7*24

Hot topic

Market

Activities

According to monitoring by Dongcha Beating, Trajectory, in collaboration with UC Berkeley's Sky Computing Lab and Anyscale, has announced the launch of the open-source multi-LoRA reinforcement learning training platform SkyRL, along with the release of the concurrent training architecture Multi-LoRA Training that supports continuous learning for large models. In traditional large model fine-tuning experiments, developers must individually schedule GPU nodes for each fine-tuning task, frequently loading and unloading the same massive model on servers. For instance, training a giant model with 397 billion parameters traditionally requires multiple sets of computing nodes to repeatedly load hundreds of GB of weight files, resulting in significant waste of computational time and resources. The Multi-LoRA architecture completely eliminates the dilemma of repeated cold starts. The system opts to keep a shared model base online in GPU memory and manages multiple different fine-tuning experiments as lightweight 'adapter modules.' During the training phase, the system serially executes forward and backward propagation by quickly swapping in and out the states of each adapter between the CPU and GPU, thus avoiding the cold start overhead of repeatedly loading the giant base model. Combined with concurrent processing during the inference phase, the overall experimental throughput for multi-tasking is significantly enhanced. Test data shows that, while ensuring no degradation in large model performance, the end-to-end experimental throughput has increased by up to 2.81 times, and the throughput within absolute time on a single node has improved by approximately 3.25 times. The training code is now open-sourced in the SkyRL repository, aiming to assist developers in achieving hourly self-evolution of large models with minimal memory cost through real-time production data.

[Large Model Achieves Hourly Self-Evolution; Open Source SkyRL Concurrent Training Stack Delivers 2.8x Efficiency Boost]According to monitoring by Dongcha Beating, Trajectory, in collaboration with UC Berkeley's Sky Computing Lab and Anyscale, has announced the launch of the open-source multi-LoRA reinforcement learning training platform SkyRL, along with the release of the concurrent training architecture Multi-LoRA Training that supports continuous learning for large models. In traditional large model fine-tuning experiments, developers must individually schedule GPU nodes for each fine-tuning task, frequently loading and unloading the same massive model on servers. For instance, training a giant model with 397 billion parameters traditionally requires multiple sets of computing nodes to repeatedly load hundreds of GB of weight files, resulting in significant waste of computational time and resources. The Multi-LoRA architecture completely eliminates the dilemma of repeated cold starts. The system opts to keep a shared model base online in GPU memory and manages multiple different fine-tuning experiments as lightweight 'adapter modules.' During the training phase, the system serially executes forward and backward propagation by quickly swapping in and out the states of each adapter between the CPU and GPU, thus avoiding the cold start overhead of repeatedly loading the giant base model. Combined with concurrent processing during the inference phase, the overall experimental throughput for multi-tasking is significantly enhanced. Test data shows that, while ensuring no degradation in large model performance, the end-to-end experimental throughput has increased by up to 2.81 times, and the throughput within absolute time on a single node has improved by approximately 3.25 times. The training code is now open-sourced in the SkyRL repository, aiming to assist developers in achieving hourly self-evolution of large models with minimal memory cost through real-time production data.

Large Model Achieves Hourly Self-Evolution; Open Source SkyRL Concurrent Training Stack Delivers 2.8x Efficiency Boost

On May 31, local time May 30, US President Trump stated in an interview aired on Fox News that the US and Iran are close to reaching a 'very good agreement.' Trump also mentioned that we will get a 'better agreement,' which should have been achieved more quickly. However, he indicated that if an agreement cannot be reached, he would ask the 'War Department' (referring to the US Department of Defense) to intervene. Trump stated that if he does not get what he wants, 'we will end it another way.' He also emphasized that by signing an agreement, we can ensure the navigation of the Strait of Hormuz and ensure that Iran does not have nuclear weapons. When discussing the agreement with Iran, Trump stressed that he is not in a hurry. 'If you are in a hurry, you cannot reach a good agreement.' (CCTV News)

[Trump: If US-Iran Agreement Fails, 'War Department' Will Intervene]On May 31, local time May 30, US President Trump stated in an interview aired on Fox News that the US and Iran are close to reaching a 'very good agreement.' Trump also mentioned that we will get a 'better agreement,' which should have been achieved more quickly. However, he indicated that if an agreement cannot be reached, he would ask the 'War Department' (referring to the US Department of Defense) to intervene. Trump stated that if he does not get what he wants, 'we will end it another way.' He also emphasized that by signing an agreement, we can ensure the navigation of the Strait of Hormuz and ensure that Iran does not have nuclear weapons. When discussing the agreement with Iran, Trump stressed that he is not in a hurry. 'If you are in a hurry, you cannot reach a good agreement.' (CCTV News)

Trump: If US-Iran Agreement Fails, 'War Department' Will Intervene

According to monitoring by Dongcha Beating, Weijie Su, an alumnus of Peking University's School of Mathematical Sciences from the class of 2007, a PhD in Statistics from Stanford University, and a full professor at the Wharton School of the University of Pennsylvania, announced on social media platform X that he has officially joined OpenAI during his sabbatical from Wharton. Su stated that he will participate in AI model training upon returning to the Bay Area after ten years. Additionally, after a decade of academic efforts, Su has been officially promoted to full professor in the Department of Statistics and Data Science at the Wharton School. Su has made significant contributions in the fields of statistics, machine learning optimization, and theoretical safety of large language models (LLMs). He received the COPSS Presidents' Award in 2026, one of the highest honors in the international statistics field. The COPSS official commendation highly praised his work in establishing the statistical foundations of generative AI (including watermarking mechanisms for LLMs, human preference alignment, and ranking theory), advancing differential privacy data analysis methods in the 2020 U.S. Census, and foundational contributions to deep learning theory, convex optimization, and high-dimensional statistical inference. In academic research, Su has made several advancements in reshaping the underlying statistics of large models. In watermark detection for large models, he formalized watermarking techniques as statistical hypothesis testing problems and proposed the 'Truncated Goodness-of-Fit Test' (Tr-GoF). Tr-GoF enhances the robustness of watermark detection for modified and trimmed texts without relying on prior knowledge of human editing processes or precise probabilities. Regarding preference alignment (RLHF), Su pointed out that traditional RLHF, based on ranking optimization mechanisms, is prone to 'Preference Collapse,' which overlooks the preferences of minority groups and leads to a homogenized reward distribution across different prompts. To preserve response diversity, the team derived a 'Preference Matching' regularization scheme by solving ordinary differential equations (ODEs), maximizing human rewards while avoiding uniform responses, and explored a game-theoretic perspective of 'Nash Learning Based on Human Feedback' (NLHF) as an alternative to traditional scalar reward alignment. Su graduated with a bachelor's degree from Peking University's School of Mathematical Sciences and is referred to as part of the 'Golden Second Generation' (or 'Platinum First Generation') of Peking University's mathematics department by media and scholars. Compared to the 'Golden Generation' of the class of 2000, known for pure mathematical research (such as Yun Zhiwei, Zhang Wei, and Xu Chenyang), Su and his peers from the class of 2007 exhibit broader interdisciplinary characteristics, expanding the boundaries of mathematics into statistics, data science, and artificial intelligence.

[Weijie Su Joins OpenAI for Model Training and Promoted to Full Professor at Wharton]According to monitoring by Dongcha Beating, Weijie Su, an alumnus of Peking University's School of Mathematical Sciences from the class of 2007, a PhD in Statistics from Stanford University, and a full professor at the Wharton School of the University of Pennsylvania, announced on social media platform X that he has officially joined OpenAI during his sabbatical from Wharton. Su stated that he will participate in AI model training upon returning to the Bay Area after ten years. Additionally, after a decade of academic efforts, Su has been officially promoted to full professor in the Department of Statistics and Data Science at the Wharton School. Su has made significant contributions in the fields of statistics, machine learning optimization, and theoretical safety of large language models (LLMs). He received the COPSS Presidents' Award in 2026, one of the highest honors in the international statistics field. The COPSS official commendation highly praised his work in establishing the statistical foundations of generative AI (including watermarking mechanisms for LLMs, human preference alignment, and ranking theory), advancing differential privacy data analysis methods in the 2020 U.S. Census, and foundational contributions to deep learning theory, convex optimization, and high-dimensional statistical inference. In academic research, Su has made several advancements in reshaping the underlying statistics of large models. In watermark detection for large models, he formalized watermarking techniques as statistical hypothesis testing problems and proposed the 'Truncated Goodness-of-Fit Test' (Tr-GoF). Tr-GoF enhances the robustness of watermark detection for modified and trimmed texts without relying on prior knowledge of human editing processes or precise probabilities. Regarding preference alignment (RLHF), Su pointed out that traditional RLHF, based on ranking optimization mechanisms, is prone to 'Preference Collapse,' which overlooks the preferences of minority groups and leads to a homogenized reward distribution across different prompts. To preserve response diversity, the team derived a 'Preference Matching' regularization scheme by solving ordinary differential equations (ODEs), maximizing human rewards while avoiding uniform responses, and explored a game-theoretic perspective of 'Nash Learning Based on Human Feedback' (NLHF) as an alternative to traditional scalar reward alignment. Su graduated with a bachelor's degree from Peking University's School of Mathematical Sciences and is referred to as part of the 'Golden Second Generation' (or 'Platinum First Generation') of Peking University's mathematics department by media and scholars. Compared to the 'Golden Generation' of the class of 2000, known for pure mathematical research (such as Yun Zhiwei, Zhang Wei, and Xu Chenyang), Su and his peers from the class of 2007 exhibit broader interdisciplinary characteristics, expanding the boundaries of mathematics into statistics, data science, and artificial intelligence.

Weijie Su Joins OpenAI for Model Training and Promoted to Full Professor at Wharton

On May 31, recent DeFi attacks targeting protocols such as Thorchain, Drift, and KelpDAO have relied on flash loans, a mechanism that does not exist on the XRP Ledger. Due to the atomic nature of XRPL transactions and the inability to include composable calls within transactions, flash loan attacks are structurally impossible on this network. As XRPL advances its AMM upgrade and the tokenization of real-world assets continues to expand, institutional investors may weigh this built-in resistance to attacks against the deeper liquidity and more mature DeFi ecosystem of Ethereum.

[New Proposal on XRP Ledger to Prevent Flash Loan Attacks Causing Hundreds of Millions in DeFi Losses]On May 31, recent DeFi attacks targeting protocols such as Thorchain, Drift, and KelpDAO have relied on flash loans, a mechanism that does not exist on the XRP Ledger. Due to the atomic nature of XRPL transactions and the inability to include composable calls within transactions, flash loan attacks are structurally impossible on this network. As XRPL advances its AMM upgrade and the tokenization of real-world assets continues to expand, institutional investors may weigh this built-in resistance to attacks against the deeper liquidity and more mature DeFi ecosystem of Ethereum.

New Proposal on XRP Ledger to Prevent Flash Loan Attacks Causing Hundreds of Millions in DeFi Losses

genius Surges 10.06% in 24H, Currently at 0.4956 USDT

【genius currently trades at 0.4956 USDT, up 10.06% in 24h, according to HTX market data.】genius Surges 10.06% in 24H, Currently at 0.4956 USDT

genius currently trades at 0.4956 USDT, up 10.06% in 24h, according to HTX market data.

On May 31, according to Tokens on Solana, Strategy's perpetual preferred stock STRCx has been launched on Kamino, becoming the first lending market on Solana to support tokenized STRC.

[Strategy's Perpetual Preferred Stock STRCx Launched on Kamino]On May 31, according to Tokens on Solana, Strategy's perpetual preferred stock STRCx has been launched on Kamino, becoming the first lending market on Solana to support tokenized STRC.

Strategy's Perpetual Preferred Stock STRCx Launched on Kamino

On May 31, it was reported by US sources on May 30 that after the Supreme Court ruled that the Trump administration lacked the authority to impose higher tariffs on nearly all trade partners, American businesses have begun receiving tariff refunds. However, this process may be hindered by the Trump administration's preparation to appeal. Reports indicate that on May 29, the Trump administration announced plans to appeal a federal judge's ruling that allows all importers who paid the invalidated tariffs to apply for refunds, not just those who have filed lawsuits. (CCTV News)

[US Government Reportedly Plans to Appeal Tariff Refund Ruling]On May 31, it was reported by US sources on May 30 that after the Supreme Court ruled that the Trump administration lacked the authority to impose higher tariffs on nearly all trade partners, American businesses have begun receiving tariff refunds. However, this process may be hindered by the Trump administration's preparation to appeal. Reports indicate that on May 29, the Trump administration announced plans to appeal a federal judge's ruling that allows all importers who paid the invalidated tariffs to apply for refunds, not just those who have filed lawsuits. (CCTV News)

US Government Reportedly Plans to Appeal Tariff Refund Ruling

mega Surges 10.02% in 24H, Currently at 0.067 USDT

【mega currently trades at 0.067 USDT, up 10.02% in 24h, according to HTX market data.】mega Surges 10.02% in 24H, Currently at 0.067 USDT

mega currently trades at 0.067 USDT, up 10.02% in 24h, according to HTX market data.

On May 31, Aave published a post-incident investigation regarding the rsETH event that occurred on April 18. The investigation revealed that the rsETH LayerZero V2 cross-chain bridge of the liquidity staking protocol Kelp accepted forged messages during the cross-chain process from Unichain to Ethereum, resulting in the Ethereum side adapter releasing 116,500 rsETH, while the corresponding destruction did not occur on the Unichain side. Aave stated that the attack took place on third-party cross-chain bridge infrastructure, but the attacker deposited the stolen rsETH into eight Aave V3 positions and borrowed 82,650 WETH and 821 wstETH, impacting the Aave market. Aave noted that the attacker’s rsETH on Arbitrum has been destroyed, and the LayerZero OFT adapter has replenished 116,131.72 rsETH in five batches, fully restoring support for rsETH assets, with the affected WETH and rsETH markets returning to normal.

[Aave Releases Post-Incident Review of April 18 rsETH Event]On May 31, Aave published a post-incident investigation regarding the rsETH event that occurred on April 18. The investigation revealed that the rsETH LayerZero V2 cross-chain bridge of the liquidity staking protocol Kelp accepted forged messages during the cross-chain process from Unichain to Ethereum, resulting in the Ethereum side adapter releasing 116,500 rsETH, while the corresponding destruction did not occur on the Unichain side. Aave stated that the attack took place on third-party cross-chain bridge infrastructure, but the attacker deposited the stolen rsETH into eight Aave V3 positions and borrowed 82,650 WETH and 821 wstETH, impacting the Aave market. Aave noted that the attacker’s rsETH on Arbitrum has been destroyed, and the LayerZero OFT adapter has replenished 116,131.72 rsETH in five batches, fully restoring support for rsETH assets, with the affected WETH and rsETH markets returning to normal.

Aave Releases Post-Incident Review of April 18 rsETH Event

According to Reuters, Facebook-parent Meta Platforms Inc (META.O) said on Tuesday it had raised $10 billion in its first-ever bond offering, as it looks to fund share buybacks and investments to revamp its business. The offering would help Meta, the only one among big technology companies without debt on its books, to build a more traditional balance sheet and fund some expensive initiatives, such as its metaverse virtual reality. Other tech giants such as Apple Inc (AAPL.O) and Intel Corp (INTC.O) also issued bonds recently, raising $5.5 billion and $6 billion, respectively.

Detail

【Meta raises $10 billion in first-ever bond offering】According to Reuters, Facebook-parent Meta Platforms Inc (META.O) said on Tuesday it had raised $10 billion in its first-ever bond offering, as it looks to fund share buybacks and investments to revamp its business. The offering would help Meta, the only one among big technology companies without debt on its books, to build a more traditional balance sheet and fund some expensive initiatives, such as its metaverse virtual reality. Other tech giants such as Apple Inc (AAPL.O) and Intel Corp (INTC.O) also issued bonds recently, raising $5.5 billion and $6 billion, respectively.

Meta raises $10 billion in first-ever bond offering

The Revived Codex, Carrying OpenAI's Hopes for IPO

Google CEO Admits Lagging Behind in Coding

Everyone Is Calling For Lower Bitcoin Price: Why This Is The Perfect Time To Go Parabolic

Bitcoin Spot ETFs Bleed $1.26 Billion In Largest Net Outflows In 3 Months – Details

Microsoft to Discontinue Using Claude: Too Expensive, or a Calculated Move?

NeoCloud Three Giants: NBIS, IREN, CRWV – Which One Has More Investment Value?

New York Lawsuit Takes Aim At 3.79 Million Dormant Bitcoin

Eric Trump Sets A “Beyond Catastrophic” Bar To Sell Bitcoin — How Far Are We From That?

Investors Frantically Snap Up AI Firms with 'No Profits': A High-Stakes Gamble on 'the Right to Define the Future'

The 'Stock Call King' Serenity: 3840% Annual Return, Beating Institutions to the Punch

Article

a16z-Linked Entity Increases Holdings by 226,121 HYPE, Totaling 3.9 Million HYPE at Average Price of $49.4

Meta raises $10 billion in first-ever bond offering

Descargo de responsabilidad：El contenido anterior no representa las posiciones de HTX.，HTX no ofrece ninguna recomendación de trading.。

Artículos relacionados

Todos los comentarios0Lo más recientePopular

Temas populares

Artículos relacionados