7*24

Hot topic

Market

Activities

According to monitoring by Dongcha Beating, MiniMax has upgraded its desktop Agent product and renamed it to Mavis (MiniMax as a Jarvis). The core new capability is Agent Teams: users can create teams composed of multiple Agents with different roles to collaboratively complete complex long tasks that a single Agent would struggle with. Additionally, the previously separate API and Agent subscriptions have been merged into one, allowing seamless integration of CLI, API, and Agent with shared quotas. MiniMax has also released a technical article explaining the design concept of Agent Teams, which was generated by an Agent Team itself—one Agent simulates user questions while another answers based on internal technical documentation. The article points out four core issues with a single Agent handling long tasks: unexpected interruptions during task execution, decreased output quality due to lengthy context, long tasks blocking user interactions, and the inability to achieve true role division in prompt-based role-playing. To address these issues, Mavis employs a code state machine instead of prompt orchestration to drive collaboration. The team has three roles: Owner responsible for task breakdown and scheduling, Worker focused on execution, and Verifier for independent acceptance. The Verifier and Worker form an adversarial mechanism, with strict isolation of context among roles, communicating only through structured summaries. MiniMax acknowledges in the article that multi-Agent collaboration introduces additional handover and aggregation costs, but for long, high-risk tasks, this structured overhead results in greater certainty in delivery outcomes.

[MiniMax Desktop Renamed to Mavis, Launches Multi-Agent Team Collaboration]According to monitoring by Dongcha Beating, MiniMax has upgraded its desktop Agent product and renamed it to Mavis (MiniMax as a Jarvis). The core new capability is Agent Teams: users can create teams composed of multiple Agents with different roles to collaboratively complete complex long tasks that a single Agent would struggle with. Additionally, the previously separate API and Agent subscriptions have been merged into one, allowing seamless integration of CLI, API, and Agent with shared quotas. MiniMax has also released a technical article explaining the design concept of Agent Teams, which was generated by an Agent Team itself—one Agent simulates user questions while another answers based on internal technical documentation. The article points out four core issues with a single Agent handling long tasks: unexpected interruptions during task execution, decreased output quality due to lengthy context, long tasks blocking user interactions, and the inability to achieve true role division in prompt-based role-playing. To address these issues, Mavis employs a code state machine instead of prompt orchestration to drive collaboration. The team has three roles: Owner responsible for task breakdown and scheduling, Worker focused on execution, and Verifier for independent acceptance. The Verifier and Worker form an adversarial mechanism, with strict isolation of context among roles, communicating only through structured summaries. MiniMax acknowledges in the article that multi-Agent collaboration introduces additional handover and aggregation costs, but for long, high-risk tasks, this structured overhead results in greater certainty in delivery outcomes.

MiniMax Desktop Renamed to Mavis, Launches Multi-Agent Team Collaboration

On May 13, the Bank of Japan announced that Governor Kazuo Ueda will deliver a speech at the monetary policy meeting hosted by the Bank of Japan on May 27. (Tokyo Shimbun)

[Bank of Japan Governor to Speak at Monetary Policy Meeting on May 27]On May 13, the Bank of Japan announced that Governor Kazuo Ueda will deliver a speech at the monetary policy meeting hosted by the Bank of Japan on May 27. (Tokyo Shimbun)

Bank of Japan Governor to Speak at Monetary Policy Meeting on May 27

According to monitoring by Dongcha Beating, NVIDIA's proxy statement submitted to the SEC for fiscal year 2026 shows that CEO Jensen Huang's total compensation is $36.3 million, a decrease of approximately $13.6 million from $49.9 million in the previous fiscal year, representing a 27% drop. The reduction in compensation is almost entirely due to the decline in the book value of equity awards. For fiscal year 2026, the equity awards are calculated at a fair value of $24.8 million on the grant date, down 36% from $38.8 million in the previous fiscal year, accounting for over 90% of the total reduction. The fixed salary portion remained largely unchanged: the base salary is $1.5 million, non-equity incentives are $6 million, and approximately $4 million is allocated for residential security, private travel safety, and driver expenses. However, the book value and actual cash received are two different matters. The equity awards recorded in the SEC proxy statement are accounted for at 'fair value on the grant date,' reflecting the valuation at the moment of issuance, not the final cash received. According to reports from South Korea's Daily Economy, Huang actually cashed out 1,113,555 performance stock units (PSUs) accumulated over the past few years this fiscal year, amounting to approximately $141 million based on the stock price at the time of cashing out. This means that although the newly issued awards have decreased in book value, the old stocks accumulated during the previous surge in stock prices are being liquidated on a large scale. This compensation structure is representative among NVIDIA executives: equity constitutes the vast majority, while cash is a very small proportion. When stock prices rise, actual earnings far exceed book compensation; when stock prices stabilize, the book numbers automatically adjust, but the equity earnings accumulated over the past few years have already been realized. NVIDIA's revenue for fiscal year 2026 increased by 65% year-on-year, and as of May 12, its market capitalization is approximately $5.36 trillion, still ranking first globally.

[NVIDIA CEO Jensen Huang's Salary Decreases by 27%, Cashes Out $141 Million in Stock This Year]According to monitoring by Dongcha Beating, NVIDIA's proxy statement submitted to the SEC for fiscal year 2026 shows that CEO Jensen Huang's total compensation is $36.3 million, a decrease of approximately $13.6 million from $49.9 million in the previous fiscal year, representing a 27% drop. The reduction in compensation is almost entirely due to the decline in the book value of equity awards. For fiscal year 2026, the equity awards are calculated at a fair value of $24.8 million on the grant date, down 36% from $38.8 million in the previous fiscal year, accounting for over 90% of the total reduction. The fixed salary portion remained largely unchanged: the base salary is $1.5 million, non-equity incentives are $6 million, and approximately $4 million is allocated for residential security, private travel safety, and driver expenses. However, the book value and actual cash received are two different matters. The equity awards recorded in the SEC proxy statement are accounted for at 'fair value on the grant date,' reflecting the valuation at the moment of issuance, not the final cash received. According to reports from South Korea's Daily Economy, Huang actually cashed out 1,113,555 performance stock units (PSUs) accumulated over the past few years this fiscal year, amounting to approximately $141 million based on the stock price at the time of cashing out. This means that although the newly issued awards have decreased in book value, the old stocks accumulated during the previous surge in stock prices are being liquidated on a large scale. This compensation structure is representative among NVIDIA executives: equity constitutes the vast majority, while cash is a very small proportion. When stock prices rise, actual earnings far exceed book compensation; when stock prices stabilize, the book numbers automatically adjust, but the equity earnings accumulated over the past few years have already been realized. NVIDIA's revenue for fiscal year 2026 increased by 65% year-on-year, and as of May 12, its market capitalization is approximately $5.36 trillion, still ranking first globally.

NVIDIA CEO Jensen Huang's Salary Decreases by 27%, Cashes Out $141 Million in Stock This Year

On May 13, South Korean President Lee Jae-myung stated that a recent post by his advisor regarding the redistribution of profits generated from AI prosperity sparked widespread discussion, with the intention of initiating a broader public debate on how to utilize potential unexpected tax wealth. Lee pointed out that the remarks made by his policy chief, Kim Yong-bum, essentially aimed to 'explore the feasibility of a plan to distribute the national surplus tax revenue generated from excessive profits in the AI sector as a 'citizen dividend' to the general public.' He emphasized that these comments never suggested using corporate profits to fund such dividends. Kim's remarks caused significant turbulence in the South Korean stock market, leaving investors confused about the specific scope of his proposal. The KOSPI index plummeted by as much as 5.1%, but the decline began to narrow after a prominent policy advisor clarified that the intention was to utilize 'excess tax revenue' generated from the AI boom, rather than imposing a new windfall tax on corporate profits. Previously, an official from the South Korean presidential office stated that Kim's comments only represented his personal views and were not a topic of formal discussion. (Dong-A Ilbo)

[South Korean President Clarifies 'Universal Dividend' Concept: No Mention of Using Corporate Profits]On May 13, South Korean President Lee Jae-myung stated that a recent post by his advisor regarding the redistribution of profits generated from AI prosperity sparked widespread discussion, with the intention of initiating a broader public debate on how to utilize potential unexpected tax wealth. Lee pointed out that the remarks made by his policy chief, Kim Yong-bum, essentially aimed to 'explore the feasibility of a plan to distribute the national surplus tax revenue generated from excessive profits in the AI sector as a 'citizen dividend' to the general public.' He emphasized that these comments never suggested using corporate profits to fund such dividends. Kim's remarks caused significant turbulence in the South Korean stock market, leaving investors confused about the specific scope of his proposal. The KOSPI index plummeted by as much as 5.1%, but the decline began to narrow after a prominent policy advisor clarified that the intention was to utilize 'excess tax revenue' generated from the AI boom, rather than imposing a new windfall tax on corporate profits. Previously, an official from the South Korean presidential office stated that Kim's comments only represented his personal views and were not a topic of formal discussion. (Dong-A Ilbo)

South Korean President Clarifies 'Universal Dividend' Concept: No Mention of Using Corporate Profits

According to market sources, Anthropic is in advanced negotiations to acquire the developer tools startup Stainless for at least $300 million. The company is utilized by OpenAI and Google. (The Information)

[Anthropic Plans to Acquire Developer Tools Startup Stainless for At Least $300 Million]According to market sources, Anthropic is in advanced negotiations to acquire the developer tools startup Stainless for at least $300 million. The company is utilized by OpenAI and Google. (The Information)

Anthropic Plans to Acquire Developer Tools Startup Stainless for At Least $300 Million

According to monitoring by Dongcha Beating, the MIT team led by He Kaiming has released the language diffusion model ELF (Embedded Language Flows). Unlike the GPT-style autoregressive approach of 'predicting the next token', it integrates text generation into a continuous embedding space until the final step, where it converts back to discrete tokens. While diffusion models have matured in image generation, applying them to text has been challenging: images are inherently continuous signals, whereas language consists of discrete tokens. Many previous continuous diffusion text models either repeatedly introduced token-level supervision during the generation trajectory or required an additional independent decoder. ELF's approach is cleaner: most steps involve denoising in the continuous vector space, with the final step using a shared weight network for discretization. The experimental results are also striking. In the unconditional generation evaluation on OpenWebText, the 105M parameter ELF-B achieved approximately 24.1 Gen. PPL with 32 sampling steps, outperforming various discrete and continuous diffusion language model baselines. More importantly, ELF-B used only about 45B training tokens, while comparative methods typically exceed 500B, indicating a reduction of nearly an order of magnitude in training tokens. This result suggests that the continuous diffusion approach in language modeling is not hindered by 'language discreteness'; the previous issues are more likely related to modeling interfaces and sampling designs.

[MIT Team Releases ELF: Language Diffusion Model Successfully Implemented]According to monitoring by Dongcha Beating, the MIT team led by He Kaiming has released the language diffusion model ELF (Embedded Language Flows). Unlike the GPT-style autoregressive approach of 'predicting the next token', it integrates text generation into a continuous embedding space until the final step, where it converts back to discrete tokens. While diffusion models have matured in image generation, applying them to text has been challenging: images are inherently continuous signals, whereas language consists of discrete tokens. Many previous continuous diffusion text models either repeatedly introduced token-level supervision during the generation trajectory or required an additional independent decoder. ELF's approach is cleaner: most steps involve denoising in the continuous vector space, with the final step using a shared weight network for discretization. The experimental results are also striking. In the unconditional generation evaluation on OpenWebText, the 105M parameter ELF-B achieved approximately 24.1 Gen. PPL with 32 sampling steps, outperforming various discrete and continuous diffusion language model baselines. More importantly, ELF-B used only about 45B training tokens, while comparative methods typically exceed 500B, indicating a reduction of nearly an order of magnitude in training tokens. This result suggests that the continuous diffusion approach in language modeling is not hindered by 'language discreteness'; the previous issues are more likely related to modeling interfaces and sampling designs.

MIT Team Releases ELF: Language Diffusion Model Successfully Implemented

On May 13, France's unemployment rate unexpectedly rose to its highest level in five years, further indicating that the economic foundation of the eurozone's second-largest economy was already fragile before the outbreak of the war in Iran. Data from the National Institute of Statistics and Economic Studies (Insee) shows that the unemployment rate reached 8.1% as unemployment rose across all age groups. Economists had previously predicted a slight decrease in the unemployment rate from 7.9% at the end of last year to 7.8%.

[France's Unemployment Rate Exceeds 8% for the First Time in Five Years]On May 13, France's unemployment rate unexpectedly rose to its highest level in five years, further indicating that the economic foundation of the eurozone's second-largest economy was already fragile before the outbreak of the war in Iran. Data from the National Institute of Statistics and Economic Studies (Insee) shows that the unemployment rate reached 8.1% as unemployment rose across all age groups. Economists had previously predicted a slight decrease in the unemployment rate from 7.9% at the end of last year to 7.8%.

France's Unemployment Rate Exceeds 8% for the First Time in Five Years

On May 13, according to monitoring by Onchain Lens, the whale Loracle.hl (@loraclexyz) has opened a 156 BTC short position with 20x leverage, valued at $12.66 million. The 10x SNDK short position opened two days ago currently has a floating profit of over $648,000, while the 5x HYPE short position has a floating profit exceeding $2.3 million. The total floating profit currently exceeds $41 million.

[Whale Opens $12.66 Million BTC Short Position, Total Floating Profit Exceeds $41 Million]On May 13, according to monitoring by Onchain Lens, the whale Loracle.hl (@loraclexyz) has opened a 156 BTC short position with 20x leverage, valued at $12.66 million. The 10x SNDK short position opened two days ago currently has a floating profit of over $648,000, while the 5x HYPE short position has a floating profit exceeding $2.3 million. The total floating profit currently exceeds $41 million.

Whale Opens $12.66 Million BTC Short Position, Total Floating Profit Exceeds $41 Million

On May 13, according to Guoshi Zhitongche, Trump boarded Air Force One on the afternoon of May 12 Eastern Time, beginning his visit to China. Before landing, Trump posted on 'Truth Social' stating, 'It is my honor to visit great China with many distinguished individuals.' Trump also expressed his hope that the accompanying business leaders would be able to 'make a significant impact' and help China reach a higher level of development! (Dongxin News Agency)

[Trump's Aerial Announcement: It Is My Honor to Visit Great China with Many Distinguished Individuals]On May 13, according to Guoshi Zhitongche, Trump boarded Air Force One on the afternoon of May 12 Eastern Time, beginning his visit to China. Before landing, Trump posted on 'Truth Social' stating, 'It is my honor to visit great China with many distinguished individuals.' Trump also expressed his hope that the accompanying business leaders would be able to 'make a significant impact' and help China reach a higher level of development! (Dongxin News Agency)

Trump's Aerial Announcement: It Is My Honor to Visit Great China with Many Distinguished Individuals

On May 12, Michael Selig, chairman of the U.S. Commodity Futures Trading Commission (CFTC), stated that the regulatory agency is in discussions with all major U.S. professional sports leagues to strengthen the regulatory framework for sports-related predictive markets and to prevent insider trading and market manipulation risks. Selig revealed at the FINRA annual meeting in Washington that the CFTC had previously signed a memorandum of understanding for information sharing with Major League Baseball, marking its first formal data collaboration mechanism with a professional sports league. This initiative comes as predictive market platforms, such as Kalshi and Polymarket, continue to expand their sports contract offerings, raising legal disputes between state-level gambling regulators and federal regulatory authority.

[CFTC Engages with Major Sports Leagues on Predictive Market Regulation Cooperation]On May 12, Michael Selig, chairman of the U.S. Commodity Futures Trading Commission (CFTC), stated that the regulatory agency is in discussions with all major U.S. professional sports leagues to strengthen the regulatory framework for sports-related predictive markets and to prevent insider trading and market manipulation risks. Selig revealed at the FINRA annual meeting in Washington that the CFTC had previously signed a memorandum of understanding for information sharing with Major League Baseball, marking its first formal data collaboration mechanism with a professional sports league. This initiative comes as predictive market platforms, such as Kalshi and Polymarket, continue to expand their sports contract offerings, raising legal disputes between state-level gambling regulators and federal regulatory authority.

CFTC Engages with Major Sports Leagues on Predictive Market Regulation Cooperation

史上最严重的存储短缺，把三星推上全球第二

租来的轨道：这一波稳定币 FX 热钱，到底在为什么买单？

MegaETH 如何让 TVL 在 TGE 一周内达到 700m？拆解包装手法

South Korea’s First Bank-Led Crypto Picks A Quantum Security Partner — Here’s Why It Matters

Bulls Approach the Ceiling

The Bitcoin Vector #54

Roobet Launches Prediction Markets on May 6, The First Major Crypto Casino to Integrate the Format

Pre-ETF Era Bitcoin Whales Surface To Cash In On $80,000 Rally

S&P Call Options Explode to Record $2.6 Trillion Volume: The Crash Risks Behind a Frenzied Gamble

Article

OpenAI Caps Microsoft Revenue Share at $38 Billion, Saving $97 Billion Long-Term but Pressuring Short-Term Cash Flow

CFTC Engages with Major Sports Leagues on Predictive Market Regulation Cooperation

Disclaimer：The content above does not represent HTX's positions.，HTX does not provide any trading recommendations.。

All Comments0LatestHot

Trending Topics