7*24

Hot topic

Market

Activities

On May 12, Michael Saylor tweeted that the CLARITY Act, being advanced by the U.S. Congress, will unlock the potential of the next phase of the 'digital capital, digital credit, and digital equity' market, and will bring stronger institutional legitimacy and adoption frameworks for BTC.

[Michael Saylor: The CLARITY Act Will Drive Institutional Adoption of Digital Capital and Bitcoin]On May 12, Michael Saylor tweeted that the CLARITY Act, being advanced by the U.S. Congress, will unlock the potential of the next phase of the 'digital capital, digital credit, and digital equity' market, and will bring stronger institutional legitimacy and adoption frameworks for BTC.

Michael Saylor: The CLARITY Act Will Drive Institutional Adoption of Digital Capital and Bitcoin

On May 12, Wells Fargo raised the target price for Nvidia (NVDA.O) from $265 to $315.

[Wells Fargo Raises Nvidia Target Price from $265 to $315]On May 12, Wells Fargo raised the target price for Nvidia (NVDA.O) from $265 to $315.

Wells Fargo Raises Nvidia Target Price from $265 to $315

According to monitoring by Dongcha Beating, Dylan Zhang, a PhD student in computer science at the University of Illinois, conducted a series of memory experiments with agents, leading to an unusual conclusion: repeatedly summarizing experiences may actually degrade the model's memory. One particularly striking result came from ARC-AGI: researchers selected 19 questions that GPT-5.4 could answer correctly without memory, then provided the model with the true solutions to these questions while it simultaneously wrote 'experience summaries.' Logically, this is akin to open-book review; however, after multiple rounds of memory compression, the same model's accuracy plummeted from 100% to 54%. The original trajectory was not incorrect; the real issue was the model rewriting the correct trajectory into generalized experiences. Worse still, this memory degradation is not an isolated case. In the WebShop online shopping task, the AWM memory method scored 0.64 after ingesting 8 expert trajectories, but when the number of trajectories increased to 128, the score dropped to 0.20, exactly returning to the no-memory baseline. This means that as memory accumulates, the benefits are effectively nullified by the model itself. The problem lies not in 'too little experience,' but in 'too frequent summarization.' The experiences written by large models are not objective logs; each summary is a form of regeneration. By the end, specific premises are discarded, rules for different tasks are blended together, and details that could guide actions transform into seemingly correct yet practically useless statements like 'prioritize the most direct action' or 'use the correct tools.' An extreme example presented in the original text shows that 50 structured memories were merged into one, compressing multiple task differences into a single generalized process, resulting in the loss of 6 to 13 successful samples in the next round of evaluation. The author's suggestion is quite restrained: do not rush to have the agent write a 'mistake book' in every round. A more stable approach is to retain filtered original operation trajectories and only abstract summaries when truly necessary. In the experiments, keeping the original episodes and disabling abstract summarization led to performance that matched or exceeded tested compressed memory methods across multiple agent benchmarks. For developers, this conclusion is straightforward: showing the model what it has actually done is usually more useful than having it memorize a bunch of abstract rules.

[AI Review Reversal: GPT-5.4's Perfect Score Solutions Drop to 54%]According to monitoring by Dongcha Beating, Dylan Zhang, a PhD student in computer science at the University of Illinois, conducted a series of memory experiments with agents, leading to an unusual conclusion: repeatedly summarizing experiences may actually degrade the model's memory. One particularly striking result came from ARC-AGI: researchers selected 19 questions that GPT-5.4 could answer correctly without memory, then provided the model with the true solutions to these questions while it simultaneously wrote 'experience summaries.' Logically, this is akin to open-book review; however, after multiple rounds of memory compression, the same model's accuracy plummeted from 100% to 54%. The original trajectory was not incorrect; the real issue was the model rewriting the correct trajectory into generalized experiences. Worse still, this memory degradation is not an isolated case. In the WebShop online shopping task, the AWM memory method scored 0.64 after ingesting 8 expert trajectories, but when the number of trajectories increased to 128, the score dropped to 0.20, exactly returning to the no-memory baseline. This means that as memory accumulates, the benefits are effectively nullified by the model itself. The problem lies not in 'too little experience,' but in 'too frequent summarization.' The experiences written by large models are not objective logs; each summary is a form of regeneration. By the end, specific premises are discarded, rules for different tasks are blended together, and details that could guide actions transform into seemingly correct yet practically useless statements like 'prioritize the most direct action' or 'use the correct tools.' An extreme example presented in the original text shows that 50 structured memories were merged into one, compressing multiple task differences into a single generalized process, resulting in the loss of 6 to 13 successful samples in the next round of evaluation. The author's suggestion is quite restrained: do not rush to have the agent write a 'mistake book' in every round. A more stable approach is to retain filtered original operation trajectories and only abstract summaries when truly necessary. In the experiments, keeping the original episodes and disabling abstract summarization led to performance that matched or exceeded tested compressed memory methods across multiple agent benchmarks. For developers, this conclusion is straightforward: showing the model what it has actually done is usually more useful than having it memorize a bunch of abstract rules.

AI Review Reversal: GPT-5.4's Perfect Score Solutions Drop to 54%

Slow Mist issued a security warning stating that Aurellion was attacked, resulting in a loss of approximately 455,003 USDC (about $455,000). Analysis indicates that the root of the vulnerability may lie in the lack of effective protection in the initialize(address) function of the SafeOwnable Facet. The Diamond contract failed to update the _initialized version slot correctly because it did not go through the initialize path when setting the owner, allowing the attacker to reinitialize the contract and override owner permissions. Slow Mist reported that the attacker has seized ownership of the Diamond contract and transferred USDC from multiple authorized addresses.

[Slow Mist: Aurellion Attacked Due to Initialization Vulnerability, Approximately $455,000 USDC Stolen]Slow Mist issued a security warning stating that Aurellion was attacked, resulting in a loss of approximately 455,003 USDC (about $455,000). Analysis indicates that the root of the vulnerability may lie in the lack of effective protection in the initialize(address) function of the SafeOwnable Facet. The Diamond contract failed to update the _initialized version slot correctly because it did not go through the initialize path when setting the owner, allowing the attacker to reinitialize the contract and override owner permissions. Slow Mist reported that the attacker has seized ownership of the Diamond contract and transferred USDC from multiple authorized addresses.

Slow Mist: Aurellion Attacked Due to Initialization Vulnerability, Approximately $455,000 USDC Stolen

According to monitoring by Dongcha Beating, NVIDIA and the MIT research team have released a new post-training framework for large language models called Lightning OPD (Offline Policy Distillation). This technology enhances training efficiency by 4 times by pre-computing the log-probabilities of the teacher model offline, eliminating the need for real-time teacher services that must remain online throughout traditional distillation training. Previously, standard Online Policy Distillation (OPD) required both student and teacher models to run simultaneously on a single machine, which often led to out-of-memory (OOM) errors as models grew larger. Lightning OPD frees up all GPU resources for the student model. In tests using 8 H100 GPUs on a single node, Lightning OPD successfully distilled the Qwen3-30B-A3B-Base (a large MoE model with a total of 30 billion parameters), achieving a score of 71.0 in the AIME 2024 test; in contrast, standard OPD encountered OOM errors under the same hardware configuration. For the smaller Qwen3-8B model, this framework achieved a score of 69.9 in just 30 GPU hours. The research team pointed out a hidden prerequisite for achieving offline distillation in their paper: 'teacher consistency.' The student model must use the same teacher model during supervised fine-tuning (SFT) and subsequent distillation phases. Failing to adhere to this principle can lead to biased gradient directions, ultimately degrading the model's performance.

[MIT and NVIDIA Launch Lightning OPD: A 4x Faster Offline Distillation Framework]According to monitoring by Dongcha Beating, NVIDIA and the MIT research team have released a new post-training framework for large language models called Lightning OPD (Offline Policy Distillation). This technology enhances training efficiency by 4 times by pre-computing the log-probabilities of the teacher model offline, eliminating the need for real-time teacher services that must remain online throughout traditional distillation training. Previously, standard Online Policy Distillation (OPD) required both student and teacher models to run simultaneously on a single machine, which often led to out-of-memory (OOM) errors as models grew larger. Lightning OPD frees up all GPU resources for the student model. In tests using 8 H100 GPUs on a single node, Lightning OPD successfully distilled the Qwen3-30B-A3B-Base (a large MoE model with a total of 30 billion parameters), achieving a score of 71.0 in the AIME 2024 test; in contrast, standard OPD encountered OOM errors under the same hardware configuration. For the smaller Qwen3-8B model, this framework achieved a score of 69.9 in just 30 GPU hours. The research team pointed out a hidden prerequisite for achieving offline distillation in their paper: 'teacher consistency.' The student model must use the same teacher model during supervised fine-tuning (SFT) and subsequent distillation phases. Failing to adhere to this principle can lead to biased gradient directions, ultimately degrading the model's performance.

MIT and NVIDIA Launch Lightning OPD: A 4x Faster Offline Distillation Framework

On May 12, Bakkt released its Q1 2026 financial report, showing a net loss attributable to the company of $11.7 million, or a loss of $0.41 per share; compared to a net profit of $7.7 million in the same period last year. The decline in crypto trading volume impacted Bakkt's crypto services revenue, which fell from $1.07 billion in the same quarter last year to $243.6 million, a 77% year-on-year decrease, although most of this was offset by crypto costs and brokerage fees. At the end of the first quarter, the company held $82.6 million in cash and had no long-term debt. Bakkt stated that it is transitioning from crypto trading infrastructure to stablecoin payments and AI financial infrastructure, having completed the acquisition of Distributed Technologies Research on April 30, gaining an AI-native payment engine and a stablecoin compliance technology stack.

[Bakkt Q1 Crypto Revenue Falls 77% Year-on-Year, Shifts Focus to Stablecoin Infrastructure]On May 12, Bakkt released its Q1 2026 financial report, showing a net loss attributable to the company of $11.7 million, or a loss of $0.41 per share; compared to a net profit of $7.7 million in the same period last year. The decline in crypto trading volume impacted Bakkt's crypto services revenue, which fell from $1.07 billion in the same quarter last year to $243.6 million, a 77% year-on-year decrease, although most of this was offset by crypto costs and brokerage fees. At the end of the first quarter, the company held $82.6 million in cash and had no long-term debt. Bakkt stated that it is transitioning from crypto trading infrastructure to stablecoin payments and AI financial infrastructure, having completed the acquisition of Distributed Technologies Research on April 30, gaining an AI-native payment engine and a stablecoin compliance technology stack.

Bakkt Q1 Crypto Revenue Falls 77% Year-on-Year, Shifts Focus to Stablecoin Infrastructure

On May 12, SlowMist released a security monitoring report indicating that the decentralized shipping project Aurellion was attacked. The attacker seized ownership of the Diamond contract and transferred a total of 455,003 USDC from multiple authorized victim addresses.

[SlowMist: Aurellion Attacked, 455,000 USDC Lost]On May 12, SlowMist released a security monitoring report indicating that the decentralized shipping project Aurellion was attacked. The attacker seized ownership of the Diamond contract and transferred a total of 455,003 USDC from multiple authorized victim addresses.

SlowMist: Aurellion Attacked, 455,000 USDC Lost

On May 12, market sources reported that eBay has rejected GameStop's acquisition offer of approximately $56 billion, citing the reasons as "neither credible nor attractive." Previously, on May 4, GameStop had proactively proposed to acquire the American online marketplace eBay for about $56 billion. GameStop's CEO Cohen stated that the company holds about 5% of eBay's shares and proposed to acquire it for $125 per share in cash and stock. If eBay does not accept this proposal, Cohen is prepared to initiate a proxy battle and present the proposal directly to shareholders.

[eBay Rejects GameStop's $56 Billion Acquisition Offer]On May 12, market sources reported that eBay has rejected GameStop's acquisition offer of approximately $56 billion, citing the reasons as "neither credible nor attractive." Previously, on May 4, GameStop had proactively proposed to acquire the American online marketplace eBay for about $56 billion. GameStop's CEO Cohen stated that the company holds about 5% of eBay's shares and proposed to acquire it for $125 per share in cash and stock. If eBay does not accept this proposal, Cohen is prepared to initiate a proxy battle and present the proposal directly to shareholders.

eBay Rejects GameStop's $56 Billion Acquisition Offer

On May 12, Gelephu Mindfulness City in Bhutan launched an expedited licensing pathway for globally regulated cryptocurrency companies, providing eligible firms with comprehensive regulatory and banking services, 0% corporate income tax, and zero capital gains tax.

[Gelephu Mindfulness City in Bhutan Offers Tax Exemptions for Eligible Cryptocurrency Companies]On May 12, Gelephu Mindfulness City in Bhutan launched an expedited licensing pathway for globally regulated cryptocurrency companies, providing eligible firms with comprehensive regulatory and banking services, 0% corporate income tax, and zero capital gains tax.

Gelephu Mindfulness City in Bhutan Offers Tax Exemptions for Eligible Cryptocurrency Companies

Florida Governor Ron DeSantis took another swing at a U.S. digital dollar today even as he defended crypto as a whole, unveiling a series of state legislative proposals that would combat "woke politics."

[Ron Desantis Renews Attack on CBDCs as Part of 'Woke Politics']Florida Governor Ron DeSantis took another swing at a U.S. digital dollar today even as he defended crypto as a whole, unveiling a series of state legislative proposals that would combat "woke politics."

Ron Desantis Renews Attack on CBDCs as Part of 'Woke Politics'

Top Ethereum Holder Makes Bold Move Says ETH Is Close to Generational Run, Here’s When

Coinbase裁员14%，熊市和AI哪个才是主因？

David Schwartz Says Ripple Stock Is His Real Bet, Not XRP

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

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

Article

ACA fluctuated over 50% in 24 hours!

EPT fluctuates over 50% in 24 hours!

Ron Desantis Renews Attack on CBDCs as Part of 'Woke Politics'

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

All Comments0LatestHot

Trending Topics