7*24

Hot topic

Market

Activities

On May 10, the latest poll from the Financial Times revealed that more than half of American voters disapprove of Trump's handling of the economy. Their main concerns include inflation, rising living costs, and the impact of the war in Iran on energy prices. Despite job reports showing new positions and a recovery in manufacturing, along with a rise in the stock market, the public's perception of actual living costs has caused Trump's economic approval rating to drop to around 37%, with disapproval exceeding 55%. The poll results highlight a disconnect between Wall Street and the general public's economic perceptions, which could affect the Republican Party's performance in the 2026 midterm elections.

[Over Half of American Voters Disapprove of Trump's Economic Policies]On May 10, the latest poll from the Financial Times revealed that more than half of American voters disapprove of Trump's handling of the economy. Their main concerns include inflation, rising living costs, and the impact of the war in Iran on energy prices. Despite job reports showing new positions and a recovery in manufacturing, along with a rise in the stock market, the public's perception of actual living costs has caused Trump's economic approval rating to drop to around 37%, with disapproval exceeding 55%. The poll results highlight a disconnect between Wall Street and the general public's economic perceptions, which could affect the Republican Party's performance in the 2026 midterm elections.

Over Half of American Voters Disapprove of Trump's Economic Policies

According to monitoring by Dongcha Beating, OpenAI has sent an official email announcing the complete shutdown of its self-service fine-tuning service for developers. As of now, new users can no longer create fine-tuning tasks; existing active users will only be able to use the service until January 6, 2027. As for the deployed fine-tuned models, their inference services will be tied to the lifecycle of the underlying base models, meaning that service will cease when the base model is retired. OpenAI's core reasoning is that the new generation of base models (such as GPT-5.5) is already extremely powerful in following instructions and formats. Compared to the expensive fine-tuning, directly writing prompts along with retrieval-augmented generation (RAG) is not only cheaper and faster but also sufficient to cover the vast majority of scenarios. Tech media Startup Fortune believes that this is not merely a product cleanup, but rather that OpenAI is shrinking its underlying model interfaces. In the past, startups often utilized low-cost self-service fine-tuning to create customized products for vertical industries. With this channel closing, developers will only be able to use generic prompt engineering, significantly increasing the difficulty for startup teams to establish customized technological barriers within the OpenAI ecosystem.

[OpenAI to Completely Shut Down Fine-Tuning API: Large Models Shift to Prompt-Based Approach, Startups Lose Customization Path]According to monitoring by Dongcha Beating, OpenAI has sent an official email announcing the complete shutdown of its self-service fine-tuning service for developers. As of now, new users can no longer create fine-tuning tasks; existing active users will only be able to use the service until January 6, 2027. As for the deployed fine-tuned models, their inference services will be tied to the lifecycle of the underlying base models, meaning that service will cease when the base model is retired. OpenAI's core reasoning is that the new generation of base models (such as GPT-5.5) is already extremely powerful in following instructions and formats. Compared to the expensive fine-tuning, directly writing prompts along with retrieval-augmented generation (RAG) is not only cheaper and faster but also sufficient to cover the vast majority of scenarios. Tech media Startup Fortune believes that this is not merely a product cleanup, but rather that OpenAI is shrinking its underlying model interfaces. In the past, startups often utilized low-cost self-service fine-tuning to create customized products for vertical industries. With this channel closing, developers will only be able to use generic prompt engineering, significantly increasing the difficulty for startup teams to establish customized technological barriers within the OpenAI ecosystem.

OpenAI to Completely Shut Down Fine-Tuning API: Large Models Shift to Prompt-Based Approach, Startups Lose Customization Path

On May 10, Korea's National Tax Service launched its first pilot program to entrust the management of seized virtual assets to private cryptocurrency custody institutions, with plans for a trial run until the end of this year. Major Korean custody institutions such as KODA, KDAC, Hecto WalletOne, BDACS, and InfiniteBlock are preparing to participate. Although the budget for this pilot project is only about $5,800, the industry views obtaining a reference case from the 'Korea National Tax Service project' as symbolically significant.

[Korea's National Tax Service Launches Pilot Program for Private Management of Seized Virtual Assets]On May 10, Korea's National Tax Service launched its first pilot program to entrust the management of seized virtual assets to private cryptocurrency custody institutions, with plans for a trial run until the end of this year. Major Korean custody institutions such as KODA, KDAC, Hecto WalletOne, BDACS, and InfiniteBlock are preparing to participate. Although the budget for this pilot project is only about $5,800, the industry views obtaining a reference case from the 'Korea National Tax Service project' as symbolically significant.

Korea's National Tax Service Launches Pilot Program for Private Management of Seized Virtual Assets

On May 10, according to monitoring by Onchain Lens, whale trader Loracle.hl has closed its positions in TON, BTC, and CL, realizing a profit of $3.9 million, while also closing approximately 96% of its long positions in ZEC. Meanwhile, this address has increased its short position in HYPE by five times, now valued at $53.23 million, bringing its total accumulated profits close to $37 million.

[Whale Trader Loracle.hl Switches to Short Positions, Accumulates Nearly $37 Million in Profits]On May 10, according to monitoring by Onchain Lens, whale trader Loracle.hl has closed its positions in TON, BTC, and CL, realizing a profit of $3.9 million, while also closing approximately 96% of its long positions in ZEC. Meanwhile, this address has increased its short position in HYPE by five times, now valued at $53.23 million, bringing its total accumulated profits close to $37 million.

Whale Trader Loracle.hl Switches to Short Positions, Accumulates Nearly $37 Million in Profits

On May 10, according to the Iranian Tasnim News Agency: A spokesperson for the Iranian military stated that countries adhering to US sanctions against Iran will "face difficulties in crossing the Strait of Hormuz."

[Iranian Military: Countries Following US Sanctions on Iran Will Face Difficulties in Crossing the Strait of Hormuz]On May 10, according to the Iranian Tasnim News Agency: A spokesperson for the Iranian military stated that countries adhering to US sanctions against Iran will "face difficulties in crossing the Strait of Hormuz."

Iranian Military: Countries Following US Sanctions on Iran Will Face Difficulties in Crossing the Strait of Hormuz

According to Dongcha Beating monitoring, Microsoft has open-sourced the Phi-Ground model family, specifically designed to address the issue of 'where to click on the screen' when AI operates a computer. Given a screenshot and a command, the model outputs precise click coordinates. The open-sourced 4 billion parameter version, when combined with large models for instruction planning, achieved a click accuracy rate exceeding that of OpenAI's Operator and Claude Computer Use in the Showdown benchmark test, and secured first place in all five evaluations below 10 billion parameters, including ScreenSpot-Pro. The team conducted large-scale validation using over 40 million data points and found that the three types of training techniques commonly used in previous academic papers became ineffective as the data volume increased. The truly effective approach is straightforward: output coordinates as ordinary numbers, such as '523, 417'. Previously, several papers invented a specialized vocabulary for coordinates, hoping the model would articulate coordinates like words, but during large-scale training, these new terms could not be learned effectively, leading to model failures. Another key factor is placing text instructions before images in the input. Large models read information unidirectionally; when they first read 'click the blue settings icon' and then look at the image, they already know what to look for when processing pixels. Conversely, if they look at the image first, the model can only scan blindly, resulting in significantly poorer performance. The team also discovered that reinforcement learning is beneficial for purely visual tasks. The specific method involves having the model make multiple click predictions on the same image, comparing correct and incorrect results for training (this method is called DPO, a type of reinforcement learning). Even after the model has been thoroughly fine-tuned, this step can still significantly improve accuracy. Previously, reinforcement learning was typically applied only to language tasks requiring reasoning, but its effectiveness in purely 'pointing and clicking' perceptual tasks was an unexpected gain. To address the issue of buttons being too small on 4K high-resolution screens (a button may occupy only 0.07% of the screen area), the team scaled down screenshots during training and placed them on a large white canvas to simulate the real scenario of extremely small elements on high-resolution screens. This approach proved particularly effective in complex professional software like Photoshop.

[Microsoft Open Sources Phi-Ground: 4 Billion Parameters Achieve Higher Click Accuracy than Operator and Claude]According to Dongcha Beating monitoring, Microsoft has open-sourced the Phi-Ground model family, specifically designed to address the issue of 'where to click on the screen' when AI operates a computer. Given a screenshot and a command, the model outputs precise click coordinates. The open-sourced 4 billion parameter version, when combined with large models for instruction planning, achieved a click accuracy rate exceeding that of OpenAI's Operator and Claude Computer Use in the Showdown benchmark test, and secured first place in all five evaluations below 10 billion parameters, including ScreenSpot-Pro. The team conducted large-scale validation using over 40 million data points and found that the three types of training techniques commonly used in previous academic papers became ineffective as the data volume increased. The truly effective approach is straightforward: output coordinates as ordinary numbers, such as '523, 417'. Previously, several papers invented a specialized vocabulary for coordinates, hoping the model would articulate coordinates like words, but during large-scale training, these new terms could not be learned effectively, leading to model failures. Another key factor is placing text instructions before images in the input. Large models read information unidirectionally; when they first read 'click the blue settings icon' and then look at the image, they already know what to look for when processing pixels. Conversely, if they look at the image first, the model can only scan blindly, resulting in significantly poorer performance. The team also discovered that reinforcement learning is beneficial for purely visual tasks. The specific method involves having the model make multiple click predictions on the same image, comparing correct and incorrect results for training (this method is called DPO, a type of reinforcement learning). Even after the model has been thoroughly fine-tuned, this step can still significantly improve accuracy. Previously, reinforcement learning was typically applied only to language tasks requiring reasoning, but its effectiveness in purely 'pointing and clicking' perceptual tasks was an unexpected gain. To address the issue of buttons being too small on 4K high-resolution screens (a button may occupy only 0.07% of the screen area), the team scaled down screenshots during training and placed them on a large white canvas to simulate the real scenario of extremely small elements on high-resolution screens. This approach proved particularly effective in complex professional software like Photoshop.

Microsoft Open Sources Phi-Ground: 4 Billion Parameters Achieve Higher Click Accuracy than Operator and Claude

Bitcoin Fear and Greed Index is 47. Neutral.
Current Price: $80696.43.

[Bitcoin Fear and Greed Index is 47]Bitcoin Fear and Greed Index is 47. Neutral.
Current Price: $80696.43.

Bitcoin Fear and Greed Index is 47

HTX data monitoring shows that UNI has broken through 4 USDT, currently reported at 4.01 USDT, with a 24H increase of 7.00%.

【UNI Breakthrough 4 USDT】HTX data monitoring shows that UNI has broken through 4 USDT, currently reported at 4.01 USDT, with a 24H increase of 7.00%.

UNI Breakthrough 4 USDT

On May 10, The Wall Street Journal reported on the 9th, citing U.S. officials and informed sources, that Israel has established a secret military base in the Iraqi desert for airstrikes against Iran. It was reported that in the early stages of the conflict with Iran, Israel launched airstrikes against Iraqi forces that 'almost discovered' the base. According to sources, the base was established shortly before the U.S. and Israel took military action against Iran this year, and the U.S. was 'aware' of it. The base also serves as a logistical hub for the Israeli Air Force. Reports indicate that Israel has deployed rescue teams at the base. Additionally, informed sources stated that Israeli Air Force special forces capable of conducting assault missions are also stationed at the base.

[US Media: Israel Establishes Secret Base in Iraqi Desert Targeting Iran]On May 10, The Wall Street Journal reported on the 9th, citing U.S. officials and informed sources, that Israel has established a secret military base in the Iraqi desert for airstrikes against Iran. It was reported that in the early stages of the conflict with Iran, Israel launched airstrikes against Iraqi forces that 'almost discovered' the base. According to sources, the base was established shortly before the U.S. and Israel took military action against Iran this year, and the U.S. was 'aware' of it. The base also serves as a logistical hub for the Israeli Air Force. Reports indicate that Israel has deployed rescue teams at the base. Additionally, informed sources stated that Israeli Air Force special forces capable of conducting assault missions are also stationed at the base.

US Media: Israel Establishes Secret Base in Iraqi Desert Targeting Iran

Coinglass data shows that the past 1 hour the whole network exploded $298 million, the past 24 hours the whole network exploded $423 million.
In the past 24 hours, ETH exploded for $177 million and BTC exploded for $99,568,900.

Detail

【The past 1 hour the whole network exploded $298 million】Coinglass data shows that the past 1 hour the whole network exploded $298 million, the past 24 hours the whole network exploded $423 million.
In the past 24 hours, ETH exploded for $177 million and BTC exploded for $99,568,900.

The past 1 hour the whole network exploded $298 million

对话Dragonfly合伙人：散户离场机构兜底，今年加密货币波动性难超美股

XRP Power Play: SBI Holdings Submits LOI To Acquire Bitbank In Major Asia Move

一周代币解锁：ENA解锁价值约1700万美元代币

This Week In Bitcoin: Top Developments That Could Signal A New Era

Bitcoin Prints Biggest Monthly Win In A Year Amid Renewed Market Optimism

十亿美元的教训：DeFi 安全重心正从代码转向运营治理

Anthropic搞了个全是AI的闲鱼群，大模型在里面互割起了韭菜

对话华尔街策略师：AI驱动通缩，加速资本流向稀缺性资产

特朗普一条推文，BTC时隔3个月重返8万美元

Article

Indian Vessel Catches Fire and Sinks Near Strait of Hormuz, Resulting in 1 Death and 4 Injuries

The past 1 hour the whole network exploded $298 million

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

Related Articles

All Comments0LatestHot

Trending Topics

Related Articles