ChatGPT Plus discontinuation turmoil: insufficient computing power is a false proposition, data is the key

Interface News Reporter | Lee Kyung-ah

At noon yesterday, the news that ChatGPT Plus was suspended caused a shock in the domestic technology circle. According to media reports such as qubits, OpenAI suspended the sale of Plus, ChatGPT no longer supports Plus payment, and it is unknown when it will open after that.

Several users told Interface News that the suspension of ChatGPT Plus did occur yesterday, including middlemen who sell ChatGPT accounts in China and middlemen who sell virtual credit cards.

This morning, a number of users told Interface News that ChatGPT members can upgrade to ChatGPT Plus normally, and everything is normal after the upgrade.

Computing power is in a hurry?

The launch of ChatGPT Plus is for users to use ChatGPT during peak hours and get a faster service response, so paid ChatGPT Plus has a larger model capacity, more advanced training and better preprocessing than free ChatGPT, and can process input and complete output more quickly and efficiently in use. OpenAI launched its ChatGPT Plus subscription plan on February 1 this year, and ChatGPT Plus costs $20 (about 135 yuan) per month.

The power of ChatGPT Plus is based on the GPT-4 model, while the ChatGPT that was popular on the whole network used the GPT-3.5 model, and a more powerful model means a larger computing power.

At present, the computing power cost of large models mainly includes two parts: initial training cost and subsequent operating cost. Soochow Securities measured the computing power required by the GPT model: in the initial training stage of the GPT-3 model, according to the data of the OpenAI official website, the training cost of each token (token is a string of strings generated by the server to be used as a token for client requests) is usually about 6N FLOPS (FLOPS REFERS TO THE NUMBER OF FLOATING-POINT OPERATIONS PER SECOND, WHICH CAN BE UNDERSTOOD AS THE CALCULATION SPEED AND CAN BE USED TO MEASURE THE PERFORMANCE OF THE HARDWARE). Then OpenAI's initial training cost for ChatGPT requires 843 NVIDIA A100 chips.

When entering the operation stage, because the computing power required in this stage is closely related to the number of users, Soochow Securities made the assumption that if GPT-3 has 50 million daily active users, each user asks 10 questions, and each question answers 400 words, then this stage requires 16255 NVIDIA A100 chips.

Based on the amount charged by ChatGPT Plus, Soochow Securities deduced that the amount of parameters of GPT-4 is more than 10 times that of GPT-3, so it is expected that the computing power demand of GPT-4 is more than 10 times that of GPT-3.

The demand for such computing power scale will naturally lead to the conjecture that computing power is urgent. With user demand exceeding expectations, OpenAI is likely to take some steps to reduce access to ChatGPT. For example, since March 31, ChatGPT has begun to focus on large-scale bans in Asia, and according to various sources, the ban has affected up to one million accounts.

OpenAI wants to use ChatGPT Plus as a starting point to open up a rich business model. However, in the real-world experience, whether it is ChatGPT or ChatGPT Plus, there are significant corresponding delays and improvement in answer error rate. Since late March, OpenAI's access threshold for Plus paid users has continuously decreased the threshold, and its application model GPT-4 access limit has been reduced from 150msg/4hr on the first day to 100msg/4hr to 50msg/3hr on the first day until the latest 25msg/3hr, which is equivalent to a continuous drop in the threshold of 4 consecutive visits in less than a month, which is regarded as a manifestation of insufficient computing power.

However, there are also professional voices that believe that computing power is not the real bottleneck in the development of large models and functional experience.

Zhu Mingjie, founder and CEO of Krypton Technology, told Interface News that he did not regard the computing power problem too much as a bottleneck, made the model bigger, and the first thing to solve was the problem of the final result.

Krypton Technology is the first batch of artificial intelligence companies to use large-scale machine learning technology at the search engine level in the financial field, Zhu Mingjie told reporters, "In terms of first principles, in essence, OpenAI solves the problem of how to compress such a large amount of data into a model and turn it into a knowledge problem, which is its ultimate goal." Whether the intermediate computing power uses A100 or A800, and whether the GPU architecture or CPU architecture is used, are all solvable things. ”

In fact, this confusing ChatGPT Plus discontinuation event is the latest movement of the recent "ChatGPT chorus".

Canada and Germany have just joined Italy's "anti-ChatGPT" camp. On April 4, local time, the Office of the Privacy Commissioner of Canada (OPC) announced that it began investigating OpenAI, the company behind ChatGPT, involving complaints that "allege OpenAI collecting, using and disclosing personal information without consent".

According to Reuters, German Federal Data Protection Commissioner Ulrich Kelber told Handelsblatt in a comment published on April 3 that Germany may follow in Italy's footsteps and "block" ChatGPT due to data security considerations.

At the end of March, the Italian Personal Data Protection Agency announced that it would temporarily ban the use of ChatGPT with immediate effect, while opening an investigation into OpenAI to restrict its processing of user information in Italy. The agency believes that on March 20, the ChatGPT platform experienced a loss of user conversation data and payment information for payment services. In addition, the platform did not inform about the collection and processing of user information, and lacked a legal basis for collecting and storing personal information in large quantities.

Two days later, OpenAI CEO Sam Altman responded that it was a "major problem" and "we feel bad about it." He said it would comply with the Italian government's request and has stopped offering ChatGPT in Italy — even though it believes it complies with all privacy laws.

Two days before Italy's decision to ban ChatGPT, Musk led a group of tech bigwigs to call for the suspension of high-level AI system training, and an open letter he published caused an uproar. The letter states that "it should only be developed if we are confident that the effects of powerful AI systems will be positive and the risks are manageable." ”

Data is key

Looking at the recent controversy surrounding ChatGPT, we will find that on the surface, computing power is the bottleneck affecting the development of OpenAI, but in fact, data is the help and barrier to whether OpenAI can continue to develop.

"The expansion of computing power will not be a bottleneck for providing services at present, and it is not difficult to overcome." Wu Shenkuo, doctoral supervisor at the Law School of Beijing Normal University and deputy director of the Research Center of the Internet Society of China, told Interface News.

"ChatGPT certainly needs new regulatory rules, but the current regulations are still in force, and in that sense, starting with data is the most solid and effective way to regulate at present." Wu Shenkuo said that whether it is Italy, Germany or Canada, their regulatory entry point is data transfer.

"However, the regulatory measures against ChatGPT in countries such as Italy have so far failed to conclude that the technology itself is illegal." Wu Shenguo emphasized.

Wu Shenguo further analyzed the different positions of Europe and the United States in ChatGPT supervision, pointing out that Europe has strict data regulations, which are convenient from the perspective of data transfer, which can reflect Europe's core position on digital sovereignty and technological sovereignty, especially for American companies, "There is a digital economy and digital development dominance here." ”

He judged that European countries are more likely to follow suit, while the United States, due to its technological first-mover advantage, will be more relaxed in regulating ChatGPT.

"There will inevitably be obvious differences in the regulatory thinking of the United States and other countries, because this technology exists in itself, which can amplify its existing data advantages, computing power advantages and algorithm advantages." Wu Shenguo said bluntly.

He expects that after that, the transparency and explainability of rules in the process of ChatGPT computing will become the focus of regulatory relations.

"The key point here is whether the data collection is legal, whether the consent of the parties has been obtained, and whether it is illegal crawling of data; the second is the processing angle, whether it exceeds or exceeds the authorization of the right holder; Third, whether the output of data processing results meets the requirements of ethics, legal rules and security, and whether it will bring results harmful to the interests of society and the public. Wu Shenkuo summarized the key for ChatGPT to overcome the current data barrier from three levels.

OpenAI may also face a challenge in the process of removing data barriers.

OpenAI already faces awkward customer competition with its own patron, Microsoft, and many enterprise customers are beginning to make trade-offs between OpenAI and Azure OpenAI. Some people point out that OpenAI has reserved some models from Azure, forcing many companies to work directly with OpenAI, such as the March 1, OpenAI began selling access to its speech recognition model Whisper, which is still not available for Azure OpenAI services.

After the large-scale ban in Asia, OpenAI's system experienced some problems, while OpenAI-related services provided on Microsoft's Azure cloud remained stable, and customers could get Microsoft Azure security features when running the same model as OpenAI. Security has become a major advantage for Microsoft to snatch customers from OpenAI.

Next, OpenAI needs to figure out where the real bottlenecks are in the midst of complexity. The advice that OpenAI co-founder John Schulman advises others applies to himself: The ability to choose what problems to solve is more important than your own skills.