AI Weekly | ByteDance's large model training was "poisoned"; Microsoft will terminate the Azure OpenAI service for individuals in China

ByteDance's large model training was "poisoned"

Recently, it was reported that ByteDance had a large model training incident that was "poisoned" by interns. It is reported that the incident occurred in the ByteDance commercialization team, because intern Tian Moumou was dissatisfied with the allocation of team resources, and took advantage of the HF (huggingface) vulnerability to destroy the code through shared model injection, resulting in damage to the team's model training results. On October 19, ByteDance issued a clarification stating that the interns involved maliciously interfered with the model training tasks of the commercialization technical team's research project, but did not affect the official commercial project and online business, nor did it involve other businesses such as ByteDance's large model. The online rumor that "involving more than 8,000 cards and losing tens of millions of dollars" is seriously exaggerated. After verification, the person involved has been interning in the commercial technology team and has no internship experience in AI Lab. ByteDance said the intern had been fired by the company in August. The company also synchronizes its actions with industry alliances and schools, and hands them over to the school to deal with.

Comments: This incident exposed the security management problems in ByteDance's technical training, including permission isolation and audit of shared code. An industry source said that conducting authority isolation and auditing is conducive to protecting the company's core data and intellectual property rights, preventing data leakage, and improving the security of data and systems. For example, real-time monitoring of permission usage to detect permission abuse and abnormal operations in a timely manner; Regularly audit permissions to check whether team members' permissions comply with authorization policies and whether there is any abuse of permissions. However, this can also be difficult, including the cost of cross-departmental cooperation and the investment of resources for regular maintenance and updates.

Microsoft will terminate the Azure OpenAI service for individuals in China

On October 17, a number of developers received an email from Microsoft saying that due to local regulatory requirements, Microsoft's Azure OpenAI service will be shut down on October 21, and only enterprise customers in Chinese mainland will be able to subscribe to Azure's OpenAI service in the future. A developer said, "This is the only window for individuals to be able to use OpenAI in China in a compliant manner, and it is now closed." At the same time, some developers are dissatisfied with Microsoft's "four days of processing time". In response to this matter, Microsoft has not made a statement for the time being.

Comments: On June 25, OpenAI sent an email to some domestic developers that it would terminate its API services in China. At that time, "Microsoft Technology" also published a document to teach users how to migrate to Microsoft Azure OpenAI. Under the agreement between OpenAI and Microsoft, OpenAI will receive a 20% share of the revenue generated from Microsoft's sales of OpenAIGPT services. At that time, there were voices saying that after OpenAI terminated its API service to China, the biggest winner was Microsoft. After just 4 months, things changed. In response to the termination of Azure's service, some developers said that this is forcing individuals and small and medium-sized developers to use the domestic large model.

Google's AI team structure has been readjusted

Following Google's merger of Google Brain and DeepMind, two AI labs, in April this year, Google's AI team structure has been adjusted again. On October 17, local time, Google officially released an open letter written by Google CEO Sun·dar Pichai (Sundar Pichai), which mentioned the merger of the Gemini App application team into Google DeepMind to further integrate the AI team. The assistant team, which focuses on home products, will be merged into the Platforms & Devices department. At the same time, Prabhakar · Raghavan, senior vice president of key businesses such as search and advertising, will move to the role of chief technology officer, while Nick Fox will take over the role of head ·of search and advertising.

Comment: This adjustment is considered another move by Google in the field of artificial intelligence. At the same time, Google's search business is in crisis, having previously been judged by the court to be a market monopoly and at risk of being broken up, and Fox looks to have high hopes. "AI is evolving faster than any technology before it. To continue to accelerate the pace of progress, we've been working hard to simplify our structure, including creating Google DeepMind and joining the platform and device teams. Pichai mentioned.

Amazon invests $500 million in nuclear energy

On October 16, local time, Amazon announced that it would invest more than $500 million to develop small modular nuclear reactors (SMRs) to meet the huge demand for clean energy as its cloud computing services expand to generative AI, and as part of achieving net-zero carbon emissions. Amazon Web Services (AWS) has entered into an agreement with Virginia-based utility Dominion Energy to explore the development of an SMR near the existing North Anna nuclear power plant. Amazon has also struck a deal with Energy Northwest to fund the development of four SMRs in Washington state, which will supply power directly to the grid, helping to meet the energy needs of Amazon's operations.

Comments: Among the major manufacturers, in addition to Amazon, Google and Microsoft have also shown strong interest in nuclear energy. Google has secured a new nuclear contract, while Microsoft plans to restart the Three Mile Island nuclear power plant by 2028. This trend shows the energy anxiety of tech companies, as well as their efforts in the pursuit of sustainability.

Nvidia shares hit a record high at one point

Boosted by TSMC's third-quarter earnings report, many semiconductor stocks rose on Thursday, and Nvidia shares hit a record high of $140.89 per share during the session, but the closing share price rose to close at $136.93 per share. Since September this year, Nvidia's stock price has begun to fluctuate and rise. Billionaire investor Stanley Druckenmiller, founder ·of Duquesne Family Office, said this week that his decision to dump Nvidia shares earlier this year was a "major mistake."

Comments: The outside world is quite concerned about the cooperation between AI chip giant Nvidia and the foundry TSMC. Whether the Nvidia Blackwell architecture chips can be successfully produced and delivered will likely affect the future performance of both companies. Previously, it was reported that the production of Nvidia Blackwell chips has been delayed recently. In order to meet the needs of customers, TSMC is maintaining a high level of capital expenditure.

Social platform X allows third parties to train AI using user data

On October 17, local time, the social networking platform X updated its privacy policy and will allow third-party "partners" to use X data to train AI models if users opt out. This change means that X, like Reddit and several media organizations, is likely to open up new revenue streams by licensing data to AI companies. The new policy will go into effect on November 15, at which time users can opt out of data sharing.

Comment: At the moment, the new privacy policy does not clearly indicate where users should set up data sharing, and the new privacy policy will take effect on November 15, and there may be options at that time. Previously, it was reported that the content of X's user's tweets was used to train the Grok AI model. Allowing third parties to train AI with user data could lead to new revenue streams for X.

Chip companies said that the increase in the semiconductor market is related to AI

From October 16th to 18th, at SEMiBAY, a number of industry insiders proposed that the global semiconductor market size is expected to exceed one trillion US dollars in 2030. Tao Tingting, marketing director of ASML, said that the growth of the semiconductor market in the next decade will be driven by AI, automotive and industrial fields. Dai Weimin, founder, chairman and president of VeriSilicon Microelectronics, believes that in the process of the semiconductor industry moving towards trillions of dollars, 70% of the increase will be related to AI, and semiconductor companies that have nothing to do with AI can only share the remaining 30%.

Comments: The recovery of the semiconductor market is not only cyclical, but also driven by the factors of industrial change. Over the past decade, one of the biggest drivers of growth in the semiconductor industry has come from mobile phones, PCs, and servers, and now from AI, automotive, and more. One of the driving forces behind the future replacement of mobile phones and PCs also comes from AI.

Ministry of Industry and Information Technology: China's generative AI service model has more than 600 million registered users

Recently, the Ministry of Industry and Information Technology said that the number of registered users of the mainland's generative AI service model has exceeded 600 million. According to Zhao Zhiguo, chief engineer of the Ministry of Industry and Information Technology, the scale of the mainland's core AI industry is constantly improving, with the number of enterprises exceeding 4,500. Nearly 200 generative AI service models have been filed and launched to provide services to the public, with more than 600 million registered users.

Comments: This year is a year to accelerate the implementation of large models, especially the application of large models promoted by platform enterprises, which may have driven a lot of users. Zhao Zhiguo mentioned that in the future, it is necessary to improve the ability of artificial intelligence infrastructure public services, build an artificial intelligence development platform with low threshold, easy adaptation and high availability, and accelerate the training, optimization and large-scale application of large models. At the same time, we will improve the supply capacity of computing facilities, accelerate the construction of a national integrated computing power system, and steadily increase the proportion of intelligent computing power.

TSMC ranked among the top eight in the world in terms of market capitalization after releasing its third-quarter earnings report

Boosted by TSMC's third-quarter earnings report, U.S. stocks TSMC closed up 9.79% on Thursday with a market capitalization of $1.07 trillion at the close. According to statistics from CompaniesMarketCap, TSMC's market value has surpassed Warren Buffett's Berkshire to become the eighth largest company in the world by market capitalization, second only to Apple, Nvidia, Microsoft, Google, Amazon, Saudi Aramco and Meta. TSMC's consolidated revenue in the third quarter was $23.504 billion, a year-on-year increase of 39.0%, and its net profit was $10.058 billion, a year-on-year increase of 54.2%.

Comments: TSMC released a signal that AI demand is still strong at the performance exchange meeting after the earnings release. TSMC management said at the performance exchange meeting that the demand for AI is real, and now the demand growth is "just beginning", and the current demand is crazy. TSMC's advanced process revenue share has also increased, with 7nm and below advanced processes accounting for 69% of total wafer revenue in the third quarter, up from 67% and 65% in the second and first quarters. At present, the fields that have rigid demand for advanced manufacturing processes include mobile phones and AI.

(This article is from Yicai)