In Zhihu, I saw the first light of ChatGPT's transformation

If there is a change, Zhihu will have a response.

Search for "ChatGPT", as if you are instantly moved to the front line where you can hear the trumpet:

Innovation Works Chairman Kai-fu Lee summarizes the salvation in the post-ChatGPT era Universal rules for your own profession;

DeepLearning.AI founder Andrew Ng reminds people to put ethical and legal questioning before fanaticism;

First class Technology founder Yuan Jinhui feels that an urgent question that needs to be studied is how to distinguish text generated by humans from text generated by ChatGPT?

Hu Yongze, author of "Digital Survival" and School of Journalism and Communication of Peking University, proposed in the "Zhichao 8:30" roundtable live broadcast that the development direction of artificial intelligence does not necessarily have to be "human-like". ”, but should be to enhance people’s abilities;

...

Through questions and answers, topics, round tables, ideas, and live broadcasts of hot topics , industry leaders have appeared one after another in the high-quality Q&A community on the Chinese Internet; investors, researchers, entrepreneurs, and practitioners have connected with each other to explore all aspects of ChatGPT’s frontier and think about the impact of their discoveries on the future.

In just four months, the popularity of the "ChatGPT" topic on Zhihu has broken the record set by the "AlphaGo" topic since 2015. The current discussion volume has reached 220,000, and the total number of views has reached 380 million. .

#The topic is extremely popular, and the presentation method is also all-round and multi-dimensional.

1. The power of starting

On January 2, 2023, a Princeton University computer science student released GPTZero, which The program can "quickly and efficiently" decipher whether the author of an article is a human or ChatGPT.

After seeing the news, the Zhihu answerer "Kaspital" felt like he was being attacked. On the 10th day after the launch of ChatGPT, "Grasspod" Guo Biyang, a third-year doctoral student in the AI ​​Lab of the School of Information Management and Engineering of Shanghai University of Finance and Economics, has already started this work with his friends. "We were actually the first to start making the ChatGPT detector. Team," he said.

More than 30,000 people tried GPTZero within a week of its release, and the app crashed for a time. Guo Biyang became nervous. The original plan was that in addition to the detector, the eight-person team would also perform manual evaluation and linguistic statistical analysis. Obviously, we can't wait any longer now.

After training several detectors using existing data sets, on January 11, they released a demo on Zhihu, which is the first ChatGPT detector in China. It was approaching the Spring Festival, and Guo Biyang, who was supposed to be chatting with his family and reminiscing about old times, kept staring at the screen and paying no attention to anyone. "My family feels that we may be doing something big."

We are in an industry where technology iteration is getting faster and faster, and the most terrifying thing is It’s not that the technology has been leaked, but that not enough people understand your technology and you, let alone the revolutionary breakthrough of ChatGPT?

"People in the AI ​​​​circle want to promote their work, and many people choose to go to Zhihu. The work is circulated by everyone, and it is easier for more people to see it. ” said Xie Lingxi, an excellent answerer on the topic of Deep Learning, who has been “age-savvy” for ten years.

When Guo Biyang and his partners were racing against time to launch the demo, a series of work related to ChatGPT also appeared on Zhihu.

PENG Bo, an excellent answerer on artificial intelligence topics, An open source project ChatRWKV that openly benchmarks ChatGPT. The author is at Zhihu calls on more people to participate in co-building the ecosystem.

# At the end of February, ChatExcel, the first work that uses natural language to command Excel work, was exclusively launched on Zhihu.

However, the ensuing community feedback surprised Guo Biyang.

Before publishing, they compared the two detectors. Because they were trained on the data set, the actual effect of their detector was much better than GPTZero. Yuan Jinhui, founder of First-Class Technology, also talked about how to distinguish text generated by humans from text generated by ChatGPT on Zhihu, "It is an urgent problem."

Initially, Arguments that the work was pointless continued.

"We focus on detecting fake news, not fake news generated by ChatGPT. If a classifier can only handle fake news generated by ChatGPT, then I will change Isn’t it just a generator?"

Some people also think that "we can avoid being detected as ChatGPT through polishing."

Guo Biyang decided to respond personally. "It can be said that our detectors are not effective, but if we say that the detectors are meaningless, then I have an opinion. There are loopholes in the law, and lawless people can take advantage of the loopholes of the law. Does this mean that the law is meaningless?"

When people polish and modify ChatGPT content in order to avoid being detected, the role of the detector has been reached. There are ten thousand ways to bypass various regulations, and all we can do is increase the "cost of irresponsibility."

As time goes by, there are more voices of support and encouragement.

"The value of the detector varies from person to person, but personally I think this data set is more valuable and can do some interesting things." Someone said.

Hu Naying, a researcher at the Content Technology Department of the Cloud Institute of China Academy of Information and Communications Technology, said in the Zhihu "Zhicha 8:30" roundtable live broadcast that anti-cheating technology currently It is true that the technology is not as powerful as ChatGPT, and a "cat-and-mouse game" will inevitably occur. "But in the end, the devil is always better than the better." To publish a work is to choose to enter a complex system. The so-called complexity means that it is not linear and will not operate according to your expectations. It has many dimensions and variables and is difficult to predict. Therefore, there will be some "caught off guard", but there are also benefits and even unexpected surprises.

Currently, the detector of Guo Biyang’s team has 6,700 stars on Github. The data set and model may have been downloaded tens of thousands of times, and the article had 20 citations in less than two months. "These are things we never thought of before." He said with some emotion, "(The growth rate of article citations) is faster than any of my previous articles."

After the exclusive first release on Zhihu It didn't take long for WPS to contact the team behind ChatExcel. The open source project ChatRWKV has also been reported by leading technology media.

In addition to being difficult to predict, the complexity of the Zhihu system also includes a kind of adaptability. Publishing a work will change the system, and the system will in turn calibrate your products or research.

In November 2022, after Xie Lingxi’s team put an important paper on the arXiv preprint website, it was also immediately published on Zhihu, with the title showing a large meteorological model, "China For the first time, the accuracy of long-term weather forecast exceeds that of traditional numerical methods."

"The title made me quickly read the original article." After reading it, a netizen expressed his appreciation for the article's achievements. After a discussion with team members in the comment area, she shared her opinion:

"The input of the AI ​​model comes from ERA5 (the analysis field assimilated from observations and models), so It also uses model forecasts. If the model does not forecast, there will be no ERA5 (that is, the input of the AI), so the model still has to run, and there is no independent alternative to the model forecast. Currently, it is (traditional model assimilates AI forecast) > (traditional model assimilates traditional model) forecast)".

The AI ​​large model is trained on ERA5 data. ERA5 data is reanalysis data, which is a complete set of reanalysis data obtained through quality control and assimilation of observation data from various sources (ground, ships, radio soundings, wind balloons, aircraft, satellites, etc.) Analyze data sets.

The assimilation process here is to turn some observation data into standard gridded meteorological data. No claim should be made to transcend traditional numerical weather prediction methods (NWP) without covering these techniques.

"Yes, this should take a lot of time. The first problem is that data seems to be difficult to obtain (so there is very little related work at home and abroad)." Team members also agree.

In fact, "AI can't do it yet, or in other words, AI hasn't officially been able to do this yet. The main reason is that there is no data." Xie Lingxi later explained to us. To do this, the input end of the AI ​​must obtain data such as satellites and weather stations. No matter which country it is in, these data are highly confidential.

"Thanks to the European Meteorological Center's release of decades of assimilation data, we have completed this work."

Soon, Xie Lingxi updated "We accept this opinion." It does not refer to the entire "field of numerical weather forecasting". Pangu has indeed surpassed traditional methods in assimilating data for the first time: we used exactly the same test environment as NVIDIA FourCastNet to ensure the fairness of the comparison and the credibility of the conclusions."

After the study was released, Xie Lingxi’s team received many exchange invitations, including a report from the China Meteorological Administration. They were also contacted by the European Meteorological Center. It is said that there is a lot of discussion within the European Meteorological Center that many existing technologies will be replaced by AI.

Some companies have also contacted Guo Biyang to explore the possibility of launching related functions on the text platform. 『Our algorithm is still improving, mainly by increasing the robustness of the model and collecting more diverse data, hoping that the next generation model will be more effective.

Accept feedback from Zhihu netizens Finally, this is Guo Biyang's next goal. 2. Encounter with the "52 Hertz Whale" Researchers are rushing to launch their works first, and capital is also intensively "killing" AI large model talents.

At 11 pm on March 27, a message quietly came through WeChat: Wang Huiwen and Yuan Jinhui, the first-class technology company, reached a merger intention to create a Chinese version of OpenAI.

In the field of large language models (LLM), a framework that was once considered the least important by many people, its value has risen to the point where military strategists must compete. "Now it seems that this is a domestic model that is truly concentrating on the underlying technology." A Tencent AI algorithm expert who followed related Zhihu topics sighed.

Six years ago, Yuan Jinhui left Microsoft Research Asia to start a business making a deep learning framework. The situation at that time was no different from that of the famous "52 Hertz whale".

Inheriting algorithm applications and underlying hardware, the deep learning framework is called the "artificial intelligence operating system" and is a cake that startups can't even imagine. At that time, with its huge influence and strong promotion capabilities, Google's deep learning framework TensorFlow was already the deep learning framework with the most users at the time (Pytorch was still in its infancy).

Google is such a big company, with hundreds of people doing the same thing, how can you compete with others? Doing low-level software, having such strong competitors, and being open source... What Yuan Jinhui heard most at the time was "hitting a rock with an egg" and "trying to use a mantis' arm as a cart."

Because the 52 Hz frequency is much higher than that of any known whale species, scientists believe that a whale detected by US military instruments The call cannot be picked up by other whales.

The system software development cycle is very long. It took four years from its launch in 2016 to July 2020 before the deep learning framework OneFlow was open sourced. Because they could not bear the pressure of high uncertainty and the impossibility of any feedback before the finished system software was released, some outstanding colleagues resigned.

In fact, once technology reaches such a deep level in the system, there will naturally be fewer "whales" that can resonate with it. In Zhihu, Yuan Jinhui still found people who care about fundamental issues as much as he does.

Many Zhihu netizens know "Teacher Mu" (Yuan Jinhui's Weibo name) in the Weibo era. "Jinhui would write some very interesting news on Weibo, and everyone was spreading rumors about who this guy was." Yang Jun, technical director of Nvidia's AI computing architecture, recalled. Not long after Yuan Jinhui started his business, Yang Jun was also considering changing jobs, and the two met through Zhihu.

In Yuan Jinhui’s mind, Yang Jun, a friend who is an excellent answerer on machine learning and deep learning topics and the 2022 new knowledge answerer, has been steadily outputting all year round. High-quality content, and I have benefited a lot from his thinking.

In Yang Jun’s eyes, Yuan Jinhui is also a relatively easy-to-talk friend. Yang Jun himself has also comparatively analyzed the two mainstream deep learning frameworks and why at this stage there are still companies willing to invest huge resources in developing AI frameworks.

When Google released MLIR in 2019, the topic of deep learning compilers attracted much attention. The two soon appeared under the question "What do you think of Google's attention to the MLIR project?" and shared their views one after another.

Yuan Jinhui did not have a high opinion of MLIR at the time, and felt that the concept of compiler-compiler was a bit redundant. MLIR only provides a scaffolding for writing deep learning compilers and does not solve any specific problems in deep learning compilers.

Yang Jun is more inclined to think that MLIR is a good thing. "His analysis of MLIR's contribution, value and shortcomings has impressed me to this day." Yuan Jinhui said.

As the thinking and discussion deepened, Yang Jun continued to update the initial answer with new ideas and gains. Yuan Jinhui also maintains cognitive flexibility. In 2022, Yuan Jinhui updated his original answer again, "The development in the past two years shows that MLIR provides a 'scaffolding'... It is very meaningful."

People need an environment to communicate and collide with people with similar tastes. Zhihu’s attributes can achieve this. Yang Jun tried to explain this wonderful fate. For example, following your questions, interests, and articles, you can naturally judge whether you can talk more.

After OneFlow was open sourced, some netizens "had a sudden enlightenment and discovered that there are such solutions to some of the problems that they had been struggling with before." Some people also praised the framework design as "fresh". When the beauty of the design was understood by third-party developers and even students, Yuan Jinhui felt, "It's like you wrote a novel and readers appreciated it."

And when Guo Biyang fell into anxiety because of ChatGPT, It was this "connection between people" that finally rescued him.

He formed a group of "Lonely AI Researchers" through Zhihu, and found that many of his peers were also living in the shadow of ChatGPT's "Sophon". It was also there that he found collaborators on the detector project. After more than forty days of hard work, eight people persevered from the beginning, and no one gave up, even during the crazy moments of the epidemic.

They call themselves insignificant researchers, but they hope that the work they do is significant work.

In another corner of Zhihu, a passionate period shared by Xie Lingxi has received 32,000 likes.

"How long would it take for China to re-develop software like MATLAB and SolidWorks?" A question asked three years ago made him recount a long-lost past event.

More than ten years ago, several students from the Department of Mathematics at Tsinghua University wanted to make a scientific computing software that could rival Mathematica, the most widely used mathematics software. . The recruitment advertisement was posted in the dormitory building of the Department of Computer Science, but no one paid any attention to it. At that time, Xie Lingxi, a junior who had just learned Java and transferred from the mathematics department to the computer department, applied to join. After more than four months of writing countless documents, the prototype was finally made. A series of honors followed, and finally won the "Challenge Cup" national special award.

"Perhaps our project can be the best result if we only work on large-scale systems with passion. Without a mature business model or a healthy ecology, the project cannot develop in the long term. Go on." Many years later, Xie Lingxi talked about the commercialization of the project in her answer.

"Our experience has positive significance. It at least proves that in any era, there is no shortage of young people who dare to pursue their dreams."

3. Looking for “Millions of Treasures”

Yuan Jinhui and First-Class Technology are classified into a new track - AI large models. The minimum monthly salary offered for ChatGPT-related positions on the recruitment website is 20,000 yuan, and the maximum monthly salary offered is 100,000 yuan. levels.fyi shows that OpenAI offers a high salary of US$900,000 for AI/ML positions (L5).

As if overnight, we have returned to the era six years ago when capital could no longer catch up with the PhDs in mathematics, computer science, and statistics one after another. At that time, Zhang Yiming offered a reward of US$1 million on Weibo to recruit top machine learning talents; in Silicon Valley, some senior managers with technical expertise could earn an annual salary (including equity incentives) of millions if they worked for large listed companies such as Google. US dollars ("Millions of Baby")

At that time, in order to build the company's algorithm team, the Zhihu answerer who had just graduated and became the chief scientist of TuSimpleNaiyan Wang (王Naiyan) also joined that talent war. However, he took a different approach and left a hero post on Zhihu when answering "If you were an interviewer, how would you judge an interviewer's deep learning level?" and claimed that these three questions could test the subject's "eight successes" :

The most successful application of CNN is in CV, so why can many problems in NLP and Speech be solved using CNN? Why is CNN also used in AlphaGo? What are the similarities between these unrelated questions? How did CNN capture this commonality?

#One more question, why do many face papers add a local connected conv at the end.

These three questions are not typical textbook questions. Just as TuSimple's autonomous driving is a "new species", pioneering exploration often requires breakthroughs. There is no precedent for reference, and no ready-made answer. Only by gaining insight into the deeper connections between seemingly unrelated things can algorithm engineers remove the false and retain the true, allowing algorithmic tools to be better used and solve real-life business problems.

"It should be said that it is a good screening question." Jia Yangqing, who was still an artificial intelligence scientist at Facebook at the time, revealed the beauty of it in his answer, "It involves a very essential question. The question is why convolution can work." There are many angles to answer this question, such as regularization, statistics, programming and even neuroscience. Answers from different angles can reflect the subject's deep learning experience from different aspects.

If someone’s answers to the three questions are basically correct, it means that he understands CNN online and is the person Wang Naiyan is looking for.

Answers came one after another and kept lengthening the progress bar. "It's close, but not accurate", "Basically reliable! HR will contact you later", "If you are interested, send a CV to **", Wang Naiyan will respond to basically irrelevant answers, but more answers are below There was silence. As expected, 80% of people don't know why convolutional neural networks work. They just see it as a tool to run open source code.

A master’s degree from a prestigious Japanese university caught Wang Naiyan’s attention. "I received my undergraduate degree from Tsinghua University, and I was looking for a job after finishing my master's degree in Japan. His answer was very close to what I was thinking," Wang Naiyan said. The subsequent interview was also very good, and I immediately sent him an offer. This is the first algorithm engineer employee hired by TuSimple. Today, he is the head of the company's Japanese operations.

The Internet has transformed the geographical meaning of "nearby" into a digital meaning of "nearby". You may not know your neighbors one step away, but you have a high degree of trust in abstract systems constructed with complex technologies, such as Zhihu. For many real AI entrepreneurial teams, when they need to rely more on personal channels to grab people, this is often a good choice.

While Wang Naiyan was looking for algorithm engineers, Yuan Jinhui was also eager for talents to develop deep learning frameworks. After registering on Zhihu, the first thing Yuan Jinhui did was to "advertise" their work. Some people saw Yuan Jinhui’s articles and interactions and realized that in addition to big Internet companies, startups can also do the underlying architecture. Several full-time colleagues in first-class technology, including interns, were recruited from Zhihu in this way.

#More often, Yuan Jinhui will take the initiative. When he sees an interesting and insightful answer, he will check the other person's Github to try to get a more comprehensive understanding. Although I failed to "dig" some of my favorite people, everyone gradually became friends and would meet to exchange some opinions.

#The longer you work in your own field, the more articles you write, the more you interact with questions, and the better the community feedback is. An undergraduate who has interviewed for internship positions in multiple companies replied to "Are there any suitable system or compiler internship positions for undergraduate students in China?":

Among them, I saw that you have the ultimate pursuit of technology/coding. I think Mr. Yuan’s oneflow is a very in-depth company. During the interview, I directly chatted with the oneflow interviewer for an afternoon about C and parallel computing, starting from various optimizations. Techniques such as sso, stack/dynamic memory, various templates, functional style programming, and some recent ml system papers.

Wang Naiyan, who has always been low-key, is particularly active on Zhihu, writing articles and sharing technology, and is also an excellent respondent in the fields of deep learning, machine learning and artificial intelligence. Many of the fans who follow him are computer science students, and many of them eventually became members of TuSimple through "following". TuSimple currently has many undergraduates in algorithm positions, which is impossible in many companies.

#Actually, they are very good. In Wang Naiyan's view, if the recruitment model of major Internet companies is followed, some "uncut gems" without a glamorous educational background or shiny papers will be directly screened out by those hard standards.

# "Many times, what they have is not that important. We care more about the person's basic abilities and potential, whether they have self-thinking and enthusiasm for technology. , even if he is an undergraduate."

Now, Wang Naiyan still looks for talents from Zhihu. But what is different from the initial stage of entrepreneurship is that people are more passively looking for them. "Whether it's the people I follow, information streams, or recommendations, it has helped me filter out a lot of invalid information." Wang Naiyan said, "Really useful information will appear repeatedly in the information stream."

Usually he browses some hot topics and clicks on interesting answers to learn more. If it meets the needs of the company, he will transfer it to the human resources department.

No matter how technology develops, the Top1% of people will not change. Their enthusiasm and firm belief in technology are still what they value most.

4. Echoes of Time

After the release of ChatGPT, the Zhihu answerer "Trinkle" suddenly appeared "How to evaluate OpenAI's super dialogue model" ChatGPT?" Under the question, he disclosed that he was "fortunate to participate in the whole process of ChatGPT training" and presented his thoughts on the future world:

##"You can start to imagine what will happen after AGI. The world, I have been thinking about it for several months......."

Answer at the bottom, at In the acknowledgments on the official website of OpenAI, "Jiayi Weng" appeared in a list of contributors and was highlighted. People gradually learned that "Trinkle" is called Weng Jiayi. He is OpenAI’s first graduate employee with a master’s degree in the past two years, and is also one of the youngest R&D engineers on the team.

#Now, the answer has received more than 3,000 likes. Few people know that he once felt that he could not get close to OpenAI. "When I submitted my resume after graduation, I also thought that I could not get close." He said.

#Weng Jiayi started getting into programming in the first grade of junior high school. At that time, his focus was on Mathematical Olympiad, and learning programming was just to expand his mathematical ideas. I really felt the charm of programming after entering Fuzhou No. 1 Middle School in high school.

#At that time, he liked Card Constant very much. "Given a fixed problem, you can write a bunch of code, write the same algorithm, and have the same time complexity, but I can coordinate some things to make the same algorithm run faster than others." This kind of PK makes him very good. A sense of accomplishment.

At that time, the information group of Fuzhou No. 1 Middle School had an internal judgment system (OJ) online assessment, which contained various historical records. Weng Jiayi often reached the 3rd Will stop once.

#In his sophomore year of high school, Weng Jiayi completely changed his focus from mathematics to programming. In order to be able to participate in the "Resumption of Diplomatic Relations between the Qing Dynasty and Northern China", he decided to participate in the Informatics Olympiad. At that time, many students in the information group were playing Zhihu, and he also registered an account. At that time, he would not have thought that a few years later he would become in the eyes of many netizens "the genius boy who started playing Zhihu in his senior year of high school."

#The year AlphaGo defeated Li Shishi, Weng Jiayi also entered Tsinghua University as he wished. Because of his poor performance in the Informatics Olympiad, he transferred to the Computer Science Department by achieving top ten GPA in the entire department during his freshman year. I got involved with reinforcement learning during my sophomore year.

#When meeting Professor Zhu Jun for a one-on-one chat, Professor Zhu Jun asked him what he wanted to do? There are three directions in the group: Bayesian, adversarial training and reinforcement learning. Although he chose reinforcement learning, he did not know what reinforcement learning was at the time.

"At first I thought it was similar to doing GAN (adversarial training)." After choosing it, I knew I wanted to play games. To get started, he later played a lot of games.

# If Weng Jiayi mainly dived and collected information on Zhihu in high school, after entering Tsinghua University, he had more desire to share. Perhaps this is related to the life goal he set in high school - to gain more influence and help more people. These all require connections with machines and people.

#The most important work he released on Zhihu is the reinforcement learning algorithm library Tianshou (Tienshou), which is his senior graduation project. It is also the research that has had the greatest impact on him so far. Later, I was able to work in Open AI and also benefited from this "first work" experience.

The original version of Tianshou was written by four people in the laboratory two years ago using Tensorflow. It was very slow and not used by many people. He tried refactoring some of the code inside, but it didn't work. Later, I just scrapped everything and started over. It turns out that the benefits brought by streamlining the framework are not only at the code level, but also at the performance level.

After the job was released, sharp-eyed netizens discovered this:

##" If the same algorithm such as dqn is also pytorch, why is your code so much faster? It feels like the logic of other codes is similar except these two parts."

"Code also has a soul (escape, it's the implementation details..." he said.

That time "really made me realize that if I want to create In order to have influence, you should write some basic things or make achievements in engineering, rather than making achievements in some research aspects." He said.

There are many low-quality implementations in the AI ​​field, probably because researchers have insufficient engineering capabilities. If you bring some engineering insights into the research, you will have different gains.

In addition to expanding the influence of his work, Weng Jiayi is also keen to participate in topics related to undergraduate life at Tsinghua University. "Do you regret choosing to study at Tsinghua University?", "Studying computer science and technology at Tsinghua University is the best What kind of experience?" The question has left traces of him. A past where he got out of confusion and gradually strengthened his direction, making his answer still popular today.

"I feel like this kind of mentality is exactly what I need. I am about to be tortured crazy by your flying bitch." A Tsinghua alumnus expressed his own voice.

"Learn to admit that you are inferior to others and reconcile with yourself." Weng Jiayi wrote. In high school, he found that no matter how hard he worked, there were always people overlooking him from a higher place, whether it was the Informatics Olympiad or the cultural class . The same is true in the first two years of college. My English is not as good as my roommates. There are always some people who can pass some classes that I don’t understand without studying.

"You must learn Define evaluation indicators and stop following the crowd." This was his suggestion. In his junior year, Weng Jiayi completely changed his evaluation indicators and returned to his original intention.

He no longer brushes up his GPA and "writes" papers step by step, preferring to do some "useless" but interesting things, such as writing code. "When I write a personal project, I feel like I am creating a work of art." Regarding writing code My love for open source projects also influenced my subsequent decision to study abroad.

Every time I move to a new site - spring recruitment, autumn recruitment, doctoral application, Looking for an internship position in China - he is not shy about sharing his experience. Whether it is the happiness of having multiple offers in hand or the gloominess of being rejected by "Quanjude" when applying for a doctoral degree, every time he answers, he gains high popularity.

Now, Weng Jiayi has contributed 33 answers, published 3 articles, accumulated more than 20,000 followers, and received a total of 28,966 likes. These numbers more or less quantify "use your own strength to help more people."

In "How is the progress of your 2022 autumn recruitment?" "Under the Q&A, he shared his experience of investing in hundreds of companies and wrote at the end of this answer with nearly a thousand likes, "Choice is greater than effort."

If I had not made the optimal decisions based on the current environment, had not participated in informatics competitions, chosen intensive learning, applied to study abroad, or persisted in studying for a Ph.D., would it have been possible to reach this point today?

Hard work cannot make up for mistakes in decision-making. Why is Google lagging so far behind OpenAI in AI? During the interview, he asked rhetorically, and then directly gave the answer, "Because they chose another direction, a different direction from OpenAI."

Recently, someone Asked "Is there any way to join OpenAI to do research?... I think doing research in the company seems to be a more effective thing. Can you give me some suggestions."

He forwarded the link of the highly praised answer to the other party.

The spectacular coral reef is a major project that the coral polyps have worked hard for many years. Coral reefs occupy only 0.5% of the global seabed and ocean floor, but are home to more than a quarter of marine life.

In Zhihu, every technological "Zhihuer" is like a small and magical coral polyp and zooxanthellae. Through questions, answers and attention, each other interacts with each other. Energy information is exchanged, and over and over again, a more advanced system emerges, attracting more cutting-edge technological "species" to inhabit it, including some top scientists.

Zhihu Vice President of Strategy and Community Business Leader Zhang Ning once said that the total number of people engaged in scientific research, study and work on the site is as high as 5.44 million, and only the technology Internet In the fields of mathematics, physics, astronomy, artificial intelligence and other fields, the average daily output of graphics and texts is more than 20,000, and the number of answers, articles and videos in mathematics, physics, astronomy, artificial intelligence and other fields exceeds 1 million.

After ChatGPT, when front-line events such as the release of Baidu’s “Wen Xin Yi Yan”, the release of GPT-4, and Microsoft’s integration of AI dialogue functions occurred, big names in the industry all Get together here to discuss it as soon as possible.

On March 28, Chinese mathematician Zhang Yitang issued an invitation on Zhihu: "I will be invited by Harvard University and several European universities to do a Live broadcast, the theme is: Non-positive sequences in analytic number theory & the Landau-Siegel zero (Non-positive sequences in analytic number theory & Landau-Siegel zero point. "

This time, in Zhihu, you can hear the sound of the horn again.

The above is the detailed content of In Zhihu, I saw the first light of ChatGPT's transformation. For more information, please follow other related articles on the PHP Chinese website!