under the advantage of occupying the scene, the core algorithm technology is synchronously developed.
"GPT is expensive to do business." Based on this recognized fact, Zhou Hongyi said that GPT services without a commercial closed-loop cannot be continued, and even the model of simply stuffing advertisements into the generated results may not be able to go far.
Finally, he concluded: GPT's future services may be chargeable.
Compared with many competitors on the same stage, 360 proposed the characteristic "two wings fly together" strategy, and Zhou Hongyi himself frankly admitted that there is a 2-3 year gap between the 360 version of GPT and GPT-4.
However, as Zhou Hongyi himself has repeatedly emphasized, in the face of this new industrial revolution that has just begun, the 360 version of GPT is still a baby, and the future is still a long way off.
In the face of this historic opportunity of the big model, 360 itself is also waiting to see whether it can usher in a new round of outbreak.