Zhihu users | The main obstacle to domestic research on ChatGPT is the fear that it will speak nonsense

Question: What are the obstacles hindering domestic teams from researching products like ChatGPT? Is it technology, money, or leadership?

Here are the responses from netizens that were censored and deleted by Zhihu:

Zhihu user: The main obstacle is the fear that ChatGPT will say inappropriate things… Large language models are completely black boxes, and you can’t guarantee that it will never say things that are not allowed. Once it says the wrong thing, it will be a devastating blow to any company. Therefore, no team will do it, and even if it is done, it is impossible to open it to ordinary people. Maybe in five years, GPT will replace Google in most parts of the world, but the mainland will still be using Baidu.

The editor of China Digital Times also excerpted some other users’ replies:

Zhihu user: Direction. This thing will definitely be used to make review robots. I have met several such teams. It is absolutely impossible to use it for large-scale model public use, otherwise the competitor will directly maliciously ask questions + report, which can make you bankrupt + eat prison food.

Jsgfery: Lack of people with a long-term vision, who can mobilize a large amount of resources, and are willing to bear a considerable high risk of failure. Innovative research requires a large investment, high risk of failure, and it is difficult to see the application value in the short term. The domestic environment is impetuous, insiders seek stability and fear risks, outsiders are arrogant and short-sighted, and the tolerance for failure is very low. If no phased results are seen in one or two years, they will be scolded as liars. Such an environment encourages incremental research and does not encourage innovative research.

Jiahui Lincoln: You: Boss, I recently want to apply for a project, a chatGPT chatbot, which can… Boss: What benefits can this project bring to the company? How much value can it create? You: Teacher, I recently want to apply for a project, a chatGPT chatbot, which can… Supervisor: Isn’t this an expert question and answer system? It’s been outdated for a long time. Everyone is chasing hot topics now, you should find a popular research field that is easy to publish papers.

import Pandora: The chatGPT language model runs in a black box, and the developers don’t know what it will say in the next sentence. Therefore, it is unlikely that we will have a team to take this risk. If I develop it, I will hire 100000000 985 graduates to directly sit in the background and answer questions manually. If the questioner wants to fish for information, I will directly hand it over to the cyber police. Is the plan feasible?

Good teacher: Not to mention how much computing power GPT3 used, do you have a concept of how many web pages it crawled? How many sub-models were invested in classification, do you have a concept? What is lacking in China is the data after calibration, massive authoritative data, which is too far from GPT. To put it bluntly, after crawling Zhihu and training, you can get a customer service that posts advertisements.

Baoshu: Combining my two years of experience and observation at Damoyuan, the gap in action is the gap in technical leadership, the gap in thinking is the gap in values, and the gap in ability is the gap in national strength.

Gu Chen: The pursuit of a free environment, first, scholars have the habit and space for free research, second, non-governmental research institutions have the space to grow and become stronger, and third, there is a policy environment and legal protection to protect the reasonable returns of contributors, including individuals and companies.

Xianyu Intelligence: The biggest obstacle is vision and atmosphere. Innovation has the risk of failure, while following the trend is more stable and good at following up. Standing on the shoulders of foreign achievements, doing a XX Chinese version, although you may not be able to eat meat, you can definitely drink soup, and you win! Unfortunately, some things really can’t be replicated. At this time, you will say, why don’t they open source!

Wu Bin: Doing this requires money, and more importantly, it requires a “crazy” founder who can persist in doing things that others think are impossible. The research and development organization of ChatGPT is OpenAI, and one of the founders of OpenAI is Musk. Musk once answered a similar question, and the answer was: “Where did Jack Ma go?” Therefore, the core answer to the question is: lunatics and aliens are difficult to survive in China for a long time.

doonny: How can technology make money faster than shooting “Manjianghong”.

Giant Nest: I want to talk about the social form, but I definitely can’t send it.

Spring and Autumn: Nothing. The domestic scientific research environment, you know.

lili: First of all, AI must have the ability to identify all sensitive points. Not only can it identify the past and the present, but it can also predict the future, at least 10 years later. It is estimated that 550W computing power is not enough, and 550WWWWWW may be enough.

Out Man: A question + an answer, maybe a product is gone, which capital dares to invest? Even if you finally manage to keep this product alive, it has already changed beyond recognition, and even the cost you spend on manual screening may exceed the R&D investment.

Zhihu user: If you, have to get to the bottom of things, have to explain the reason for every single thing. Then I can only say that in the end, you will find that it is an ideological issue. This is not my subjective guess, this is an objective fact, and most people know it, but they have to defend it. This is also normal, the ass determines the head.

Baladin: If you want to build a high-rise building on flat ground, you must first lay a good foundation. If you don’t have a foundation, you can only build a three-story small building.

Zhang Yiling: Regulatory documents for AI synthesis were quickly released within a week: https://www.gov.cn/zhengce/zhengceku/2022-12/12/content_5731431.htm After reading the details of the document, does a product similar to chatgpt have a way out? But anyway, ordinary companies or ordinary research groups don’t have the money to burn large-scale models. The problem is that this is the full-chain management of AI synthesis. Simply put, you are doing an aicg annotation outsourcing within the scope of management and filing permits.

Jia Xintang P useless: I have an idea, we can research an AI that reviews what chatgpt says, we can’t say it casually, but we can not say it casually! Review AI-chineseGPT, after the project is established, you only need to add a sensitive word library, the funds come!