Original link: https://www.latepost.com/news/dj_detail?id=1725
Bytebig model strategy: self-developed platform
This week, ByteDance’s cloud computing service Volcano Engine launched the Ark large-scale model platform, which introduced the large-scale models of 7 start-up companies/research institutions such as Zhipu AI and MiniMax, and sold services to the outside world. Relevant companies and institutions set up booths at the Volcano Engine event venue, and the founders or co-founders announced that they would have more in-depth cooperation with Volcano Engine in the future.
In order to attract start-up companies to put their models on the Volcano Engine, ByteDance quickly dispatches idle computing resources for businesses such as Douyin to the Volcano Engine, and sells computing power services at a lower price than its peers. Tan Dai, President of Volcano Engine, said that most large-scale model companies in China use the Volcano Engine for training, and it is logical to do reasoning on the Volcano Engine.
At the beginning of this year, in order to seize the opportunity of large-scale artificial intelligence models, ByteDance formed at least three teams to develop large-scale models, and asked Nvidia to place an order for GPUs exceeding US$1 billion. The founder Zhang Yiming, who stepped down as CEO two years ago, also began to look at related Papers, exchange ideas with some teams.
ByteDance, which has invested a lot of resources, aims not only to develop a large model like OpenAI, but also hopes to build a platform with its rich GPU reserves to help startups train and sell large models as a middleman.
According to Tan Dai, they will introduce more large-scale models in the future. The large-scale models developed by ByteDance will not only be applied to business, but will also be sold on the platform. He said that this is based on two judgments: the large-scale model market will not have a situation where a few large-scale models take all, and enterprises will use multiple models to develop applications or transform businesses.
“Super large models are effective, but cost-effective. Not all problems require super models,” Tan said in an interview. In addition, different industries solve different problems, and the model training data is also different, so there will be specific industries or models. Large models exist with varying amounts of parameters (determining the cost).
The big model is an opportunity for Chinese cloud computing companies, which has formed an industry consensus. But there are differences in how they seize opportunities—Baidu and Ali both develop their own large-scale models first, and then provide external services. Tencent has not yet released its self-developed model. Ma Huateng said, “There is no rush.” What Tencent does first is to build a platform to attract large-scale models from the industry to provide services to the outside world. (He Qianming)
Temu recruits compliance positions in the United States to deal with regulatory risks such as anti-money laundering
Recently, Temu posted recruitment information on LinkedIn. He wants to recruit compliance posts in Boston headquarters, mainly to respond to regulatory requirements related to financial operations such as anti-money laundering. Responsibilities include formulating and implementing compliance policies and procedures, training employees, and following up on policy changes. Timely response to regulatory agency requests and more.
Temu is also recruiting senior legal counsel in the United States, whose responsibilities cover product compliance, consumer privacy, intellectual property, trade compliance and other fields. In Hong Kong, Temu is looking for legal counsel with e-commerce or online payment regulatory compliance experience. There are other hiring needs scattered across the U.S. and Europe for accounting and tax positions.
The compliance difficulty for Chinese companies in cross-border e-commerce lies not only in the cross-border business itself, but also in meeting a set of legal and regulatory systems in developed countries in Europe and the United States, involving taxation, accounting standards, advertising laws, intellectual property rights, consumer privacy, and data compliance. regulations and so on. In terms of financial compliance, it is necessary to comply with a series of regulations such as anti-money laundering, such as conducting due diligence on customers, monitoring and reporting suspicious transactions, and so on.
For example, in Europe, from Google to Facebook to TikTok, they have all been punished by European regulators for data privacy issues. In the United States, anti-money laundering has always been one of the high-voltage lines of business operations. The local government has regulatory functions from the Ministry of Finance, the Ministry of Justice, the General Post Office, the General Administration of Customs, and the Drug Enforcement Administration.
According to China News Weekly, a seller on an independent website claimed that because the sales of goods rose from US$500 to US$10,000 the day before, the local regulatory authorities believed that he was suspected of money laundering, resulting in his account being blocked. In contrast, large platforms such as Temu and SHEIN are more difficult to identify and have higher risks because they involve more transactions and more intensive funds.
And as the volume grows, Temu and SHEIN are also subject to more and more regulatory pressures such as trade tariffs and supply chains. According to data from market agency Yipit, Temu’s total sales in the US, Canada, Australia, and New Zealand reached US$630 million in April this year. SHEIN had sales of US$22.7 billion last year.
Since April this year, U.S. government agencies have issued reports one after another, accusing Temu and SHEIN of alleged tax evasion, supply chain violations, and so on. In the middle of this month, some U.S. lawmakers proposed a bill to terminate the duty-free benefits for packages from China that are less than US$800.
In fact, SHEIN realized the challenges in this area earlier and has begun to fix the deficiencies in compliance. It appointed an ESG global director at the end of 2021. In the 2022 ESG report, it disclosed in detail the progress of issues such as supply chain labor rights and environmental impacts, with granularity as fine as “2.1% of employees’ wages are lower than the local minimum wage standard” and “upstream transportation greenhouse.” Gas emissions increased by 31% year-on-year” and so on. In April this year, SHEIN also recruited an anti-money laundering compliance officer in the United States, which is considered to be in preparation for the local launch of the platform model. (Lin Guangying)
How to grow lettuce in a container
On June 27, the third “Duoduo Agricultural Research Science and Technology Competition” came to an end. After growing strawberries and tomatoes in the greenhouse in the previous two sessions, the topic this time is to grow lettuce in a fully enclosed container. The purpose is to combine digital technology with agronomy to pursue better quality and higher yield with lower energy consumption.
According to reports, growing vegetables in container-type plant factories can maximize the isolation from the natural environment and make the growth factors of crops controllable. Here, the participating teams can match the spectrum by themselves, precisely control the temperature, light, water, humidity, fertilizer, etc., and observe the growth status of the crops under different parameters to continuously optimize the plan. To a large extent, agriculture is being developed in a standardized and industrialized manner.
The final winner was the Shanghai Academy of Agricultural Sciences team, which developed a planting decision-making system. After inputting data such as the environment and crops, it can output a series of planting suggestions, including how to adjust the light formula, water and fertilize, and pick in time. Members said that if the final planting plan is applied to a large-scale production plant factory, the production efficiency may be equivalent to the international advanced level of growing 400 grams of lettuce per square meter per day.
The “Cyber Farmers” team from China Agricultural University won the second place. They mentioned that when the conditions are right, the lettuce will grow quickly, but if it grows too fast, the lettuce is prone to heartburn and affects the quality, so in the end it was decided to sacrifice part of the yield and improve the quality to maximize economic benefits.
The “Endless Life” team from Shanghai Jiaotong University and the “Lettuce Growth” team from Polestar Agriculture tied for third place. The former has built an information platform to assist smart cultivation, and with the iteration of the platform, the planting plan can be quickly optimized. The latter integrates the “production-sales docking” link into the system design, by predicting the growth, arranging harvesting in advance, or receiving market demand, and customizing lettuce in turn.
It is not new to use algorithmic models to assist agricultural production. One of the major difficulties is that crops have a growth cycle, which makes it difficult to quickly iterate and optimize the planting plan. In this competition, 15 teams participated in the preliminary competition, and 4 teams entered the final. During this period, three crops of lettuce had to be planted, and the final crop was recorded in the results. This “horse racing” mechanism can be regarded as a way to improve efficiency and spread costs.
Cultivation costs are another hurdle. Taking plant factories as an example, the difficulty lies in the high cost. High-level plant factories consume 10 kilowatt-hours of electricity to produce 1 kg of lettuce, and the total cost may reach more than 20 yuan. However, the wholesale price of lettuce in Shanghai this month is only 3.6 yuan per kilogram. In contrast, strawberries with higher production value are more smoothly commercialized.
Agriculture is a slow industry, its short-term benefits are often overestimated, and its long-term benefits are easily underestimated. Referring to the global agricultural giant Monsanto, as early as 2012, it began to apply information technology to the agricultural field, and successively established agricultural data platforms and service platforms, such as using sensors and satellite images to collect field data, guiding farmers to plant, and optimizing models based on feedback data , so as to continuously save resources and increase production.
The Pinduoduo Agricultural Research Science and Technology Competition has been held for three times. The winning team “Zhiduoberry” in the first “Duoduo Agricultural Research Technology Competition” saw the market prospect of technology productization during the competition, and decided to establish the “Zhiduoberry” company. Help small and medium growers improve efficiency.
It is understood that “Zhiduoberry” has formed hardware, software, and algorithm products such as intelligent irrigation and intelligent greenhouse environmental control. As of the first quarter of this year, 40 sets of systems have been exported in Liaoning, Yunnan, Anhui, Inner Mongolia, Shanghai, Beijing and other places. It is mainly used to assist strawberry and blueberry production.
In Laowo Village, Nujiang Prefecture, Yunnan Province, which is a key support area for rural revitalization, “Zhiduoberry” has built a digital strawberry production system for the local area, which has reduced the common labor cost of the strawberry industry in Laowo Village by more than 30%, including reducing fertilizer expenditure by 2,500 yuan/mu, plant protection The expenditure was reduced by 1,000 yuan/mu, and the strawberry output increased by 30%, effectively helping local farmers increase their income and industrial development.
The dynamic adjustment technology of nutrient solution formula accumulated by the “Cyber Farmers” team in previous competitions has been written into a popular science paper, and all data has also been embedded in the model to form a standard algorithm, which is being applied at the Xiaotangshan Base in Beijing.
In the latest performance meeting, in addition to various agricultural research competitions, Pinduoduo also mentioned that it will cooperate with scientific research institutions to promote the transformation of achievements, help farmers connect with the storage and cold chain logistics system, and assist in building agricultural product brands, etc. (Lin Guangying)
Liu Qiangdong plans to sell $2.79 billion worth of JD shares
According to the SEC document (Form 144) submitted by JD.com, Liu Qiangdong, chairman of the group, intends to sell 800,000 ADSs (1 ADS = 2 ordinary shares) through the affiliated company Fortune Rising. According to the closing price of the day before the document was made public, this batch of shares Worth approximately $2.79 billion.
This is Liu Qiangdong’s seventh publicly recorded equity sale this year. A total of 44,066 JD ADSs were sold in the first 6 times, with a cumulative value of about 1.85 million US dollars.
“Wan Lian Cai” also learned that a group of Internet executives recently listed to sell their own company’s stock:
- Keike founder Zuo Hui’s family fund plans to sell ADS worth about 179 million US dollars;
- Hu Min, CFO of Tencent Music, intends to sell ADS worth about US$781,100;
- Guo Kongfeng, director of Donghai Group, intends to sell ADSs worth approximately US$22.79 million;
- Zhao Peng, the founder of BOSS, intends to sell ADS worth about 8.85 million US dollars;
- Liu Jun, vice president of finance of Pinduoduo, intends to sell ADSs worth approximately US$1.369 million.
A 62-Employee AI Startup Sold $1.3 Billion
This week, software company Databricks announced the acquisition of AI startup MosaicML for $1.3 billion.
MosaicML was founded in 2021 and now has only 62 employees. Its co-founder and CEO, Naveen Rao, launched Neon, the deep learning underlying framework that went out of the circle in 2014, and sold it to Intel for $400 million. He also joined Intel as the head of AI products.
Later, MosaicML was newly established to provide language models like OpenAI. The difference is that GPT parameters are larger, more versatile, and more expensive. MosaicML can also provide enterprises with dedicated, lower-cost and safer AI models trained with specific data. Its latest MPT-30B model is said to be comparable to GPT-3, and the training cost is only a fraction of GPT.
It’s a lot like Facebook’s acquisition of Instagram more than a decade ago. Instagram, which was established less than two years ago, had only 13 employees, and generated almost no revenue, was eventually sold to Facebook for $1 billion.
Instagram was founded in 2010 by Kevin Systrom and Mike Krieger. Two years later, Instagram had 25 million users, and Systrom turned down a $500 million to $700 million acquisition offer from Twitter in favor of $50 million in venture capital from Sequoia.
After that, Zuckerberg told the founder of Instagram directly on the phone that he wanted to buy the company at twice Instagram’s valuation at the time. After promising to give Instagram Facebook resources and operate independently, Zuckerberg successfully completed the acquisition.
The Instagram deal sets an important precedent for tech M&A by tech giants. Over the next few years, Twitter acquired Vine and Periscope in a similar fashion, Google acquired Nest, Amazon acquired Whole Foods, and Facebook itself went on to acquire WhatsApp and Oculus VR, all operating independently, at least initially. Guaranteed.
Now that these companies have matured, people once thought that there would be no more exponential growth opportunities like the mobile Internet. Until the emergence of ChatGPT and generative AI, new opportunities began to emerge again. A more extreme example than MosaicML is Midjourney, an AI drawing tool that was launched less than two years ago. It has only 11 full-time employees and has not raised funds so far. It earns $100 million a year from paid subscriptions.
Ten years later, we can still see the commonality of these entrepreneurs.
They are not as lucky as others, maybe they just saw potential opportunities earlier than most people, survived the unknown investment period, kept making products better and easier to use, and finally waited for the right opportunity.
In the beginning, Instagram was just a check-in tool, and the product design direction was changed several times. It was not until the release of the iPhone 4, which greatly improved the camera performance, that more people realized its unique value; Midjourney has updated 5 versions so far. The initial training model At that time, Holtz had no money and no results. He borrowed 10,000 GPUs from a cloud provider based on his personal reputation. It was not until ChatGPT became popular that Midjourney started to go out of the circle and continuously optimize and use it more conveniently. score. (Qiu Hao)
CHART OF THE DAY | Accounts receivable of several nucleic acid testing companies doubled
According to Jiemian News, in just six months, Daan Gene, once a leading listed company in nucleic acid testing for the new crown, has brought four third-party medical laboratories to court.
In the past three years, most of Daan Gene’s revenue came from the new crown nucleic acid testing. The company’s revenue increased from 1.098 billion yuan in 2019 to 12.046 billion yuan in 2022, and its net profit increased from 92 million yuan in 2019 to 5.412 billion yuan . However, its accounts receivable reached 4.066 billion yuan last year, accounting for more than one-third of its total revenue and nearly 30% of its total assets. In addition, Daan Gene made a total of 523 million yuan in bad debt provisions that year.
The situation of Daan Gene is not an isolated case. Listed nucleic acid testing companies generally face greater pressure on accounts receivable. According to incomplete statistics from Wandian Finance, as of the end of last year, there were 8 related listed companies whose accounts receivable exceeded 1 billion yuan. (Qiu Hao)
OTHER NEWS
In the first five months of the year, the decline in the profits of industrial enterprises across the country narrowed.
According to the National Bureau of Statistics, in the first five months of this year, the total profit of industrial enterprises across the country was about 2.67 trillion yuan, a year-on-year decrease of 18.8% (the previous value was -20.6%). Classified by the nature of enterprises, the profits of foreign and Hong Kong, Macao and Taiwan companies have the smallest decline, followed by state-owned holding companies, joint-stock companies, and private industrial companies.
Cainiao launched its own express delivery service.
Cainiao Group announced on the 28th the launch of its self-operated express delivery service, Cainiao Express. This business was originally a Tmall supermarket delivery business. Now it has been upgraded to a national express network, providing services such as half-day delivery, same-day delivery, door-to-door delivery, and overnight collection. Cainiao CEO Wan Lin said that he will create more cost-effective quality services. In May of this year, Ali announced that since “1+6+N”, Cainiao will take the lead in launching the listing plan.
The Ministry of Education will strictly investigate subject training and related online advertisements during the summer vacation.
According to the “China Education News”, the Ministry of Education will focus on rectifying discipline-related training in the name of “live-in teachers”, “high-end housekeeping” and research, summer camps, etc. during the summer vacation. At the same time, it will strictly investigate issues such as online training and advertising. The major platforms pay special attention to the content and advertisements related to education during the summer vacation. In the third quarter of last year, Douyin took down nearly 150,000 education and training advertisements selling educational anxiety, promises of results, and false propaganda.
Bancassurance institutions use corporate WeChat for risks, and were named by the State Administration of Financial Supervision.
The State Administration of Financial Supervision recently notified the risks of banking and insurance institutions using corporate WeChat services. According to the notification document, an agent provides corporate WeChat services for multiple banks, and saves chat content containing sensitive personal information of customers on a rented public cloud server. The agency trained the model with more than 6 million pieces of archived data and provided it to affiliates without informing the banks.
The stock prices of many overseas chip companies plummeted.
Today, Nvidia’s stock price fell more than 3% after hours, AMD fell 2.4%, and China’s domestic AI chip manufacturers Cambrian and Jingjiawei closed up 3.18% and 4.78% respectively. The market is mainly worried about the United States’ escalation of restrictions on China’s export of AI chips. In October last year, the U.S. government announced the export control of AI chips to China. For this purpose, Nvidia specially customized the GPU A800 and H800 for China with a slightly weaker performance. The new restrictions could make exports of the two chips to China also require a Commerce Department license.
Meituan launched the “Yuan Shen” cloud game portal, trying to make users stay longer.
Meituan recently launched the cloud game portal of “Yuan Shen” in the app, which is operated by the third-party platform “Haimawan”. In this way, besides reading novels to make money, watching short videos to make money and casual games, Meituan users have another reason to stay on the platform. According to QuestMobile data, the number of monthly active users of “Yuan Shen” reached 21.45 million in December last year. Games may be one of the most time-consuming forms of online entertainment. In order to attract more people to play games on Meituan, Meituan has recently promoted its game services on Douyin.
The entrance of the “Yuan Shen” cloud game in Meituan’s internal test, the contact email in the lower right corner indicates that it is operated by a third-party platform Haimawan. Image source: NGA
KeeTa, a subsidiary of Meituan, expanded its range of food delivery in Hong Kong, saying that user support exceeded expectations.
KeeTa has been operating in Tsim Sha Tsui, West Kowloon, Hung Hom and other areas since the 27th, and will enter Kowloon City, Kwun Tong and other places in July, and is expected to basically cover the urban areas of the Kowloon Peninsula. KeeTa was launched in Hong Kong on the 22nd of last month. Initially, it only operated on a small scale in Mong Kok and other places. According to the platform data, the average delivery time during the period was 23.91 minutes, and the on-time rate was 98%. The spokesperson said that in the past month, the enthusiastic support and recognition of local users has exceeded expectations, and more regions will be launched as soon as possible.
Nearly 90% of China’s 737 MAX fleet resumed flying.
According to Boeing China, as of the end of June this year, about 90% of China’s 737 MAX fleet has resumed commercial operations. Some planes have been put on regional international routes to Central Asia and Southeast Asia. After the two crashes of Indonesian Lion Air and Ethiopian Airlines from 2018 to 2019, the Civil Aviation Administration of China announced the grounding of Boeing 737 MAX models. On January 13 this year, Boeing 737 MAX aircraft resumed commercial operations in China. Since 2018, only the Greater Bay Area airlines among Chinese-funded airlines have purchased 15 737 MAXs in March this year.
New energy subsidies have been reduced, and NIO’s battery replacement model continues to receive policy support.
The Department of Taxation and Administration of the Ministry of Finance answered questions from reporters on the continuation and optimization of the new energy vehicle purchase tax reduction and exemption policy, and explained the policy in more detail. It is mentioned that new energy vehicles that meet the standard “battery swap mode” are allowed to use new energy vehicles without power batteries as the object of vehicle purchase tax, but they need to issue separate invoices for new energy vehicles and power batteries.
At present, NIO is the main new energy car company that adopts the battery swap model. After the price reduction this month, NIO’s ET5 and ES6 naked car prices (excluding tax) are 263,700 yuan and 299,100 yuan respectively, after deducting the power battery. , the vehicle purchase tax was further reduced.
The refinancing projects of four listed real estate companies obtained registration approval.
After China Merchants Shekou’s refinancing project was approved by the China Securities Regulatory Commission on the 16th of this month, four state-owned enterprises including Poly Development, China Communications Real Estate, Daming City and Fuxing Co., Ltd. also obtained the approval for refinancing project registration on June 27. Among them, Poly Development plans to raise a total of no more than 12.5 billion yuan, which is expected to be the largest among companies approved by the stock exchange so far. As of now, 17 other A-share real estate companies have disclosed their refinancing plans.
Hot weather boosted the sales of air conditioners.
According to industry online data, due to the high temperature weather, the domestic sales of Chinese air-conditioning manufacturers reached 12.636 million units in May this year, a year-on-year increase of nearly 30%, reaching a new high in five years. In the past two days, cities in Heilongjiang have appeared in the list of the hottest cities in the country for several days. According to the report of Aoweiyun.com, the omni-channel retail volume of air conditioners during the “618” period was 11.95 million units, a year-on-year increase of 35.9%. In addition, since February this year, sales of air conditioners have increased compared with previous years, which may be related to the partial recovery of the property market.
Silicon Valley startups are hot again, and 35% of companies admitted by YC are doing AI.
The latest phase of YC, a well-known Silicon Valley incubator, received 24,000 applications from companies, a record high, nearly 7,000 more than the same period last year. In the end, less than 1% were selected to join. Of these, 35 percent are AI-focused, and half involve AI. YC runs two incubation programs per year, training entrepreneurial teams and giving them a small investment. Airbnb, which is currently worth tens of billions of dollars, and Stripe, an online payment platform, are all incubated by YC.
Ford will lay off more than 1,000 employees in North America.
To offset the huge cost of investing in electric vehicles, Ford will lay off at least 1,000 contract and salaried workers in the United States and Canada. Ford CEO Jim Farley said that as the auto industry shifts to electric vehicles and digital services, Ford has too many employees and too few employees with the skills required for electric vehicles. Ford cut 173,000 employees worldwide after layoffs last year, following another 3,800 layoffs in Europe in February.
But layoffs are likely to be hindered, and the Ford union will most likely negotiate with management. Former Volkswagen CEO Herbert Diess considered laying off up to 30,000 employees in 2021. It was also because the company shifted to the electric vehicle business that required fewer parts, which triggered a strike by labor unions and finally announced his resignation.
It may be difficult to find users for the new Mac Pro, because the MacBook Pro is enough.
According to technology media The Verge, Apple’s new Mac Pro for professionals, starting at 55,999 yuan, may not be a good sell, because the cheaper and more portable MacBook Pro is enough to meet the needs of these people. The Verge interviewed more than 20 professionals who fit the Mac Pro audience, and only one person was interested in buying a Mac Pro. In addition, Apple’s new Mac Studio performance is also very powerful, and the price is only less than a third of the Mac Pro. Since it has just been released, there is no more detailed sales data yet.
This article is transferred from: https://www.latepost.com/news/dj_detail?id=1725
This site is only for collection, and the copyright belongs to the original author.