7*24

Hot topic

Market

Activities

According to PolyBeats monitoring, on the prediction market Polymarket, a new address has invested $48,000 in betting 'No' on the event 'Will Trump visit China before May'. The current probability for 'No' is 62%. Trump's originally scheduled trip to Beijing from March 31st to April 2nd has been postponed. Trump himself stated that his new Asia trip would be in about five to six weeks. If calculated strictly based on this statement, it would fall around April 21st to April 28th, which highly overlaps with the unbookable period at the China World Hotel in Beijing from April 22nd to April 26th. However, as of now, the White House has stated that both sides are still negotiating a new date; the Chinese Embassy in the US maintains that they 'will continue to communicate regarding Trump's visit to China'. Account: 0x126091a58494557662a8285c35f3f4b5e4f107ce Total Investment: $48k --------------------------------- See the future earlier, follow @PolyBeats_Bot

[New Address Bets $48K Trump Won't Visit China Before May]According to PolyBeats monitoring, on the prediction market Polymarket, a new address has invested $48,000 in betting 'No' on the event 'Will Trump visit China before May'. The current probability for 'No' is 62%. Trump's originally scheduled trip to Beijing from March 31st to April 2nd has been postponed. Trump himself stated that his new Asia trip would be in about five to six weeks. If calculated strictly based on this statement, it would fall around April 21st to April 28th, which highly overlaps with the unbookable period at the China World Hotel in Beijing from April 22nd to April 26th. However, as of now, the White House has stated that both sides are still negotiating a new date; the Chinese Embassy in the US maintains that they 'will continue to communicate regarding Trump's visit to China'. Account: 0x126091a58494557662a8285c35f3f4b5e4f107ce Total Investment: $48k --------------------------------- See the future earlier, follow @PolyBeats_Bot

New Address Bets $48K Trump Won't Visit China Before May

According to 1M AI News monitoring, Chen Cheng, an engineer at Ant Group and author of the front-end framework Umi.js, has reverse-engineered the source code of Claude Code 2.1.81, fully reconstructing the events that occur after the 'auto mode' is activated. Key findings reveal a four-layer decision pipeline for every tool invocation. Only when the first three layers cannot make a determination is an independent AI classifier invoked for a safety review. The four layers are as follows: The first layer checks existing permission rules; if a match is found, it's directly approved. The second layer simulates the 'acceptEdits' mode (permission level allowing file edits); if it passes under this mode, indicating low risk, it bypasses the classifier. The third layer checks a whitelist of read-only tools (Read, Grep, Glob, LSP, WebSearch, etc.) which do not modify any state and are unconditionally approved. If none of the above conditions are met, the process enters the fourth layer, making an independent API request to Claude Sonnet for safety classification. Key design details of the classifier include: always using Sonnet instead of Opus as a trade-off between cost and latency; setting the temperature to 0 to ensure deterministic output; defining the classifier as a 'safety monitor for autonomous AI programming agents' to protect against three types of risks (prompt injection, scope creep, and unintended damage); and injecting the user's CLAUDE.md configuration file into the classifier's context as a basis for judging user intent. Interception rules cover over 22 categories, including force push, direct push to the main branch, downloading and executing external code, production deployment, data exfiltration, self-modification of permissions, creation of remote code execution surfaces, and credential leakage. Exceptions for approval include 7 categories: hardcoded test keys, local file operations within the working directory, read-only GET requests, installation of declared dependencies, official toolchain installation, reading configuration credentials to send to the target provider, and pushing to the current working branch. There is also a circuit breaker mechanism: after 3 consecutive rejections or a cumulative 20 rejections, the system degrades to manual confirmation; in headless mode, the agent is directly aborted. When the classifier is unavailable, a feature flag controls whether it 'fails closed' (rejects directly) or 'fails open' (degrades to manual confirmation). Prompt injection in 'auto mode' also has fine-grained frequency control: injected once every 5 conversation rounds. The first injection in every 5-injection cycle is the full version (approximately 800 words, including six instructions such as 'execute immediately, minimize interruptions, action over planning'), while the other 4 are single-line condensed versions, balancing context window usage with behavioral stability.

[Ant Engineer Reverse-Engineers Claude Code's Auto Mode: Four-Layer Decision Pipeline and Safety Classifier Revealed]According to 1M AI News monitoring, Chen Cheng, an engineer at Ant Group and author of the front-end framework Umi.js, has reverse-engineered the source code of Claude Code 2.1.81, fully reconstructing the events that occur after the 'auto mode' is activated. Key findings reveal a four-layer decision pipeline for every tool invocation. Only when the first three layers cannot make a determination is an independent AI classifier invoked for a safety review. The four layers are as follows: The first layer checks existing permission rules; if a match is found, it's directly approved. The second layer simulates the 'acceptEdits' mode (permission level allowing file edits); if it passes under this mode, indicating low risk, it bypasses the classifier. The third layer checks a whitelist of read-only tools (Read, Grep, Glob, LSP, WebSearch, etc.) which do not modify any state and are unconditionally approved. If none of the above conditions are met, the process enters the fourth layer, making an independent API request to Claude Sonnet for safety classification. Key design details of the classifier include: always using Sonnet instead of Opus as a trade-off between cost and latency; setting the temperature to 0 to ensure deterministic output; defining the classifier as a 'safety monitor for autonomous AI programming agents' to protect against three types of risks (prompt injection, scope creep, and unintended damage); and injecting the user's CLAUDE.md configuration file into the classifier's context as a basis for judging user intent. Interception rules cover over 22 categories, including force push, direct push to the main branch, downloading and executing external code, production deployment, data exfiltration, self-modification of permissions, creation of remote code execution surfaces, and credential leakage. Exceptions for approval include 7 categories: hardcoded test keys, local file operations within the working directory, read-only GET requests, installation of declared dependencies, official toolchain installation, reading configuration credentials to send to the target provider, and pushing to the current working branch. There is also a circuit breaker mechanism: after 3 consecutive rejections or a cumulative 20 rejections, the system degrades to manual confirmation; in headless mode, the agent is directly aborted. When the classifier is unavailable, a feature flag controls whether it 'fails closed' (rejects directly) or 'fails open' (degrades to manual confirmation). Prompt injection in 'auto mode' also has fine-grained frequency control: injected once every 5 conversation rounds. The first injection in every 5-injection cycle is the full version (approximately 800 words, including six instructions such as 'execute immediately, minimize interruptions, action over planning'), while the other 4 are single-line condensed versions, balancing context window usage with behavioral stability.

Ant Engineer Reverse-Engineers Claude Code's Auto Mode: Four-Layer Decision Pipeline and Safety Classifier Revealed

Market data shows that BTC has broken through $72,000, currently trading at $72,007.52. It has seen a 24-hour increase of 1.24%. The market is experiencing significant volatility, so please implement risk control measures.

[BTC Surpasses $72,000]Market data shows that BTC has broken through $72,000, currently trading at $72,007.52. It has seen a 24-hour increase of 1.24%. The market is experiencing significant volatility, so please implement risk control measures.

BTC Surpasses $72,000

On March 25th, veteran investor Simon Dixon publicly criticized Michael Saylor's strategy of aggressively accumulating Bitcoin, stating that it could exacerbate the centralization risks of the cryptocurrency. Currently, Strategy holds approximately 762,000 BTC, representing over 3% of the total supply. The company recently launched a $42 billion ATM financing plan, intending to further increase its Bitcoin holdings. Dixon pointed out that this model deeply intertwines Bitcoin with the traditional financial system, making it more susceptible to short-term trading and price manipulation, thus deviating from its original decentralized ethos. He advocates for a return to self-custody and long-term holding strategies, rather than centralized allocation through publicly traded companies.

[Veteran Investor Criticizes Michael Saylor's Bitcoin Accumulation Strategy Amid Centralization Concerns]On March 25th, veteran investor Simon Dixon publicly criticized Michael Saylor's strategy of aggressively accumulating Bitcoin, stating that it could exacerbate the centralization risks of the cryptocurrency. Currently, Strategy holds approximately 762,000 BTC, representing over 3% of the total supply. The company recently launched a $42 billion ATM financing plan, intending to further increase its Bitcoin holdings. Dixon pointed out that this model deeply intertwines Bitcoin with the traditional financial system, making it more susceptible to short-term trading and price manipulation, thus deviating from its original decentralized ethos. He advocates for a return to self-custody and long-term holding strategies, rather than centralized allocation through publicly traded companies.

Veteran Investor Criticizes Michael Saylor's Bitcoin Accumulation Strategy Amid Centralization Concerns

According to 1M AI News monitoring, Liu Chengcheng, founder of tech media 36Kr, currently heads ByteDance's AI hardware team, Ocean. He reports directly to Zhu Jun, head of the AI product team, Flow, and is in charge of the Doubao mobile assistant and ByteDance's entire AI hardware business. After leading 36Kr to list on Nasdaq at the age of 30 in 2019, Liu Chengcheng secretly founded Whale Dolphin Technology to develop a domestic operating system and tablet hardware, primarily selling overseas. The company was later acquired by ByteDance, and Liu Chengcheng and his founding team subsequently joined ByteDance. After joining, he drove ByteDance's full acquisition of open-ear headphone brand Oladance (approximately $50 million) and then led the launch of the first AI hardware product, Ola Friend. After the AI glasses project was internally halted in February this year due to hardware homogenization, the Ocean team's strategic focus has shifted to the Doubao mobile assistant. This product, capable of executing complex tasks across applications at the system level, is considered the true breakthrough for ByteDance's AI hardware line. The Flow team (including Ocean) holds a priority level within ByteDance equivalent to that of the Douyin team.

[ByteDance's Doubao Assistant Led by 36Kr Founder Liu Chengcheng; Team Operates at Same Level as Douyin Internally]According to 1M AI News monitoring, Liu Chengcheng, founder of tech media 36Kr, currently heads ByteDance's AI hardware team, Ocean. He reports directly to Zhu Jun, head of the AI product team, Flow, and is in charge of the Doubao mobile assistant and ByteDance's entire AI hardware business. After leading 36Kr to list on Nasdaq at the age of 30 in 2019, Liu Chengcheng secretly founded Whale Dolphin Technology to develop a domestic operating system and tablet hardware, primarily selling overseas. The company was later acquired by ByteDance, and Liu Chengcheng and his founding team subsequently joined ByteDance. After joining, he drove ByteDance's full acquisition of open-ear headphone brand Oladance (approximately $50 million) and then led the launch of the first AI hardware product, Ola Friend. After the AI glasses project was internally halted in February this year due to hardware homogenization, the Ocean team's strategic focus has shifted to the Doubao mobile assistant. This product, capable of executing complex tasks across applications at the system level, is considered the true breakthrough for ByteDance's AI hardware line. The Flow team (including Ocean) holds a priority level within ByteDance equivalent to that of the Douyin team.

ByteDance's Doubao Assistant Led by 36Kr Founder Liu Chengcheng; Team Operates at Same Level as Douyin Internally

According to a report by the Financial Times on March 25, Manus CEO Hongchao Xiao and Chief Scientist Yichao Ji were summoned to Beijing for a meeting this month. After the meeting, they were informed that they are prohibited from leaving China due to regulatory review requirements, but they are still free to move within the country.
Previously, officials from China's National Development and Reform Commission (NDRC) had convened a meeting with executives from Meta and Manus to express concerns about the deal announced last December. No formal investigation has been launched, nor have any charges been filed. Manus is actively seeking assistance from law firms and consulting companies to address the matter.

[AI Unicorn Manus Founders Face Exit Ban in China]According to a report by the Financial Times on March 25, Manus CEO Hongchao Xiao and Chief Scientist Yichao Ji were summoned to Beijing for a meeting this month. After the meeting, they were informed that they are prohibited from leaving China due to regulatory review requirements, but they are still free to move within the country.
Previously, officials from China's National Development and Reform Commission (NDRC) had convened a meeting with executives from Meta and Manus to express concerns about the deal announced last December. No formal investigation has been launched, nor have any charges been filed. Manus is actively seeking assistance from law firms and consulting companies to address the matter.

AI Unicorn Manus Founders Face Exit Ban in China

HTX data showed that BTC broke through 72000  USDT and is now at 72006.61 USDT, with a increase of 1.21% in 24H.

[BTC broke through 72000  USDT] HTX data showed that BTC broke through 72000  USDT and is now at 72006.61 USDT, with a increase of 1.21% in 24H.

BTC broke through 72000  USDT

According to 1M AI News monitoring, DeepSeek has opened 17 recruitment positions, with a clear shift in core research and development focus from foundational model research to Agent productization. Three dedicated Agent positions cover the entire chain of algorithm research, data evaluation, and infrastructure. The Algorithm Researcher role focuses on the application of reinforcement learning in large model alignment (including RLHF/RLAIF, process rewards, preference learning, etc.). The Data Evaluation Specialist will be responsible for building evaluation datasets and designing test cases for Agent capabilities such as planning, tool invocation, multi-turn interaction, and long-term memory. The Infrastructure Engineer will focus on building the foundational runtime for Agents, including integrating external tools into internal reinforcement learning infrastructure and establishing evaluation platforms. Two notable signals are present in the recruitment requirements. Several positions explicitly mention a preference for candidates who have 'heavily used AI coding tools such as Claude Code, Cursor, and Copilot' as a bonus. The Full-Stack Development Engineer role includes an uncommon responsibility: 'As a heavy user of Vibe Coding, continuously explore innovative applications of model capabilities in products.' The core work direction for the same role is to build 'the next-generation container scheduling and isolation platform supporting the operation of massive AI Agents.' The Model Strategy Product Manager position has a dedicated Agent track, requiring candidates to be familiar with Agent products like Claude Code, OpenClaw, and Manus. They must identify high-value application scenarios (including personal assistants, Deep Research, automated workflows, and multimodal device control) and lead the design of Agent evaluation systems and training data schemes. Compared to January of this year, when core positions were concentrated in general areas like 'Deep Learning Researcher - AGI,' the current recruitment focus has clearly shifted towards Agent productization.

[DeepSeek Focuses on 17 Agent-Centric Roles, Prioritizing Heavy AI Coding Tool Usage]According to 1M AI News monitoring, DeepSeek has opened 17 recruitment positions, with a clear shift in core research and development focus from foundational model research to Agent productization. Three dedicated Agent positions cover the entire chain of algorithm research, data evaluation, and infrastructure. The Algorithm Researcher role focuses on the application of reinforcement learning in large model alignment (including RLHF/RLAIF, process rewards, preference learning, etc.). The Data Evaluation Specialist will be responsible for building evaluation datasets and designing test cases for Agent capabilities such as planning, tool invocation, multi-turn interaction, and long-term memory. The Infrastructure Engineer will focus on building the foundational runtime for Agents, including integrating external tools into internal reinforcement learning infrastructure and establishing evaluation platforms. Two notable signals are present in the recruitment requirements. Several positions explicitly mention a preference for candidates who have 'heavily used AI coding tools such as Claude Code, Cursor, and Copilot' as a bonus. The Full-Stack Development Engineer role includes an uncommon responsibility: 'As a heavy user of Vibe Coding, continuously explore innovative applications of model capabilities in products.' The core work direction for the same role is to build 'the next-generation container scheduling and isolation platform supporting the operation of massive AI Agents.' The Model Strategy Product Manager position has a dedicated Agent track, requiring candidates to be familiar with Agent products like Claude Code, OpenClaw, and Manus. They must identify high-value application scenarios (including personal assistants, Deep Research, automated workflows, and multimodal device control) and lead the design of Agent evaluation systems and training data schemes. Compared to January of this year, when core positions were concentrated in general areas like 'Deep Learning Researcher - AGI,' the current recruitment focus has clearly shifted towards Agent productization.

DeepSeek Focuses on 17 Agent-Centric Roles, Prioritizing Heavy AI Coding Tool Usage

A poll released by Reuters and Ipsos on March 24 shows that U.S. President Trump's approval rating has dropped from 40% last week to 36%, the lowest level since he returned to the White House. Reuters reported that the main reasons for the decline in Trump's approval rating are soaring fuel prices and widespread public dissatisfaction with the U.S. military action against Iran. (CCTV International News)

[Trump's Approval Rating Hits Lowest Point of Second Term]A poll released by Reuters and Ipsos on March 24 shows that U.S. President Trump's approval rating has dropped from 40% last week to 36%, the lowest level since he returned to the White House. Reuters reported that the main reasons for the decline in Trump's approval rating are soaring fuel prices and widespread public dissatisfaction with the U.S. military action against Iran. (CCTV International News)

Trump's Approval Rating Hits Lowest Point of Second Term

&nbsp;according to AFP: The Iranian Deputy Foreign Minister stated that Iran has not laid mines in the Strait of Hormuz.

[Iranian Deputy Foreign Minister: Iran has not laid mines in the Strait of Hormuz]&nbsp;according to AFP: The Iranian Deputy Foreign Minister stated that Iran has not laid mines in the Strait of Hormuz.

Iranian Deputy Foreign Minister: Iran has not laid mines in the Strait of Hormuz

Article

Solana Foundation Launches Developer Platform for the AI Era

High Stakes Capital Sells Over 600,000 HYPE Tokens for Nearly $22.94 Million

Tencent Recruits Key Personnel from ByteDance's Seed Unit, Including Xiao Xuefeng and Zhang Chi, Who Will Report to Yao Shunyu

OpenAI Aims to Raise $10 Billion from Investors Including MGX, Coatue, and Thrive

U.S. Stock Indices Close Lower

SOL broke through 90  USDT

Iranian Deputy Foreign Minister: Iran has not laid mines in the Strait of Hormuz

Disclaimer：Il contenuto di cui sopra non rappresenta le posizioni di HTX.，HTX non fornisce alcuna raccomandazione di trading.。

Tutti i commenti0RecentePopolare

Temi di tendenza