7*24

Hot topic

Market

Activities

According to monitoring by Beating, Tilde Research has discovered a hidden flaw in the optimizer Muon, used by leading models such as DeepSeek V4, Kimi K2.5, and GLM-5: it causes over a quarter of the neurons in the MLP layer to permanently die early in training. Based on this finding, the team designed an alternative optimizer, Aurora, and made it open-source. A 1.1B model, trained with approximately 100B tokens, matched the performance of Qwen3-1.7B, which was trained with 36T tokens, on language understanding benchmarks like HellaSwag and Winogrande. The issue lies in a mathematical characteristic of how Muon handles MLP weight matrices. Early in training, some neurons happen to receive weaker gradient signals. Traditional optimizers like AdamW normalize per parameter, naturally leveling out these differences; however, Muon's orthogonalization step passes the weak signals unchanged. Weak neurons continue to receive weak updates, becoming increasingly silent, creating a 'rich get richer' death spiral. By the 500th training step, over a quarter of the neurons are effectively dead, wasting parameter capacity. The previous improved version, NorMuon, alleviated this by forcing uniformity in update magnitudes for each row, but at the cost of disrupting the orthogonality of the update matrix (orthogonalization makes each step as efficient as possible, which is a core advantage of Muon), resulting in a loss of optimization precision. Aurora sets 'uniform updates' and 'orthogonality' as joint constraints, alternating iterations to satisfy both: ensuring that each neuron gets a fair learning opportunity without sacrificing update precision. Un-tuned, Aurora incurs only 6% more computational overhead than Muon and can be directly replaced. In modded-nanoGPT optimization benchmarks, Aurora set a new record with 3175 steps. The advantages of Aurora will also amplify with increased MLP width; the higher the expansion factor, the more pronounced the improvements. The code and the 1.1B pre-trained model have been made open-source.

[Muon Silently 'Starves' 25% of Neurons: Aurora Increases Data Efficiency by Over 100 Times After Fix]According to monitoring by Beating, Tilde Research has discovered a hidden flaw in the optimizer Muon, used by leading models such as DeepSeek V4, Kimi K2.5, and GLM-5: it causes over a quarter of the neurons in the MLP layer to permanently die early in training. Based on this finding, the team designed an alternative optimizer, Aurora, and made it open-source. A 1.1B model, trained with approximately 100B tokens, matched the performance of Qwen3-1.7B, which was trained with 36T tokens, on language understanding benchmarks like HellaSwag and Winogrande. The issue lies in a mathematical characteristic of how Muon handles MLP weight matrices. Early in training, some neurons happen to receive weaker gradient signals. Traditional optimizers like AdamW normalize per parameter, naturally leveling out these differences; however, Muon's orthogonalization step passes the weak signals unchanged. Weak neurons continue to receive weak updates, becoming increasingly silent, creating a 'rich get richer' death spiral. By the 500th training step, over a quarter of the neurons are effectively dead, wasting parameter capacity. The previous improved version, NorMuon, alleviated this by forcing uniformity in update magnitudes for each row, but at the cost of disrupting the orthogonality of the update matrix (orthogonalization makes each step as efficient as possible, which is a core advantage of Muon), resulting in a loss of optimization precision. Aurora sets 'uniform updates' and 'orthogonality' as joint constraints, alternating iterations to satisfy both: ensuring that each neuron gets a fair learning opportunity without sacrificing update precision. Un-tuned, Aurora incurs only 6% more computational overhead than Muon and can be directly replaced. In modded-nanoGPT optimization benchmarks, Aurora set a new record with 3175 steps. The advantages of Aurora will also amplify with increased MLP width; the higher the expansion factor, the more pronounced the improvements. The code and the 1.1B pre-trained model have been made open-source.

Muon Silently 'Starves' 25% of Neurons: Aurora Increases Data Efficiency by Over 100 Times After Fix

As of May 10, according to Coinglass data, the cryptocurrency Fear and Greed Index is currently at 48, an increase of 11 points from yesterday. The average over the past 7 days is 43, while the average over the past 30 days is 32.

[Current Crypto Fear and Greed Index at 48, Neutral State]As of May 10, according to Coinglass data, the cryptocurrency Fear and Greed Index is currently at 48, an increase of 11 points from yesterday. The average over the past 7 days is 43, while the average over the past 30 days is 32.

Current Crypto Fear and Greed Index at 48, Neutral State

On May 10, according to Barchart data, the S&P 500 index has now risen for six consecutive weeks, marking the longest streak since 2024. On the other hand, the SOX semiconductor index weekly chart shows that the overbought index RSI(14) has reached 85.54, the highest level since the peak of the 2000 internet bubble, raising concerns about the semiconductor sector being overbought. However, some analysts point out that the current strong rise in semiconductor stocks is different from the internet bubble era, as this increase is supported by real demand from AI, data centers, and corporate cash flow, rather than mere speculation.

[S&P 500 Posts Six Consecutive Weekly Gains; Semiconductor Stocks Reach Overbought Levels Not Seen Since Internet Bubble Peak]On May 10, according to Barchart data, the S&P 500 index has now risen for six consecutive weeks, marking the longest streak since 2024. On the other hand, the SOX semiconductor index weekly chart shows that the overbought index RSI(14) has reached 85.54, the highest level since the peak of the 2000 internet bubble, raising concerns about the semiconductor sector being overbought. However, some analysts point out that the current strong rise in semiconductor stocks is different from the internet bubble era, as this increase is supported by real demand from AI, data centers, and corporate cash flow, rather than mere speculation.

S&P 500 Posts Six Consecutive Weekly Gains; Semiconductor Stocks Reach Overbought Levels Not Seen Since Internet Bubble Peak

On May 10, significant signs of peace emerged in both the US-Iran and Russia-Ukraine conflicts this week, leading to a substantial easing of geopolitical risks. Next week, it will be crucial to monitor whether these two geopolitical conflicts can further de-escalate. Additionally, macroeconomic events will focus on the US April CPI data, with the following specifics: On Tuesday at 15:15, FOMC permanent voting member and New York Fed President Williams will participate in a panel discussion on monetary policy; on Tuesday at 20:15, the ADP employment change for the week ending April 25 will be released; on Tuesday at 20:30, the US April CPI data will be announced; on Wednesday at 04:30, API crude oil inventories for the week ending May 8 will be reported; on Wednesday at 20:30, the US April PPI year-on-year and month-on-month data will be released; on Friday at 05:30, Fed Governor Barr will give a speech; and on Friday at 21:15, the US April industrial production month-on-month data will be published. Lastly, next week will see significant personnel changes at the Fed, with incoming Chair Kevin Walsh expected to be confirmed by the Senate on Monday and officially taking over from Powell on May 15. In the stock market, as of this Friday, a new surge has driven the S&P 500 index to an 8% cumulative increase in 2026, continuing to rise on the foundation of double-digit returns over the past three years. The tech-heavy Nasdaq Composite index has risen nearly 13% year-to-date, with both indices reaching historic highs. Although the first quarter earnings season is nearing its end, corporate reports will remain a key driver of stock prices in the coming days.

[Macroeconomic Outlook for Next Week: Potential Peace Talks in US-Iran and Russia-Ukraine Conflicts, Walsh Officially Takes Over as Fed Chair]On May 10, significant signs of peace emerged in both the US-Iran and Russia-Ukraine conflicts this week, leading to a substantial easing of geopolitical risks. Next week, it will be crucial to monitor whether these two geopolitical conflicts can further de-escalate. Additionally, macroeconomic events will focus on the US April CPI data, with the following specifics: On Tuesday at 15:15, FOMC permanent voting member and New York Fed President Williams will participate in a panel discussion on monetary policy; on Tuesday at 20:15, the ADP employment change for the week ending April 25 will be released; on Tuesday at 20:30, the US April CPI data will be announced; on Wednesday at 04:30, API crude oil inventories for the week ending May 8 will be reported; on Wednesday at 20:30, the US April PPI year-on-year and month-on-month data will be released; on Friday at 05:30, Fed Governor Barr will give a speech; and on Friday at 21:15, the US April industrial production month-on-month data will be published. Lastly, next week will see significant personnel changes at the Fed, with incoming Chair Kevin Walsh expected to be confirmed by the Senate on Monday and officially taking over from Powell on May 15. In the stock market, as of this Friday, a new surge has driven the S&P 500 index to an 8% cumulative increase in 2026, continuing to rise on the foundation of double-digit returns over the past three years. The tech-heavy Nasdaq Composite index has risen nearly 13% year-to-date, with both indices reaching historic highs. Although the first quarter earnings season is nearing its end, corporate reports will remain a key driver of stock prices in the coming days.

Macroeconomic Outlook for Next Week: Potential Peace Talks in US-Iran and Russia-Ukraine Conflicts, Walsh Officially Takes Over as Fed Chair

On May 10, insiders revealed that Alibaba is preparing to announce the integration plan of its artificial intelligence platform Qwen with the e-commerce platform Taobao, aiming to enhance shopping through conversation rather than keyword searches. The integration will enable consumers to interact with an AI assistant via the Qwen app, allowing them to browse, compare, and purchase products without manually sifting through product lists. The Qwen app is expected to access the complete catalog of over 4 billion items from Taobao and Tmall, supported by a 'skill library' capable of managing logistics and after-sales services. It will also provide shopping recommendations based on users' order history and shopping preferences. Additionally, within the Taobao platform, Alibaba will launch an AI shopping assistant powered by Qwen, featuring tools such as virtual try-ons and 30-day price tracking.

[Sources: Alibaba to Integrate Qwen AI with Taobao for Smart Shopping Service]On May 10, insiders revealed that Alibaba is preparing to announce the integration plan of its artificial intelligence platform Qwen with the e-commerce platform Taobao, aiming to enhance shopping through conversation rather than keyword searches. The integration will enable consumers to interact with an AI assistant via the Qwen app, allowing them to browse, compare, and purchase products without manually sifting through product lists. The Qwen app is expected to access the complete catalog of over 4 billion items from Taobao and Tmall, supported by a 'skill library' capable of managing logistics and after-sales services. It will also provide shopping recommendations based on users' order history and shopping preferences. Additionally, within the Taobao platform, Alibaba will launch an AI shopping assistant powered by Qwen, featuring tools such as virtual try-ons and 30-day price tracking.

Sources: Alibaba to Integrate Qwen AI with Taobao for Smart Shopping Service

"AI Geek Championship" has officially launched today. This event is produced by HTX DAO and jointly organized by Golden Finance, Twinkle, and Dushu Finance, bringing together top industry resources to explore the innovative power of the integration of AI and blockchain. The competition focuses on the cross-disciplinary collision of AI technology and Web3, symbolizing the underlying logic and innovative pulse of the digital intelligence era. This is not only a contest of technology but also a foresight into future trends. Scan the QR code now to follow the latest updates on the event and witness the birth of AI geeks.

[AI Geek Championship Officially Launched on May 10]"AI Geek Championship" has officially launched today. This event is produced by HTX DAO and jointly organized by Golden Finance, Twinkle, and Dushu Finance, bringing together top industry resources to explore the innovative power of the integration of AI and blockchain. The competition focuses on the cross-disciplinary collision of AI technology and Web3, symbolizing the underlying logic and innovative pulse of the digital intelligence era. This is not only a contest of technology but also a foresight into future trends. Scan the QR code now to follow the latest updates on the event and witness the birth of AI geeks.

AI Geek Championship Officially Launched on May 10

On May 10, according to the Islamic Republic News Agency (IRNA), Khalil Sherkhani, Iran's Ambassador to Armenia, emphasized that an agreement could be reached if countries that have illegally aggressed against Iran return to rational and logical policies and actions. In an interview with Armenian television on Saturday, Sherkhani highlighted that the aggression of the U.S. and Israeli regimes against Iran clearly indicates that military means, pressure, or threats cannot force Iran to surrender or accept extreme demands from either side. He stated that Iran is willing to reach an agreement through negotiations, but this requires clear guarantees that the U.S. will adopt rational and logical approaches to ensure that military aggression against Iran will not occur again. He also stressed that Iran opposes selective interpretations of international law and will never allow the Iranian nation to be deprived of its right to peacefully utilize nuclear capabilities.

[Iranian Envoy: Agreement Possible if U.S. Acts Rationally]On May 10, according to the Islamic Republic News Agency (IRNA), Khalil Sherkhani, Iran's Ambassador to Armenia, emphasized that an agreement could be reached if countries that have illegally aggressed against Iran return to rational and logical policies and actions. In an interview with Armenian television on Saturday, Sherkhani highlighted that the aggression of the U.S. and Israeli regimes against Iran clearly indicates that military means, pressure, or threats cannot force Iran to surrender or accept extreme demands from either side. He stated that Iran is willing to reach an agreement through negotiations, but this requires clear guarantees that the U.S. will adopt rational and logical approaches to ensure that military aggression against Iran will not occur again. He also stressed that Iran opposes selective interpretations of international law and will never allow the Iranian nation to be deprived of its right to peacefully utilize nuclear capabilities.

Iranian Envoy: Agreement Possible if U.S. Acts Rationally

On May 10, Wasabi Protocol released an update regarding a security incident, stating that attackers exploited a configuration vulnerability in the Spring Boot Actuator within its AWS infrastructure to steal the private keys controlling EVM smart contracts, resulting in the theft of approximately $4.8 million in user funds and $900,000 from the protocol's treasury. The attack chain began with a public server used for analysis, whose Actuator heap dump was not properly password protected, allowing the attackers to obtain credentials from another server and ultimately gain control over the smart contract private keys. This incident only affected EVM deployments, including certain treasuries on Ethereum, Base, Blast, and Berachain, while Solana deployments and Prop AMM were not impacted. Wasabi Protocol stated that a final compensation plan for affected users has not yet been established, but ensuring that all harmed users receive compensation remains the team's top priority, with future updates on the investigation to be released in their Discord community.

[Wasabi Protocol Updates on Security Incident: No Final User Compensation Plan Yet]On May 10, Wasabi Protocol released an update regarding a security incident, stating that attackers exploited a configuration vulnerability in the Spring Boot Actuator within its AWS infrastructure to steal the private keys controlling EVM smart contracts, resulting in the theft of approximately $4.8 million in user funds and $900,000 from the protocol's treasury. The attack chain began with a public server used for analysis, whose Actuator heap dump was not properly password protected, allowing the attackers to obtain credentials from another server and ultimately gain control over the smart contract private keys. This incident only affected EVM deployments, including certain treasuries on Ethereum, Base, Blast, and Berachain, while Solana deployments and Prop AMM were not impacted. Wasabi Protocol stated that a final compensation plan for affected users has not yet been established, but ensuring that all harmed users receive compensation remains the team's top priority, with future updates on the investigation to be released in their Discord community.

Wasabi Protocol Updates on Security Incident: No Final User Compensation Plan Yet

On May 10, Wasabi Protocol released an update regarding a security incident, stating that attackers exploited a configuration vulnerability in the Spring Boot Actuator within its AWS infrastructure to steal the private keys controlling EVM smart contracts, resulting in the theft of approximately $4.8 million in user funds and $900,000 in protocol treasury funds. The attack chain began with a public server used for analysis, whose Actuator heap dump was not properly password protected, allowing the attackers to obtain credentials for another server and ultimately gain control over the smart contract private keys. This incident only affected EVM deployments, including certain treasuries on Ethereum, Base, Blast, and Berachain, while Solana deployments and Prop AMM were not impacted. Wasabi Protocol stated that a final compensation plan for users has not yet been determined, but ensuring that all affected users are compensated remains the team's highest priority, with updates on the investigation progress to be shared in the Discord community in the future.

[Wasabi Protocol Updates on Security Incident: No Final User Compensation Plan Yet]On May 10, Wasabi Protocol released an update regarding a security incident, stating that attackers exploited a configuration vulnerability in the Spring Boot Actuator within its AWS infrastructure to steal the private keys controlling EVM smart contracts, resulting in the theft of approximately $4.8 million in user funds and $900,000 in protocol treasury funds. The attack chain began with a public server used for analysis, whose Actuator heap dump was not properly password protected, allowing the attackers to obtain credentials for another server and ultimately gain control over the smart contract private keys. This incident only affected EVM deployments, including certain treasuries on Ethereum, Base, Blast, and Berachain, while Solana deployments and Prop AMM were not impacted. Wasabi Protocol stated that a final compensation plan for users has not yet been determined, but ensuring that all affected users are compensated remains the team's highest priority, with updates on the investigation progress to be shared in the Discord community in the future.

Immunefi posted that Aptos launched a bug bounty program with a maximum reward of $1 million, paid in USDC and USDT, and the vulnerability types are blockchain and smart contracts.

Detail

【Aptos launched a bug bounty program with a maximum reward of $1 million】Immunefi posted that Aptos launched a bug bounty program with a maximum reward of $1 million, paid in USDC and USDT, and the vulnerability types are blockchain and smart contracts.

Aptos launched a bug bounty program with a maximum reward of $1 million

BVNK 创始人：稳定币发展的三个阶段

Bitcoin 2026大会洞察：挖矿衰落，AI崛起，一场结构性大迁徙

Sam Altman最新专访自白：其实我也不太懂AI内部在发生什么

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

Anthropic搞了个全是AI的闲鱼群，大模型在里面互割起了韭菜

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

US Court Approves Aave's Transfer of $71 Million ETH Related to North Korean Hacking Incident

Aptos launched a bug bounty program with a maximum reward of $1 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